xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Utils/InlineFunction.cpp (revision 6966ac055c3b7a39266fb982493330df7a097997)
1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
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 inlining of a function into a call site, resolving
10 // parameters and the return value as appropriate.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/DenseMap.h"
15 #include "llvm/ADT/None.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/StringExtras.h"
22 #include "llvm/ADT/iterator_range.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/BlockFrequencyInfo.h"
26 #include "llvm/Analysis/CallGraph.h"
27 #include "llvm/Analysis/CaptureTracking.h"
28 #include "llvm/Analysis/EHPersonalities.h"
29 #include "llvm/Analysis/InstructionSimplify.h"
30 #include "llvm/Analysis/ProfileSummaryInfo.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/Analysis/VectorUtils.h"
34 #include "llvm/IR/Argument.h"
35 #include "llvm/IR/BasicBlock.h"
36 #include "llvm/IR/CFG.h"
37 #include "llvm/IR/CallSite.h"
38 #include "llvm/IR/Constant.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DIBuilder.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DebugInfoMetadata.h"
43 #include "llvm/IR/DebugLoc.h"
44 #include "llvm/IR/DerivedTypes.h"
45 #include "llvm/IR/Dominators.h"
46 #include "llvm/IR/Function.h"
47 #include "llvm/IR/IRBuilder.h"
48 #include "llvm/IR/InstrTypes.h"
49 #include "llvm/IR/Instruction.h"
50 #include "llvm/IR/Instructions.h"
51 #include "llvm/IR/IntrinsicInst.h"
52 #include "llvm/IR/Intrinsics.h"
53 #include "llvm/IR/LLVMContext.h"
54 #include "llvm/IR/MDBuilder.h"
55 #include "llvm/IR/Metadata.h"
56 #include "llvm/IR/Module.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/User.h"
59 #include "llvm/IR/Value.h"
60 #include "llvm/Support/Casting.h"
61 #include "llvm/Support/CommandLine.h"
62 #include "llvm/Support/ErrorHandling.h"
63 #include "llvm/Transforms/Utils/Cloning.h"
64 #include "llvm/Transforms/Utils/ValueMapper.h"
65 #include <algorithm>
66 #include <cassert>
67 #include <cstdint>
68 #include <iterator>
69 #include <limits>
70 #include <string>
71 #include <utility>
72 #include <vector>
73 
74 using namespace llvm;
75 using ProfileCount = Function::ProfileCount;
76 
77 static cl::opt<bool>
78 EnableNoAliasConversion("enable-noalias-to-md-conversion", cl::init(true),
79   cl::Hidden,
80   cl::desc("Convert noalias attributes to metadata during inlining."));
81 
82 static cl::opt<bool>
83 PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining",
84   cl::init(true), cl::Hidden,
85   cl::desc("Convert align attributes to assumptions during inlining."));
86 
87 llvm::InlineResult llvm::InlineFunction(CallBase *CB, InlineFunctionInfo &IFI,
88                                         AAResults *CalleeAAR,
89                                         bool InsertLifetime) {
90   return InlineFunction(CallSite(CB), IFI, CalleeAAR, InsertLifetime);
91 }
92 
93 namespace {
94 
95   /// A class for recording information about inlining a landing pad.
96   class LandingPadInliningInfo {
97     /// Destination of the invoke's unwind.
98     BasicBlock *OuterResumeDest;
99 
100     /// Destination for the callee's resume.
101     BasicBlock *InnerResumeDest = nullptr;
102 
103     /// LandingPadInst associated with the invoke.
104     LandingPadInst *CallerLPad = nullptr;
105 
106     /// PHI for EH values from landingpad insts.
107     PHINode *InnerEHValuesPHI = nullptr;
108 
109     SmallVector<Value*, 8> UnwindDestPHIValues;
110 
111   public:
112     LandingPadInliningInfo(InvokeInst *II)
113         : OuterResumeDest(II->getUnwindDest()) {
114       // If there are PHI nodes in the unwind destination block, we need to keep
115       // track of which values came into them from the invoke before removing
116       // the edge from this block.
117       BasicBlock *InvokeBB = II->getParent();
118       BasicBlock::iterator I = OuterResumeDest->begin();
119       for (; isa<PHINode>(I); ++I) {
120         // Save the value to use for this edge.
121         PHINode *PHI = cast<PHINode>(I);
122         UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
123       }
124 
125       CallerLPad = cast<LandingPadInst>(I);
126     }
127 
128     /// The outer unwind destination is the target of
129     /// unwind edges introduced for calls within the inlined function.
130     BasicBlock *getOuterResumeDest() const {
131       return OuterResumeDest;
132     }
133 
134     BasicBlock *getInnerResumeDest();
135 
136     LandingPadInst *getLandingPadInst() const { return CallerLPad; }
137 
138     /// Forward the 'resume' instruction to the caller's landing pad block.
139     /// When the landing pad block has only one predecessor, this is
140     /// a simple branch. When there is more than one predecessor, we need to
141     /// split the landing pad block after the landingpad instruction and jump
142     /// to there.
143     void forwardResume(ResumeInst *RI,
144                        SmallPtrSetImpl<LandingPadInst*> &InlinedLPads);
145 
146     /// Add incoming-PHI values to the unwind destination block for the given
147     /// basic block, using the values for the original invoke's source block.
148     void addIncomingPHIValuesFor(BasicBlock *BB) const {
149       addIncomingPHIValuesForInto(BB, OuterResumeDest);
150     }
151 
152     void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const {
153       BasicBlock::iterator I = dest->begin();
154       for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
155         PHINode *phi = cast<PHINode>(I);
156         phi->addIncoming(UnwindDestPHIValues[i], src);
157       }
158     }
159   };
160 
161 } // end anonymous namespace
162 
163 /// Get or create a target for the branch from ResumeInsts.
164 BasicBlock *LandingPadInliningInfo::getInnerResumeDest() {
165   if (InnerResumeDest) return InnerResumeDest;
166 
167   // Split the landing pad.
168   BasicBlock::iterator SplitPoint = ++CallerLPad->getIterator();
169   InnerResumeDest =
170     OuterResumeDest->splitBasicBlock(SplitPoint,
171                                      OuterResumeDest->getName() + ".body");
172 
173   // The number of incoming edges we expect to the inner landing pad.
174   const unsigned PHICapacity = 2;
175 
176   // Create corresponding new PHIs for all the PHIs in the outer landing pad.
177   Instruction *InsertPoint = &InnerResumeDest->front();
178   BasicBlock::iterator I = OuterResumeDest->begin();
179   for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) {
180     PHINode *OuterPHI = cast<PHINode>(I);
181     PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity,
182                                         OuterPHI->getName() + ".lpad-body",
183                                         InsertPoint);
184     OuterPHI->replaceAllUsesWith(InnerPHI);
185     InnerPHI->addIncoming(OuterPHI, OuterResumeDest);
186   }
187 
188   // Create a PHI for the exception values.
189   InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity,
190                                      "eh.lpad-body", InsertPoint);
191   CallerLPad->replaceAllUsesWith(InnerEHValuesPHI);
192   InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest);
193 
194   // All done.
195   return InnerResumeDest;
196 }
197 
198 /// Forward the 'resume' instruction to the caller's landing pad block.
199 /// When the landing pad block has only one predecessor, this is a simple
200 /// branch. When there is more than one predecessor, we need to split the
201 /// landing pad block after the landingpad instruction and jump to there.
202 void LandingPadInliningInfo::forwardResume(
203     ResumeInst *RI, SmallPtrSetImpl<LandingPadInst *> &InlinedLPads) {
204   BasicBlock *Dest = getInnerResumeDest();
205   BasicBlock *Src = RI->getParent();
206 
207   BranchInst::Create(Dest, Src);
208 
209   // Update the PHIs in the destination. They were inserted in an order which
210   // makes this work.
211   addIncomingPHIValuesForInto(Src, Dest);
212 
213   InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src);
214   RI->eraseFromParent();
215 }
216 
217 /// Helper for getUnwindDestToken/getUnwindDestTokenHelper.
218 static Value *getParentPad(Value *EHPad) {
219   if (auto *FPI = dyn_cast<FuncletPadInst>(EHPad))
220     return FPI->getParentPad();
221   return cast<CatchSwitchInst>(EHPad)->getParentPad();
222 }
223 
224 using UnwindDestMemoTy = DenseMap<Instruction *, Value *>;
225 
226 /// Helper for getUnwindDestToken that does the descendant-ward part of
227 /// the search.
228 static Value *getUnwindDestTokenHelper(Instruction *EHPad,
229                                        UnwindDestMemoTy &MemoMap) {
230   SmallVector<Instruction *, 8> Worklist(1, EHPad);
231 
232   while (!Worklist.empty()) {
233     Instruction *CurrentPad = Worklist.pop_back_val();
234     // We only put pads on the worklist that aren't in the MemoMap.  When
235     // we find an unwind dest for a pad we may update its ancestors, but
236     // the queue only ever contains uncles/great-uncles/etc. of CurrentPad,
237     // so they should never get updated while queued on the worklist.
238     assert(!MemoMap.count(CurrentPad));
239     Value *UnwindDestToken = nullptr;
240     if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(CurrentPad)) {
241       if (CatchSwitch->hasUnwindDest()) {
242         UnwindDestToken = CatchSwitch->getUnwindDest()->getFirstNonPHI();
243       } else {
244         // Catchswitch doesn't have a 'nounwind' variant, and one might be
245         // annotated as "unwinds to caller" when really it's nounwind (see
246         // e.g. SimplifyCFGOpt::SimplifyUnreachable), so we can't infer the
247         // parent's unwind dest from this.  We can check its catchpads'
248         // descendants, since they might include a cleanuppad with an
249         // "unwinds to caller" cleanupret, which can be trusted.
250         for (auto HI = CatchSwitch->handler_begin(),
251                   HE = CatchSwitch->handler_end();
252              HI != HE && !UnwindDestToken; ++HI) {
253           BasicBlock *HandlerBlock = *HI;
254           auto *CatchPad = cast<CatchPadInst>(HandlerBlock->getFirstNonPHI());
255           for (User *Child : CatchPad->users()) {
256             // Intentionally ignore invokes here -- since the catchswitch is
257             // marked "unwind to caller", it would be a verifier error if it
258             // contained an invoke which unwinds out of it, so any invoke we'd
259             // encounter must unwind to some child of the catch.
260             if (!isa<CleanupPadInst>(Child) && !isa<CatchSwitchInst>(Child))
261               continue;
262 
263             Instruction *ChildPad = cast<Instruction>(Child);
264             auto Memo = MemoMap.find(ChildPad);
265             if (Memo == MemoMap.end()) {
266               // Haven't figured out this child pad yet; queue it.
267               Worklist.push_back(ChildPad);
268               continue;
269             }
270             // We've already checked this child, but might have found that
271             // it offers no proof either way.
272             Value *ChildUnwindDestToken = Memo->second;
273             if (!ChildUnwindDestToken)
274               continue;
275             // We already know the child's unwind dest, which can either
276             // be ConstantTokenNone to indicate unwind to caller, or can
277             // be another child of the catchpad.  Only the former indicates
278             // the unwind dest of the catchswitch.
279             if (isa<ConstantTokenNone>(ChildUnwindDestToken)) {
280               UnwindDestToken = ChildUnwindDestToken;
281               break;
282             }
283             assert(getParentPad(ChildUnwindDestToken) == CatchPad);
284           }
285         }
286       }
287     } else {
288       auto *CleanupPad = cast<CleanupPadInst>(CurrentPad);
289       for (User *U : CleanupPad->users()) {
290         if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(U)) {
291           if (BasicBlock *RetUnwindDest = CleanupRet->getUnwindDest())
292             UnwindDestToken = RetUnwindDest->getFirstNonPHI();
293           else
294             UnwindDestToken = ConstantTokenNone::get(CleanupPad->getContext());
295           break;
296         }
297         Value *ChildUnwindDestToken;
298         if (auto *Invoke = dyn_cast<InvokeInst>(U)) {
299           ChildUnwindDestToken = Invoke->getUnwindDest()->getFirstNonPHI();
300         } else if (isa<CleanupPadInst>(U) || isa<CatchSwitchInst>(U)) {
301           Instruction *ChildPad = cast<Instruction>(U);
302           auto Memo = MemoMap.find(ChildPad);
303           if (Memo == MemoMap.end()) {
304             // Haven't resolved this child yet; queue it and keep searching.
305             Worklist.push_back(ChildPad);
306             continue;
307           }
308           // We've checked this child, but still need to ignore it if it
309           // had no proof either way.
310           ChildUnwindDestToken = Memo->second;
311           if (!ChildUnwindDestToken)
312             continue;
313         } else {
314           // Not a relevant user of the cleanuppad
315           continue;
316         }
317         // In a well-formed program, the child/invoke must either unwind to
318         // an(other) child of the cleanup, or exit the cleanup.  In the
319         // first case, continue searching.
320         if (isa<Instruction>(ChildUnwindDestToken) &&
321             getParentPad(ChildUnwindDestToken) == CleanupPad)
322           continue;
323         UnwindDestToken = ChildUnwindDestToken;
324         break;
325       }
326     }
327     // If we haven't found an unwind dest for CurrentPad, we may have queued its
328     // children, so move on to the next in the worklist.
329     if (!UnwindDestToken)
330       continue;
331 
332     // Now we know that CurrentPad unwinds to UnwindDestToken.  It also exits
333     // any ancestors of CurrentPad up to but not including UnwindDestToken's
334     // parent pad.  Record this in the memo map, and check to see if the
335     // original EHPad being queried is one of the ones exited.
336     Value *UnwindParent;
337     if (auto *UnwindPad = dyn_cast<Instruction>(UnwindDestToken))
338       UnwindParent = getParentPad(UnwindPad);
339     else
340       UnwindParent = nullptr;
341     bool ExitedOriginalPad = false;
342     for (Instruction *ExitedPad = CurrentPad;
343          ExitedPad && ExitedPad != UnwindParent;
344          ExitedPad = dyn_cast<Instruction>(getParentPad(ExitedPad))) {
345       // Skip over catchpads since they just follow their catchswitches.
346       if (isa<CatchPadInst>(ExitedPad))
347         continue;
348       MemoMap[ExitedPad] = UnwindDestToken;
349       ExitedOriginalPad |= (ExitedPad == EHPad);
350     }
351 
352     if (ExitedOriginalPad)
353       return UnwindDestToken;
354 
355     // Continue the search.
356   }
357 
358   // No definitive information is contained within this funclet.
359   return nullptr;
360 }
361 
362 /// Given an EH pad, find where it unwinds.  If it unwinds to an EH pad,
363 /// return that pad instruction.  If it unwinds to caller, return
364 /// ConstantTokenNone.  If it does not have a definitive unwind destination,
365 /// return nullptr.
366 ///
367 /// This routine gets invoked for calls in funclets in inlinees when inlining
368 /// an invoke.  Since many funclets don't have calls inside them, it's queried
369 /// on-demand rather than building a map of pads to unwind dests up front.
370 /// Determining a funclet's unwind dest may require recursively searching its
371 /// descendants, and also ancestors and cousins if the descendants don't provide
372 /// an answer.  Since most funclets will have their unwind dest immediately
373 /// available as the unwind dest of a catchswitch or cleanupret, this routine
374 /// searches top-down from the given pad and then up. To avoid worst-case
375 /// quadratic run-time given that approach, it uses a memo map to avoid
376 /// re-processing funclet trees.  The callers that rewrite the IR as they go
377 /// take advantage of this, for correctness, by checking/forcing rewritten
378 /// pads' entries to match the original callee view.
379 static Value *getUnwindDestToken(Instruction *EHPad,
380                                  UnwindDestMemoTy &MemoMap) {
381   // Catchpads unwind to the same place as their catchswitch;
382   // redirct any queries on catchpads so the code below can
383   // deal with just catchswitches and cleanuppads.
384   if (auto *CPI = dyn_cast<CatchPadInst>(EHPad))
385     EHPad = CPI->getCatchSwitch();
386 
387   // Check if we've already determined the unwind dest for this pad.
388   auto Memo = MemoMap.find(EHPad);
389   if (Memo != MemoMap.end())
390     return Memo->second;
391 
392   // Search EHPad and, if necessary, its descendants.
393   Value *UnwindDestToken = getUnwindDestTokenHelper(EHPad, MemoMap);
394   assert((UnwindDestToken == nullptr) != (MemoMap.count(EHPad) != 0));
395   if (UnwindDestToken)
396     return UnwindDestToken;
397 
398   // No information is available for this EHPad from itself or any of its
399   // descendants.  An unwind all the way out to a pad in the caller would
400   // need also to agree with the unwind dest of the parent funclet, so
401   // search up the chain to try to find a funclet with information.  Put
402   // null entries in the memo map to avoid re-processing as we go up.
403   MemoMap[EHPad] = nullptr;
404 #ifndef NDEBUG
405   SmallPtrSet<Instruction *, 4> TempMemos;
406   TempMemos.insert(EHPad);
407 #endif
408   Instruction *LastUselessPad = EHPad;
409   Value *AncestorToken;
410   for (AncestorToken = getParentPad(EHPad);
411        auto *AncestorPad = dyn_cast<Instruction>(AncestorToken);
412        AncestorToken = getParentPad(AncestorToken)) {
413     // Skip over catchpads since they just follow their catchswitches.
414     if (isa<CatchPadInst>(AncestorPad))
415       continue;
416     // If the MemoMap had an entry mapping AncestorPad to nullptr, since we
417     // haven't yet called getUnwindDestTokenHelper for AncestorPad in this
418     // call to getUnwindDestToken, that would mean that AncestorPad had no
419     // information in itself, its descendants, or its ancestors.  If that
420     // were the case, then we should also have recorded the lack of information
421     // for the descendant that we're coming from.  So assert that we don't
422     // find a null entry in the MemoMap for AncestorPad.
423     assert(!MemoMap.count(AncestorPad) || MemoMap[AncestorPad]);
424     auto AncestorMemo = MemoMap.find(AncestorPad);
425     if (AncestorMemo == MemoMap.end()) {
426       UnwindDestToken = getUnwindDestTokenHelper(AncestorPad, MemoMap);
427     } else {
428       UnwindDestToken = AncestorMemo->second;
429     }
430     if (UnwindDestToken)
431       break;
432     LastUselessPad = AncestorPad;
433     MemoMap[LastUselessPad] = nullptr;
434 #ifndef NDEBUG
435     TempMemos.insert(LastUselessPad);
436 #endif
437   }
438 
439   // We know that getUnwindDestTokenHelper was called on LastUselessPad and
440   // returned nullptr (and likewise for EHPad and any of its ancestors up to
441   // LastUselessPad), so LastUselessPad has no information from below.  Since
442   // getUnwindDestTokenHelper must investigate all downward paths through
443   // no-information nodes to prove that a node has no information like this,
444   // and since any time it finds information it records it in the MemoMap for
445   // not just the immediately-containing funclet but also any ancestors also
446   // exited, it must be the case that, walking downward from LastUselessPad,
447   // visiting just those nodes which have not been mapped to an unwind dest
448   // by getUnwindDestTokenHelper (the nullptr TempMemos notwithstanding, since
449   // they are just used to keep getUnwindDestTokenHelper from repeating work),
450   // any node visited must have been exhaustively searched with no information
451   // for it found.
452   SmallVector<Instruction *, 8> Worklist(1, LastUselessPad);
453   while (!Worklist.empty()) {
454     Instruction *UselessPad = Worklist.pop_back_val();
455     auto Memo = MemoMap.find(UselessPad);
456     if (Memo != MemoMap.end() && Memo->second) {
457       // Here the name 'UselessPad' is a bit of a misnomer, because we've found
458       // that it is a funclet that does have information about unwinding to
459       // a particular destination; its parent was a useless pad.
460       // Since its parent has no information, the unwind edge must not escape
461       // the parent, and must target a sibling of this pad.  This local unwind
462       // gives us no information about EHPad.  Leave it and the subtree rooted
463       // at it alone.
464       assert(getParentPad(Memo->second) == getParentPad(UselessPad));
465       continue;
466     }
467     // We know we don't have information for UselesPad.  If it has an entry in
468     // the MemoMap (mapping it to nullptr), it must be one of the TempMemos
469     // added on this invocation of getUnwindDestToken; if a previous invocation
470     // recorded nullptr, it would have had to prove that the ancestors of
471     // UselessPad, which include LastUselessPad, had no information, and that
472     // in turn would have required proving that the descendants of
473     // LastUselesPad, which include EHPad, have no information about
474     // LastUselessPad, which would imply that EHPad was mapped to nullptr in
475     // the MemoMap on that invocation, which isn't the case if we got here.
476     assert(!MemoMap.count(UselessPad) || TempMemos.count(UselessPad));
477     // Assert as we enumerate users that 'UselessPad' doesn't have any unwind
478     // information that we'd be contradicting by making a map entry for it
479     // (which is something that getUnwindDestTokenHelper must have proved for
480     // us to get here).  Just assert on is direct users here; the checks in
481     // this downward walk at its descendants will verify that they don't have
482     // any unwind edges that exit 'UselessPad' either (i.e. they either have no
483     // unwind edges or unwind to a sibling).
484     MemoMap[UselessPad] = UnwindDestToken;
485     if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(UselessPad)) {
486       assert(CatchSwitch->getUnwindDest() == nullptr && "Expected useless pad");
487       for (BasicBlock *HandlerBlock : CatchSwitch->handlers()) {
488         auto *CatchPad = HandlerBlock->getFirstNonPHI();
489         for (User *U : CatchPad->users()) {
490           assert(
491               (!isa<InvokeInst>(U) ||
492                (getParentPad(
493                     cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) ==
494                 CatchPad)) &&
495               "Expected useless pad");
496           if (isa<CatchSwitchInst>(U) || isa<CleanupPadInst>(U))
497             Worklist.push_back(cast<Instruction>(U));
498         }
499       }
500     } else {
501       assert(isa<CleanupPadInst>(UselessPad));
502       for (User *U : UselessPad->users()) {
503         assert(!isa<CleanupReturnInst>(U) && "Expected useless pad");
504         assert((!isa<InvokeInst>(U) ||
505                 (getParentPad(
506                      cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) ==
507                  UselessPad)) &&
508                "Expected useless pad");
509         if (isa<CatchSwitchInst>(U) || isa<CleanupPadInst>(U))
510           Worklist.push_back(cast<Instruction>(U));
511       }
512     }
513   }
514 
515   return UnwindDestToken;
516 }
517 
518 /// When we inline a basic block into an invoke,
519 /// we have to turn all of the calls that can throw into invokes.
520 /// This function analyze BB to see if there are any calls, and if so,
521 /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI
522 /// nodes in that block with the values specified in InvokeDestPHIValues.
523 static BasicBlock *HandleCallsInBlockInlinedThroughInvoke(
524     BasicBlock *BB, BasicBlock *UnwindEdge,
525     UnwindDestMemoTy *FuncletUnwindMap = nullptr) {
526   for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
527     Instruction *I = &*BBI++;
528 
529     // We only need to check for function calls: inlined invoke
530     // instructions require no special handling.
531     CallInst *CI = dyn_cast<CallInst>(I);
532 
533     if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue()))
534       continue;
535 
536     // We do not need to (and in fact, cannot) convert possibly throwing calls
537     // to @llvm.experimental_deoptimize (resp. @llvm.experimental.guard) into
538     // invokes.  The caller's "segment" of the deoptimization continuation
539     // attached to the newly inlined @llvm.experimental_deoptimize
540     // (resp. @llvm.experimental.guard) call should contain the exception
541     // handling logic, if any.
542     if (auto *F = CI->getCalledFunction())
543       if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize ||
544           F->getIntrinsicID() == Intrinsic::experimental_guard)
545         continue;
546 
547     if (auto FuncletBundle = CI->getOperandBundle(LLVMContext::OB_funclet)) {
548       // This call is nested inside a funclet.  If that funclet has an unwind
549       // destination within the inlinee, then unwinding out of this call would
550       // be UB.  Rewriting this call to an invoke which targets the inlined
551       // invoke's unwind dest would give the call's parent funclet multiple
552       // unwind destinations, which is something that subsequent EH table
553       // generation can't handle and that the veirifer rejects.  So when we
554       // see such a call, leave it as a call.
555       auto *FuncletPad = cast<Instruction>(FuncletBundle->Inputs[0]);
556       Value *UnwindDestToken =
557           getUnwindDestToken(FuncletPad, *FuncletUnwindMap);
558       if (UnwindDestToken && !isa<ConstantTokenNone>(UnwindDestToken))
559         continue;
560 #ifndef NDEBUG
561       Instruction *MemoKey;
562       if (auto *CatchPad = dyn_cast<CatchPadInst>(FuncletPad))
563         MemoKey = CatchPad->getCatchSwitch();
564       else
565         MemoKey = FuncletPad;
566       assert(FuncletUnwindMap->count(MemoKey) &&
567              (*FuncletUnwindMap)[MemoKey] == UnwindDestToken &&
568              "must get memoized to avoid confusing later searches");
569 #endif // NDEBUG
570     }
571 
572     changeToInvokeAndSplitBasicBlock(CI, UnwindEdge);
573     return BB;
574   }
575   return nullptr;
576 }
577 
578 /// If we inlined an invoke site, we need to convert calls
579 /// in the body of the inlined function into invokes.
580 ///
581 /// II is the invoke instruction being inlined.  FirstNewBlock is the first
582 /// block of the inlined code (the last block is the end of the function),
583 /// and InlineCodeInfo is information about the code that got inlined.
584 static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock,
585                                     ClonedCodeInfo &InlinedCodeInfo) {
586   BasicBlock *InvokeDest = II->getUnwindDest();
587 
588   Function *Caller = FirstNewBlock->getParent();
589 
590   // The inlined code is currently at the end of the function, scan from the
591   // start of the inlined code to its end, checking for stuff we need to
592   // rewrite.
593   LandingPadInliningInfo Invoke(II);
594 
595   // Get all of the inlined landing pad instructions.
596   SmallPtrSet<LandingPadInst*, 16> InlinedLPads;
597   for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end();
598        I != E; ++I)
599     if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator()))
600       InlinedLPads.insert(II->getLandingPadInst());
601 
602   // Append the clauses from the outer landing pad instruction into the inlined
603   // landing pad instructions.
604   LandingPadInst *OuterLPad = Invoke.getLandingPadInst();
605   for (LandingPadInst *InlinedLPad : InlinedLPads) {
606     unsigned OuterNum = OuterLPad->getNumClauses();
607     InlinedLPad->reserveClauses(OuterNum);
608     for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx)
609       InlinedLPad->addClause(OuterLPad->getClause(OuterIdx));
610     if (OuterLPad->isCleanup())
611       InlinedLPad->setCleanup(true);
612   }
613 
614   for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
615        BB != E; ++BB) {
616     if (InlinedCodeInfo.ContainsCalls)
617       if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
618               &*BB, Invoke.getOuterResumeDest()))
619         // Update any PHI nodes in the exceptional block to indicate that there
620         // is now a new entry in them.
621         Invoke.addIncomingPHIValuesFor(NewBB);
622 
623     // Forward any resumes that are remaining here.
624     if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator()))
625       Invoke.forwardResume(RI, InlinedLPads);
626   }
627 
628   // Now that everything is happy, we have one final detail.  The PHI nodes in
629   // the exception destination block still have entries due to the original
630   // invoke instruction. Eliminate these entries (which might even delete the
631   // PHI node) now.
632   InvokeDest->removePredecessor(II->getParent());
633 }
634 
635 /// If we inlined an invoke site, we need to convert calls
636 /// in the body of the inlined function into invokes.
637 ///
638 /// II is the invoke instruction being inlined.  FirstNewBlock is the first
639 /// block of the inlined code (the last block is the end of the function),
640 /// and InlineCodeInfo is information about the code that got inlined.
641 static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
642                                ClonedCodeInfo &InlinedCodeInfo) {
643   BasicBlock *UnwindDest = II->getUnwindDest();
644   Function *Caller = FirstNewBlock->getParent();
645 
646   assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!");
647 
648   // If there are PHI nodes in the unwind destination block, we need to keep
649   // track of which values came into them from the invoke before removing the
650   // edge from this block.
651   SmallVector<Value *, 8> UnwindDestPHIValues;
652   BasicBlock *InvokeBB = II->getParent();
653   for (Instruction &I : *UnwindDest) {
654     // Save the value to use for this edge.
655     PHINode *PHI = dyn_cast<PHINode>(&I);
656     if (!PHI)
657       break;
658     UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB));
659   }
660 
661   // Add incoming-PHI values to the unwind destination block for the given basic
662   // block, using the values for the original invoke's source block.
663   auto UpdatePHINodes = [&](BasicBlock *Src) {
664     BasicBlock::iterator I = UnwindDest->begin();
665     for (Value *V : UnwindDestPHIValues) {
666       PHINode *PHI = cast<PHINode>(I);
667       PHI->addIncoming(V, Src);
668       ++I;
669     }
670   };
671 
672   // This connects all the instructions which 'unwind to caller' to the invoke
673   // destination.
674   UnwindDestMemoTy FuncletUnwindMap;
675   for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end();
676        BB != E; ++BB) {
677     if (auto *CRI = dyn_cast<CleanupReturnInst>(BB->getTerminator())) {
678       if (CRI->unwindsToCaller()) {
679         auto *CleanupPad = CRI->getCleanupPad();
680         CleanupReturnInst::Create(CleanupPad, UnwindDest, CRI);
681         CRI->eraseFromParent();
682         UpdatePHINodes(&*BB);
683         // Finding a cleanupret with an unwind destination would confuse
684         // subsequent calls to getUnwindDestToken, so map the cleanuppad
685         // to short-circuit any such calls and recognize this as an "unwind
686         // to caller" cleanup.
687         assert(!FuncletUnwindMap.count(CleanupPad) ||
688                isa<ConstantTokenNone>(FuncletUnwindMap[CleanupPad]));
689         FuncletUnwindMap[CleanupPad] =
690             ConstantTokenNone::get(Caller->getContext());
691       }
692     }
693 
694     Instruction *I = BB->getFirstNonPHI();
695     if (!I->isEHPad())
696       continue;
697 
698     Instruction *Replacement = nullptr;
699     if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) {
700       if (CatchSwitch->unwindsToCaller()) {
701         Value *UnwindDestToken;
702         if (auto *ParentPad =
703                 dyn_cast<Instruction>(CatchSwitch->getParentPad())) {
704           // This catchswitch is nested inside another funclet.  If that
705           // funclet has an unwind destination within the inlinee, then
706           // unwinding out of this catchswitch would be UB.  Rewriting this
707           // catchswitch to unwind to the inlined invoke's unwind dest would
708           // give the parent funclet multiple unwind destinations, which is
709           // something that subsequent EH table generation can't handle and
710           // that the veirifer rejects.  So when we see such a call, leave it
711           // as "unwind to caller".
712           UnwindDestToken = getUnwindDestToken(ParentPad, FuncletUnwindMap);
713           if (UnwindDestToken && !isa<ConstantTokenNone>(UnwindDestToken))
714             continue;
715         } else {
716           // This catchswitch has no parent to inherit constraints from, and
717           // none of its descendants can have an unwind edge that exits it and
718           // targets another funclet in the inlinee.  It may or may not have a
719           // descendant that definitively has an unwind to caller.  In either
720           // case, we'll have to assume that any unwinds out of it may need to
721           // be routed to the caller, so treat it as though it has a definitive
722           // unwind to caller.
723           UnwindDestToken = ConstantTokenNone::get(Caller->getContext());
724         }
725         auto *NewCatchSwitch = CatchSwitchInst::Create(
726             CatchSwitch->getParentPad(), UnwindDest,
727             CatchSwitch->getNumHandlers(), CatchSwitch->getName(),
728             CatchSwitch);
729         for (BasicBlock *PadBB : CatchSwitch->handlers())
730           NewCatchSwitch->addHandler(PadBB);
731         // Propagate info for the old catchswitch over to the new one in
732         // the unwind map.  This also serves to short-circuit any subsequent
733         // checks for the unwind dest of this catchswitch, which would get
734         // confused if they found the outer handler in the callee.
735         FuncletUnwindMap[NewCatchSwitch] = UnwindDestToken;
736         Replacement = NewCatchSwitch;
737       }
738     } else if (!isa<FuncletPadInst>(I)) {
739       llvm_unreachable("unexpected EHPad!");
740     }
741 
742     if (Replacement) {
743       Replacement->takeName(I);
744       I->replaceAllUsesWith(Replacement);
745       I->eraseFromParent();
746       UpdatePHINodes(&*BB);
747     }
748   }
749 
750   if (InlinedCodeInfo.ContainsCalls)
751     for (Function::iterator BB = FirstNewBlock->getIterator(),
752                             E = Caller->end();
753          BB != E; ++BB)
754       if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke(
755               &*BB, UnwindDest, &FuncletUnwindMap))
756         // Update any PHI nodes in the exceptional block to indicate that there
757         // is now a new entry in them.
758         UpdatePHINodes(NewBB);
759 
760   // Now that everything is happy, we have one final detail.  The PHI nodes in
761   // the exception destination block still have entries due to the original
762   // invoke instruction. Eliminate these entries (which might even delete the
763   // PHI node) now.
764   UnwindDest->removePredecessor(InvokeBB);
765 }
766 
767 /// When inlining a call site that has !llvm.mem.parallel_loop_access or
768 /// llvm.access.group metadata, that metadata should be propagated to all
769 /// memory-accessing cloned instructions.
770 static void PropagateParallelLoopAccessMetadata(CallSite CS,
771                                                 ValueToValueMapTy &VMap) {
772   MDNode *M =
773     CS.getInstruction()->getMetadata(LLVMContext::MD_mem_parallel_loop_access);
774   MDNode *CallAccessGroup =
775       CS.getInstruction()->getMetadata(LLVMContext::MD_access_group);
776   if (!M && !CallAccessGroup)
777     return;
778 
779   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
780        VMI != VMIE; ++VMI) {
781     if (!VMI->second)
782       continue;
783 
784     Instruction *NI = dyn_cast<Instruction>(VMI->second);
785     if (!NI)
786       continue;
787 
788     if (M) {
789       if (MDNode *PM =
790               NI->getMetadata(LLVMContext::MD_mem_parallel_loop_access)) {
791         M = MDNode::concatenate(PM, M);
792       NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M);
793       } else if (NI->mayReadOrWriteMemory()) {
794         NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M);
795       }
796     }
797 
798     if (NI->mayReadOrWriteMemory()) {
799       MDNode *UnitedAccGroups = uniteAccessGroups(
800           NI->getMetadata(LLVMContext::MD_access_group), CallAccessGroup);
801       NI->setMetadata(LLVMContext::MD_access_group, UnitedAccGroups);
802     }
803   }
804 }
805 
806 /// When inlining a function that contains noalias scope metadata,
807 /// this metadata needs to be cloned so that the inlined blocks
808 /// have different "unique scopes" at every call site. Were this not done, then
809 /// aliasing scopes from a function inlined into a caller multiple times could
810 /// not be differentiated (and this would lead to miscompiles because the
811 /// non-aliasing property communicated by the metadata could have
812 /// call-site-specific control dependencies).
813 static void CloneAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap) {
814   const Function *CalledFunc = CS.getCalledFunction();
815   SetVector<const MDNode *> MD;
816 
817   // Note: We could only clone the metadata if it is already used in the
818   // caller. I'm omitting that check here because it might confuse
819   // inter-procedural alias analysis passes. We can revisit this if it becomes
820   // an efficiency or overhead problem.
821 
822   for (const BasicBlock &I : *CalledFunc)
823     for (const Instruction &J : I) {
824       if (const MDNode *M = J.getMetadata(LLVMContext::MD_alias_scope))
825         MD.insert(M);
826       if (const MDNode *M = J.getMetadata(LLVMContext::MD_noalias))
827         MD.insert(M);
828     }
829 
830   if (MD.empty())
831     return;
832 
833   // Walk the existing metadata, adding the complete (perhaps cyclic) chain to
834   // the set.
835   SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end());
836   while (!Queue.empty()) {
837     const MDNode *M = cast<MDNode>(Queue.pop_back_val());
838     for (unsigned i = 0, ie = M->getNumOperands(); i != ie; ++i)
839       if (const MDNode *M1 = dyn_cast<MDNode>(M->getOperand(i)))
840         if (MD.insert(M1))
841           Queue.push_back(M1);
842   }
843 
844   // Now we have a complete set of all metadata in the chains used to specify
845   // the noalias scopes and the lists of those scopes.
846   SmallVector<TempMDTuple, 16> DummyNodes;
847   DenseMap<const MDNode *, TrackingMDNodeRef> MDMap;
848   for (const MDNode *I : MD) {
849     DummyNodes.push_back(MDTuple::getTemporary(CalledFunc->getContext(), None));
850     MDMap[I].reset(DummyNodes.back().get());
851   }
852 
853   // Create new metadata nodes to replace the dummy nodes, replacing old
854   // metadata references with either a dummy node or an already-created new
855   // node.
856   for (const MDNode *I : MD) {
857     SmallVector<Metadata *, 4> NewOps;
858     for (unsigned i = 0, ie = I->getNumOperands(); i != ie; ++i) {
859       const Metadata *V = I->getOperand(i);
860       if (const MDNode *M = dyn_cast<MDNode>(V))
861         NewOps.push_back(MDMap[M]);
862       else
863         NewOps.push_back(const_cast<Metadata *>(V));
864     }
865 
866     MDNode *NewM = MDNode::get(CalledFunc->getContext(), NewOps);
867     MDTuple *TempM = cast<MDTuple>(MDMap[I]);
868     assert(TempM->isTemporary() && "Expected temporary node");
869 
870     TempM->replaceAllUsesWith(NewM);
871   }
872 
873   // Now replace the metadata in the new inlined instructions with the
874   // repacements from the map.
875   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
876        VMI != VMIE; ++VMI) {
877     if (!VMI->second)
878       continue;
879 
880     Instruction *NI = dyn_cast<Instruction>(VMI->second);
881     if (!NI)
882       continue;
883 
884     if (MDNode *M = NI->getMetadata(LLVMContext::MD_alias_scope)) {
885       MDNode *NewMD = MDMap[M];
886       // If the call site also had alias scope metadata (a list of scopes to
887       // which instructions inside it might belong), propagate those scopes to
888       // the inlined instructions.
889       if (MDNode *CSM =
890               CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
891         NewMD = MDNode::concatenate(NewMD, CSM);
892       NI->setMetadata(LLVMContext::MD_alias_scope, NewMD);
893     } else if (NI->mayReadOrWriteMemory()) {
894       if (MDNode *M =
895               CS.getInstruction()->getMetadata(LLVMContext::MD_alias_scope))
896         NI->setMetadata(LLVMContext::MD_alias_scope, M);
897     }
898 
899     if (MDNode *M = NI->getMetadata(LLVMContext::MD_noalias)) {
900       MDNode *NewMD = MDMap[M];
901       // If the call site also had noalias metadata (a list of scopes with
902       // which instructions inside it don't alias), propagate those scopes to
903       // the inlined instructions.
904       if (MDNode *CSM =
905               CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
906         NewMD = MDNode::concatenate(NewMD, CSM);
907       NI->setMetadata(LLVMContext::MD_noalias, NewMD);
908     } else if (NI->mayReadOrWriteMemory()) {
909       if (MDNode *M = CS.getInstruction()->getMetadata(LLVMContext::MD_noalias))
910         NI->setMetadata(LLVMContext::MD_noalias, M);
911     }
912   }
913 }
914 
915 /// If the inlined function has noalias arguments,
916 /// then add new alias scopes for each noalias argument, tag the mapped noalias
917 /// parameters with noalias metadata specifying the new scope, and tag all
918 /// non-derived loads, stores and memory intrinsics with the new alias scopes.
919 static void AddAliasScopeMetadata(CallSite CS, ValueToValueMapTy &VMap,
920                                   const DataLayout &DL, AAResults *CalleeAAR) {
921   if (!EnableNoAliasConversion)
922     return;
923 
924   const Function *CalledFunc = CS.getCalledFunction();
925   SmallVector<const Argument *, 4> NoAliasArgs;
926 
927   for (const Argument &Arg : CalledFunc->args())
928     if (Arg.hasNoAliasAttr() && !Arg.use_empty())
929       NoAliasArgs.push_back(&Arg);
930 
931   if (NoAliasArgs.empty())
932     return;
933 
934   // To do a good job, if a noalias variable is captured, we need to know if
935   // the capture point dominates the particular use we're considering.
936   DominatorTree DT;
937   DT.recalculate(const_cast<Function&>(*CalledFunc));
938 
939   // noalias indicates that pointer values based on the argument do not alias
940   // pointer values which are not based on it. So we add a new "scope" for each
941   // noalias function argument. Accesses using pointers based on that argument
942   // become part of that alias scope, accesses using pointers not based on that
943   // argument are tagged as noalias with that scope.
944 
945   DenseMap<const Argument *, MDNode *> NewScopes;
946   MDBuilder MDB(CalledFunc->getContext());
947 
948   // Create a new scope domain for this function.
949   MDNode *NewDomain =
950     MDB.createAnonymousAliasScopeDomain(CalledFunc->getName());
951   for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) {
952     const Argument *A = NoAliasArgs[i];
953 
954     std::string Name = CalledFunc->getName();
955     if (A->hasName()) {
956       Name += ": %";
957       Name += A->getName();
958     } else {
959       Name += ": argument ";
960       Name += utostr(i);
961     }
962 
963     // Note: We always create a new anonymous root here. This is true regardless
964     // of the linkage of the callee because the aliasing "scope" is not just a
965     // property of the callee, but also all control dependencies in the caller.
966     MDNode *NewScope = MDB.createAnonymousAliasScope(NewDomain, Name);
967     NewScopes.insert(std::make_pair(A, NewScope));
968   }
969 
970   // Iterate over all new instructions in the map; for all memory-access
971   // instructions, add the alias scope metadata.
972   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
973        VMI != VMIE; ++VMI) {
974     if (const Instruction *I = dyn_cast<Instruction>(VMI->first)) {
975       if (!VMI->second)
976         continue;
977 
978       Instruction *NI = dyn_cast<Instruction>(VMI->second);
979       if (!NI)
980         continue;
981 
982       bool IsArgMemOnlyCall = false, IsFuncCall = false;
983       SmallVector<const Value *, 2> PtrArgs;
984 
985       if (const LoadInst *LI = dyn_cast<LoadInst>(I))
986         PtrArgs.push_back(LI->getPointerOperand());
987       else if (const StoreInst *SI = dyn_cast<StoreInst>(I))
988         PtrArgs.push_back(SI->getPointerOperand());
989       else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(I))
990         PtrArgs.push_back(VAAI->getPointerOperand());
991       else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
992         PtrArgs.push_back(CXI->getPointerOperand());
993       else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
994         PtrArgs.push_back(RMWI->getPointerOperand());
995       else if (const auto *Call = dyn_cast<CallBase>(I)) {
996         // If we know that the call does not access memory, then we'll still
997         // know that about the inlined clone of this call site, and we don't
998         // need to add metadata.
999         if (Call->doesNotAccessMemory())
1000           continue;
1001 
1002         IsFuncCall = true;
1003         if (CalleeAAR) {
1004           FunctionModRefBehavior MRB = CalleeAAR->getModRefBehavior(Call);
1005           if (MRB == FMRB_OnlyAccessesArgumentPointees ||
1006               MRB == FMRB_OnlyReadsArgumentPointees)
1007             IsArgMemOnlyCall = true;
1008         }
1009 
1010         for (Value *Arg : Call->args()) {
1011           // We need to check the underlying objects of all arguments, not just
1012           // the pointer arguments, because we might be passing pointers as
1013           // integers, etc.
1014           // However, if we know that the call only accesses pointer arguments,
1015           // then we only need to check the pointer arguments.
1016           if (IsArgMemOnlyCall && !Arg->getType()->isPointerTy())
1017             continue;
1018 
1019           PtrArgs.push_back(Arg);
1020         }
1021       }
1022 
1023       // If we found no pointers, then this instruction is not suitable for
1024       // pairing with an instruction to receive aliasing metadata.
1025       // However, if this is a call, this we might just alias with none of the
1026       // noalias arguments.
1027       if (PtrArgs.empty() && !IsFuncCall)
1028         continue;
1029 
1030       // It is possible that there is only one underlying object, but you
1031       // need to go through several PHIs to see it, and thus could be
1032       // repeated in the Objects list.
1033       SmallPtrSet<const Value *, 4> ObjSet;
1034       SmallVector<Metadata *, 4> Scopes, NoAliases;
1035 
1036       SmallSetVector<const Argument *, 4> NAPtrArgs;
1037       for (const Value *V : PtrArgs) {
1038         SmallVector<const Value *, 4> Objects;
1039         GetUnderlyingObjects(V, Objects, DL, /* LI = */ nullptr);
1040 
1041         for (const Value *O : Objects)
1042           ObjSet.insert(O);
1043       }
1044 
1045       // Figure out if we're derived from anything that is not a noalias
1046       // argument.
1047       bool CanDeriveViaCapture = false, UsesAliasingPtr = false;
1048       for (const Value *V : ObjSet) {
1049         // Is this value a constant that cannot be derived from any pointer
1050         // value (we need to exclude constant expressions, for example, that
1051         // are formed from arithmetic on global symbols).
1052         bool IsNonPtrConst = isa<ConstantInt>(V) || isa<ConstantFP>(V) ||
1053                              isa<ConstantPointerNull>(V) ||
1054                              isa<ConstantDataVector>(V) || isa<UndefValue>(V);
1055         if (IsNonPtrConst)
1056           continue;
1057 
1058         // If this is anything other than a noalias argument, then we cannot
1059         // completely describe the aliasing properties using alias.scope
1060         // metadata (and, thus, won't add any).
1061         if (const Argument *A = dyn_cast<Argument>(V)) {
1062           if (!A->hasNoAliasAttr())
1063             UsesAliasingPtr = true;
1064         } else {
1065           UsesAliasingPtr = true;
1066         }
1067 
1068         // If this is not some identified function-local object (which cannot
1069         // directly alias a noalias argument), or some other argument (which,
1070         // by definition, also cannot alias a noalias argument), then we could
1071         // alias a noalias argument that has been captured).
1072         if (!isa<Argument>(V) &&
1073             !isIdentifiedFunctionLocal(const_cast<Value*>(V)))
1074           CanDeriveViaCapture = true;
1075       }
1076 
1077       // A function call can always get captured noalias pointers (via other
1078       // parameters, globals, etc.).
1079       if (IsFuncCall && !IsArgMemOnlyCall)
1080         CanDeriveViaCapture = true;
1081 
1082       // First, we want to figure out all of the sets with which we definitely
1083       // don't alias. Iterate over all noalias set, and add those for which:
1084       //   1. The noalias argument is not in the set of objects from which we
1085       //      definitely derive.
1086       //   2. The noalias argument has not yet been captured.
1087       // An arbitrary function that might load pointers could see captured
1088       // noalias arguments via other noalias arguments or globals, and so we
1089       // must always check for prior capture.
1090       for (const Argument *A : NoAliasArgs) {
1091         if (!ObjSet.count(A) && (!CanDeriveViaCapture ||
1092                                  // It might be tempting to skip the
1093                                  // PointerMayBeCapturedBefore check if
1094                                  // A->hasNoCaptureAttr() is true, but this is
1095                                  // incorrect because nocapture only guarantees
1096                                  // that no copies outlive the function, not
1097                                  // that the value cannot be locally captured.
1098                                  !PointerMayBeCapturedBefore(A,
1099                                    /* ReturnCaptures */ false,
1100                                    /* StoreCaptures */ false, I, &DT)))
1101           NoAliases.push_back(NewScopes[A]);
1102       }
1103 
1104       if (!NoAliases.empty())
1105         NI->setMetadata(LLVMContext::MD_noalias,
1106                         MDNode::concatenate(
1107                             NI->getMetadata(LLVMContext::MD_noalias),
1108                             MDNode::get(CalledFunc->getContext(), NoAliases)));
1109 
1110       // Next, we want to figure out all of the sets to which we might belong.
1111       // We might belong to a set if the noalias argument is in the set of
1112       // underlying objects. If there is some non-noalias argument in our list
1113       // of underlying objects, then we cannot add a scope because the fact
1114       // that some access does not alias with any set of our noalias arguments
1115       // cannot itself guarantee that it does not alias with this access
1116       // (because there is some pointer of unknown origin involved and the
1117       // other access might also depend on this pointer). We also cannot add
1118       // scopes to arbitrary functions unless we know they don't access any
1119       // non-parameter pointer-values.
1120       bool CanAddScopes = !UsesAliasingPtr;
1121       if (CanAddScopes && IsFuncCall)
1122         CanAddScopes = IsArgMemOnlyCall;
1123 
1124       if (CanAddScopes)
1125         for (const Argument *A : NoAliasArgs) {
1126           if (ObjSet.count(A))
1127             Scopes.push_back(NewScopes[A]);
1128         }
1129 
1130       if (!Scopes.empty())
1131         NI->setMetadata(
1132             LLVMContext::MD_alias_scope,
1133             MDNode::concatenate(NI->getMetadata(LLVMContext::MD_alias_scope),
1134                                 MDNode::get(CalledFunc->getContext(), Scopes)));
1135     }
1136   }
1137 }
1138 
1139 /// If the inlined function has non-byval align arguments, then
1140 /// add @llvm.assume-based alignment assumptions to preserve this information.
1141 static void AddAlignmentAssumptions(CallSite CS, InlineFunctionInfo &IFI) {
1142   if (!PreserveAlignmentAssumptions || !IFI.GetAssumptionCache)
1143     return;
1144 
1145   AssumptionCache *AC = &(*IFI.GetAssumptionCache)(*CS.getCaller());
1146   auto &DL = CS.getCaller()->getParent()->getDataLayout();
1147 
1148   // To avoid inserting redundant assumptions, we should check for assumptions
1149   // already in the caller. To do this, we might need a DT of the caller.
1150   DominatorTree DT;
1151   bool DTCalculated = false;
1152 
1153   Function *CalledFunc = CS.getCalledFunction();
1154   for (Argument &Arg : CalledFunc->args()) {
1155     unsigned Align = Arg.getType()->isPointerTy() ? Arg.getParamAlignment() : 0;
1156     if (Align && !Arg.hasByValOrInAllocaAttr() && !Arg.hasNUses(0)) {
1157       if (!DTCalculated) {
1158         DT.recalculate(*CS.getCaller());
1159         DTCalculated = true;
1160       }
1161 
1162       // If we can already prove the asserted alignment in the context of the
1163       // caller, then don't bother inserting the assumption.
1164       Value *ArgVal = CS.getArgument(Arg.getArgNo());
1165       if (getKnownAlignment(ArgVal, DL, CS.getInstruction(), AC, &DT) >= Align)
1166         continue;
1167 
1168       CallInst *NewAsmp = IRBuilder<>(CS.getInstruction())
1169                               .CreateAlignmentAssumption(DL, ArgVal, Align);
1170       AC->registerAssumption(NewAsmp);
1171     }
1172   }
1173 }
1174 
1175 /// Once we have cloned code over from a callee into the caller,
1176 /// update the specified callgraph to reflect the changes we made.
1177 /// Note that it's possible that not all code was copied over, so only
1178 /// some edges of the callgraph may remain.
1179 static void UpdateCallGraphAfterInlining(CallSite CS,
1180                                          Function::iterator FirstNewBlock,
1181                                          ValueToValueMapTy &VMap,
1182                                          InlineFunctionInfo &IFI) {
1183   CallGraph &CG = *IFI.CG;
1184   const Function *Caller = CS.getCaller();
1185   const Function *Callee = CS.getCalledFunction();
1186   CallGraphNode *CalleeNode = CG[Callee];
1187   CallGraphNode *CallerNode = CG[Caller];
1188 
1189   // Since we inlined some uninlined call sites in the callee into the caller,
1190   // add edges from the caller to all of the callees of the callee.
1191   CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
1192 
1193   // Consider the case where CalleeNode == CallerNode.
1194   CallGraphNode::CalledFunctionsVector CallCache;
1195   if (CalleeNode == CallerNode) {
1196     CallCache.assign(I, E);
1197     I = CallCache.begin();
1198     E = CallCache.end();
1199   }
1200 
1201   for (; I != E; ++I) {
1202     const Value *OrigCall = I->first;
1203 
1204     ValueToValueMapTy::iterator VMI = VMap.find(OrigCall);
1205     // Only copy the edge if the call was inlined!
1206     if (VMI == VMap.end() || VMI->second == nullptr)
1207       continue;
1208 
1209     // If the call was inlined, but then constant folded, there is no edge to
1210     // add.  Check for this case.
1211     auto *NewCall = dyn_cast<CallBase>(VMI->second);
1212     if (!NewCall)
1213       continue;
1214 
1215     // We do not treat intrinsic calls like real function calls because we
1216     // expect them to become inline code; do not add an edge for an intrinsic.
1217     if (NewCall->getCalledFunction() &&
1218         NewCall->getCalledFunction()->isIntrinsic())
1219       continue;
1220 
1221     // Remember that this call site got inlined for the client of
1222     // InlineFunction.
1223     IFI.InlinedCalls.push_back(NewCall);
1224 
1225     // It's possible that inlining the callsite will cause it to go from an
1226     // indirect to a direct call by resolving a function pointer.  If this
1227     // happens, set the callee of the new call site to a more precise
1228     // destination.  This can also happen if the call graph node of the caller
1229     // was just unnecessarily imprecise.
1230     if (!I->second->getFunction())
1231       if (Function *F = NewCall->getCalledFunction()) {
1232         // Indirect call site resolved to direct call.
1233         CallerNode->addCalledFunction(NewCall, CG[F]);
1234 
1235         continue;
1236       }
1237 
1238     CallerNode->addCalledFunction(NewCall, I->second);
1239   }
1240 
1241   // Update the call graph by deleting the edge from Callee to Caller.  We must
1242   // do this after the loop above in case Caller and Callee are the same.
1243   CallerNode->removeCallEdgeFor(*cast<CallBase>(CS.getInstruction()));
1244 }
1245 
1246 static void HandleByValArgumentInit(Value *Dst, Value *Src, Module *M,
1247                                     BasicBlock *InsertBlock,
1248                                     InlineFunctionInfo &IFI) {
1249   Type *AggTy = cast<PointerType>(Src->getType())->getElementType();
1250   IRBuilder<> Builder(InsertBlock, InsertBlock->begin());
1251 
1252   Value *Size = Builder.getInt64(M->getDataLayout().getTypeStoreSize(AggTy));
1253 
1254   // Always generate a memcpy of alignment 1 here because we don't know
1255   // the alignment of the src pointer.  Other optimizations can infer
1256   // better alignment.
1257   Builder.CreateMemCpy(Dst, /*DstAlign*/1, Src, /*SrcAlign*/1, Size);
1258 }
1259 
1260 /// When inlining a call site that has a byval argument,
1261 /// we have to make the implicit memcpy explicit by adding it.
1262 static Value *HandleByValArgument(Value *Arg, Instruction *TheCall,
1263                                   const Function *CalledFunc,
1264                                   InlineFunctionInfo &IFI,
1265                                   unsigned ByValAlignment) {
1266   PointerType *ArgTy = cast<PointerType>(Arg->getType());
1267   Type *AggTy = ArgTy->getElementType();
1268 
1269   Function *Caller = TheCall->getFunction();
1270   const DataLayout &DL = Caller->getParent()->getDataLayout();
1271 
1272   // If the called function is readonly, then it could not mutate the caller's
1273   // copy of the byval'd memory.  In this case, it is safe to elide the copy and
1274   // temporary.
1275   if (CalledFunc->onlyReadsMemory()) {
1276     // If the byval argument has a specified alignment that is greater than the
1277     // passed in pointer, then we either have to round up the input pointer or
1278     // give up on this transformation.
1279     if (ByValAlignment <= 1)  // 0 = unspecified, 1 = no particular alignment.
1280       return Arg;
1281 
1282     AssumptionCache *AC =
1283         IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr;
1284 
1285     // If the pointer is already known to be sufficiently aligned, or if we can
1286     // round it up to a larger alignment, then we don't need a temporary.
1287     if (getOrEnforceKnownAlignment(Arg, ByValAlignment, DL, TheCall, AC) >=
1288         ByValAlignment)
1289       return Arg;
1290 
1291     // Otherwise, we have to make a memcpy to get a safe alignment.  This is bad
1292     // for code quality, but rarely happens and is required for correctness.
1293   }
1294 
1295   // Create the alloca.  If we have DataLayout, use nice alignment.
1296   unsigned Align = DL.getPrefTypeAlignment(AggTy);
1297 
1298   // If the byval had an alignment specified, we *must* use at least that
1299   // alignment, as it is required by the byval argument (and uses of the
1300   // pointer inside the callee).
1301   Align = std::max(Align, ByValAlignment);
1302 
1303   Value *NewAlloca = new AllocaInst(AggTy, DL.getAllocaAddrSpace(),
1304                                     nullptr, Align, Arg->getName(),
1305                                     &*Caller->begin()->begin());
1306   IFI.StaticAllocas.push_back(cast<AllocaInst>(NewAlloca));
1307 
1308   // Uses of the argument in the function should use our new alloca
1309   // instead.
1310   return NewAlloca;
1311 }
1312 
1313 // Check whether this Value is used by a lifetime intrinsic.
1314 static bool isUsedByLifetimeMarker(Value *V) {
1315   for (User *U : V->users())
1316     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U))
1317       if (II->isLifetimeStartOrEnd())
1318         return true;
1319   return false;
1320 }
1321 
1322 // Check whether the given alloca already has
1323 // lifetime.start or lifetime.end intrinsics.
1324 static bool hasLifetimeMarkers(AllocaInst *AI) {
1325   Type *Ty = AI->getType();
1326   Type *Int8PtrTy = Type::getInt8PtrTy(Ty->getContext(),
1327                                        Ty->getPointerAddressSpace());
1328   if (Ty == Int8PtrTy)
1329     return isUsedByLifetimeMarker(AI);
1330 
1331   // Do a scan to find all the casts to i8*.
1332   for (User *U : AI->users()) {
1333     if (U->getType() != Int8PtrTy) continue;
1334     if (U->stripPointerCasts() != AI) continue;
1335     if (isUsedByLifetimeMarker(U))
1336       return true;
1337   }
1338   return false;
1339 }
1340 
1341 /// Return the result of AI->isStaticAlloca() if AI were moved to the entry
1342 /// block. Allocas used in inalloca calls and allocas of dynamic array size
1343 /// cannot be static.
1344 static bool allocaWouldBeStaticInEntry(const AllocaInst *AI ) {
1345   return isa<Constant>(AI->getArraySize()) && !AI->isUsedWithInAlloca();
1346 }
1347 
1348 /// Returns a DebugLoc for a new DILocation which is a clone of \p OrigDL
1349 /// inlined at \p InlinedAt. \p IANodes is an inlined-at cache.
1350 static DebugLoc inlineDebugLoc(DebugLoc OrigDL, DILocation *InlinedAt,
1351                                LLVMContext &Ctx,
1352                                DenseMap<const MDNode *, MDNode *> &IANodes) {
1353   auto IA = DebugLoc::appendInlinedAt(OrigDL, InlinedAt, Ctx, IANodes);
1354   return DebugLoc::get(OrigDL.getLine(), OrigDL.getCol(), OrigDL.getScope(),
1355                        IA);
1356 }
1357 
1358 /// Returns the LoopID for a loop which has has been cloned from another
1359 /// function for inlining with the new inlined-at start and end locs.
1360 static MDNode *inlineLoopID(const MDNode *OrigLoopId, DILocation *InlinedAt,
1361                             LLVMContext &Ctx,
1362                             DenseMap<const MDNode *, MDNode *> &IANodes) {
1363   assert(OrigLoopId && OrigLoopId->getNumOperands() > 0 &&
1364          "Loop ID needs at least one operand");
1365   assert(OrigLoopId && OrigLoopId->getOperand(0).get() == OrigLoopId &&
1366          "Loop ID should refer to itself");
1367 
1368   // Save space for the self-referential LoopID.
1369   SmallVector<Metadata *, 4> MDs = {nullptr};
1370 
1371   for (unsigned i = 1; i < OrigLoopId->getNumOperands(); ++i) {
1372     Metadata *MD = OrigLoopId->getOperand(i);
1373     // Update the DILocations to encode the inlined-at metadata.
1374     if (DILocation *DL = dyn_cast<DILocation>(MD))
1375       MDs.push_back(inlineDebugLoc(DL, InlinedAt, Ctx, IANodes));
1376     else
1377       MDs.push_back(MD);
1378   }
1379 
1380   MDNode *NewLoopID = MDNode::getDistinct(Ctx, MDs);
1381   // Insert the self-referential LoopID.
1382   NewLoopID->replaceOperandWith(0, NewLoopID);
1383   return NewLoopID;
1384 }
1385 
1386 /// Update inlined instructions' line numbers to
1387 /// to encode location where these instructions are inlined.
1388 static void fixupLineNumbers(Function *Fn, Function::iterator FI,
1389                              Instruction *TheCall, bool CalleeHasDebugInfo) {
1390   const DebugLoc &TheCallDL = TheCall->getDebugLoc();
1391   if (!TheCallDL)
1392     return;
1393 
1394   auto &Ctx = Fn->getContext();
1395   DILocation *InlinedAtNode = TheCallDL;
1396 
1397   // Create a unique call site, not to be confused with any other call from the
1398   // same location.
1399   InlinedAtNode = DILocation::getDistinct(
1400       Ctx, InlinedAtNode->getLine(), InlinedAtNode->getColumn(),
1401       InlinedAtNode->getScope(), InlinedAtNode->getInlinedAt());
1402 
1403   // Cache the inlined-at nodes as they're built so they are reused, without
1404   // this every instruction's inlined-at chain would become distinct from each
1405   // other.
1406   DenseMap<const MDNode *, MDNode *> IANodes;
1407 
1408   for (; FI != Fn->end(); ++FI) {
1409     for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
1410          BI != BE; ++BI) {
1411       // Loop metadata needs to be updated so that the start and end locs
1412       // reference inlined-at locations.
1413       if (MDNode *LoopID = BI->getMetadata(LLVMContext::MD_loop)) {
1414         MDNode *NewLoopID =
1415             inlineLoopID(LoopID, InlinedAtNode, BI->getContext(), IANodes);
1416         BI->setMetadata(LLVMContext::MD_loop, NewLoopID);
1417       }
1418 
1419       if (DebugLoc DL = BI->getDebugLoc()) {
1420         DebugLoc IDL =
1421             inlineDebugLoc(DL, InlinedAtNode, BI->getContext(), IANodes);
1422         BI->setDebugLoc(IDL);
1423         continue;
1424       }
1425 
1426       if (CalleeHasDebugInfo)
1427         continue;
1428 
1429       // If the inlined instruction has no line number, make it look as if it
1430       // originates from the call location. This is important for
1431       // ((__always_inline__, __nodebug__)) functions which must use caller
1432       // location for all instructions in their function body.
1433 
1434       // Don't update static allocas, as they may get moved later.
1435       if (auto *AI = dyn_cast<AllocaInst>(BI))
1436         if (allocaWouldBeStaticInEntry(AI))
1437           continue;
1438 
1439       BI->setDebugLoc(TheCallDL);
1440     }
1441   }
1442 }
1443 
1444 /// Update the block frequencies of the caller after a callee has been inlined.
1445 ///
1446 /// Each block cloned into the caller has its block frequency scaled by the
1447 /// ratio of CallSiteFreq/CalleeEntryFreq. This ensures that the cloned copy of
1448 /// callee's entry block gets the same frequency as the callsite block and the
1449 /// relative frequencies of all cloned blocks remain the same after cloning.
1450 static void updateCallerBFI(BasicBlock *CallSiteBlock,
1451                             const ValueToValueMapTy &VMap,
1452                             BlockFrequencyInfo *CallerBFI,
1453                             BlockFrequencyInfo *CalleeBFI,
1454                             const BasicBlock &CalleeEntryBlock) {
1455   SmallPtrSet<BasicBlock *, 16> ClonedBBs;
1456   for (auto const &Entry : VMap) {
1457     if (!isa<BasicBlock>(Entry.first) || !Entry.second)
1458       continue;
1459     auto *OrigBB = cast<BasicBlock>(Entry.first);
1460     auto *ClonedBB = cast<BasicBlock>(Entry.second);
1461     uint64_t Freq = CalleeBFI->getBlockFreq(OrigBB).getFrequency();
1462     if (!ClonedBBs.insert(ClonedBB).second) {
1463       // Multiple blocks in the callee might get mapped to one cloned block in
1464       // the caller since we prune the callee as we clone it. When that happens,
1465       // we want to use the maximum among the original blocks' frequencies.
1466       uint64_t NewFreq = CallerBFI->getBlockFreq(ClonedBB).getFrequency();
1467       if (NewFreq > Freq)
1468         Freq = NewFreq;
1469     }
1470     CallerBFI->setBlockFreq(ClonedBB, Freq);
1471   }
1472   BasicBlock *EntryClone = cast<BasicBlock>(VMap.lookup(&CalleeEntryBlock));
1473   CallerBFI->setBlockFreqAndScale(
1474       EntryClone, CallerBFI->getBlockFreq(CallSiteBlock).getFrequency(),
1475       ClonedBBs);
1476 }
1477 
1478 /// Update the branch metadata for cloned call instructions.
1479 static void updateCallProfile(Function *Callee, const ValueToValueMapTy &VMap,
1480                               const ProfileCount &CalleeEntryCount,
1481                               const Instruction *TheCall,
1482                               ProfileSummaryInfo *PSI,
1483                               BlockFrequencyInfo *CallerBFI) {
1484   if (!CalleeEntryCount.hasValue() || CalleeEntryCount.isSynthetic() ||
1485       CalleeEntryCount.getCount() < 1)
1486     return;
1487   auto CallSiteCount = PSI ? PSI->getProfileCount(TheCall, CallerBFI) : None;
1488   int64_t CallCount =
1489       std::min(CallSiteCount.hasValue() ? CallSiteCount.getValue() : 0,
1490                CalleeEntryCount.getCount());
1491   updateProfileCallee(Callee, -CallCount, &VMap);
1492 }
1493 
1494 void llvm::updateProfileCallee(
1495     Function *Callee, int64_t entryDelta,
1496     const ValueMap<const Value *, WeakTrackingVH> *VMap) {
1497   auto CalleeCount = Callee->getEntryCount();
1498   if (!CalleeCount.hasValue())
1499     return;
1500 
1501   uint64_t priorEntryCount = CalleeCount.getCount();
1502   uint64_t newEntryCount;
1503 
1504   // Since CallSiteCount is an estimate, it could exceed the original callee
1505   // count and has to be set to 0 so guard against underflow.
1506   if (entryDelta < 0 && static_cast<uint64_t>(-entryDelta) > priorEntryCount)
1507     newEntryCount = 0;
1508   else
1509     newEntryCount = priorEntryCount + entryDelta;
1510 
1511   Callee->setEntryCount(newEntryCount);
1512 
1513   // During inlining ?
1514   if (VMap) {
1515     uint64_t cloneEntryCount = priorEntryCount - newEntryCount;
1516     for (auto const &Entry : *VMap)
1517       if (isa<CallInst>(Entry.first))
1518         if (auto *CI = dyn_cast_or_null<CallInst>(Entry.second))
1519           CI->updateProfWeight(cloneEntryCount, priorEntryCount);
1520   }
1521   for (BasicBlock &BB : *Callee)
1522     // No need to update the callsite if it is pruned during inlining.
1523     if (!VMap || VMap->count(&BB))
1524       for (Instruction &I : BB)
1525         if (CallInst *CI = dyn_cast<CallInst>(&I))
1526           CI->updateProfWeight(newEntryCount, priorEntryCount);
1527 }
1528 
1529 /// This function inlines the called function into the basic block of the
1530 /// caller. This returns false if it is not possible to inline this call.
1531 /// The program is still in a well defined state if this occurs though.
1532 ///
1533 /// Note that this only does one level of inlining.  For example, if the
1534 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
1535 /// exists in the instruction stream.  Similarly this will inline a recursive
1536 /// function by one level.
1537 llvm::InlineResult llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI,
1538                                         AAResults *CalleeAAR,
1539                                         bool InsertLifetime,
1540                                         Function *ForwardVarArgsTo) {
1541   Instruction *TheCall = CS.getInstruction();
1542   assert(TheCall->getParent() && TheCall->getFunction()
1543          && "Instruction not in function!");
1544 
1545   // FIXME: we don't inline callbr yet.
1546   if (isa<CallBrInst>(TheCall))
1547     return false;
1548 
1549   // If IFI has any state in it, zap it before we fill it in.
1550   IFI.reset();
1551 
1552   Function *CalledFunc = CS.getCalledFunction();
1553   if (!CalledFunc ||               // Can't inline external function or indirect
1554       CalledFunc->isDeclaration()) // call!
1555     return "external or indirect";
1556 
1557   // The inliner does not know how to inline through calls with operand bundles
1558   // in general ...
1559   if (CS.hasOperandBundles()) {
1560     for (int i = 0, e = CS.getNumOperandBundles(); i != e; ++i) {
1561       uint32_t Tag = CS.getOperandBundleAt(i).getTagID();
1562       // ... but it knows how to inline through "deopt" operand bundles ...
1563       if (Tag == LLVMContext::OB_deopt)
1564         continue;
1565       // ... and "funclet" operand bundles.
1566       if (Tag == LLVMContext::OB_funclet)
1567         continue;
1568 
1569       return "unsupported operand bundle";
1570     }
1571   }
1572 
1573   // If the call to the callee cannot throw, set the 'nounwind' flag on any
1574   // calls that we inline.
1575   bool MarkNoUnwind = CS.doesNotThrow();
1576 
1577   BasicBlock *OrigBB = TheCall->getParent();
1578   Function *Caller = OrigBB->getParent();
1579 
1580   // GC poses two hazards to inlining, which only occur when the callee has GC:
1581   //  1. If the caller has no GC, then the callee's GC must be propagated to the
1582   //     caller.
1583   //  2. If the caller has a differing GC, it is invalid to inline.
1584   if (CalledFunc->hasGC()) {
1585     if (!Caller->hasGC())
1586       Caller->setGC(CalledFunc->getGC());
1587     else if (CalledFunc->getGC() != Caller->getGC())
1588       return "incompatible GC";
1589   }
1590 
1591   // Get the personality function from the callee if it contains a landing pad.
1592   Constant *CalledPersonality =
1593       CalledFunc->hasPersonalityFn()
1594           ? CalledFunc->getPersonalityFn()->stripPointerCasts()
1595           : nullptr;
1596 
1597   // Find the personality function used by the landing pads of the caller. If it
1598   // exists, then check to see that it matches the personality function used in
1599   // the callee.
1600   Constant *CallerPersonality =
1601       Caller->hasPersonalityFn()
1602           ? Caller->getPersonalityFn()->stripPointerCasts()
1603           : nullptr;
1604   if (CalledPersonality) {
1605     if (!CallerPersonality)
1606       Caller->setPersonalityFn(CalledPersonality);
1607     // If the personality functions match, then we can perform the
1608     // inlining. Otherwise, we can't inline.
1609     // TODO: This isn't 100% true. Some personality functions are proper
1610     //       supersets of others and can be used in place of the other.
1611     else if (CalledPersonality != CallerPersonality)
1612       return "incompatible personality";
1613   }
1614 
1615   // We need to figure out which funclet the callsite was in so that we may
1616   // properly nest the callee.
1617   Instruction *CallSiteEHPad = nullptr;
1618   if (CallerPersonality) {
1619     EHPersonality Personality = classifyEHPersonality(CallerPersonality);
1620     if (isScopedEHPersonality(Personality)) {
1621       Optional<OperandBundleUse> ParentFunclet =
1622           CS.getOperandBundle(LLVMContext::OB_funclet);
1623       if (ParentFunclet)
1624         CallSiteEHPad = cast<FuncletPadInst>(ParentFunclet->Inputs.front());
1625 
1626       // OK, the inlining site is legal.  What about the target function?
1627 
1628       if (CallSiteEHPad) {
1629         if (Personality == EHPersonality::MSVC_CXX) {
1630           // The MSVC personality cannot tolerate catches getting inlined into
1631           // cleanup funclets.
1632           if (isa<CleanupPadInst>(CallSiteEHPad)) {
1633             // Ok, the call site is within a cleanuppad.  Let's check the callee
1634             // for catchpads.
1635             for (const BasicBlock &CalledBB : *CalledFunc) {
1636               if (isa<CatchSwitchInst>(CalledBB.getFirstNonPHI()))
1637                 return "catch in cleanup funclet";
1638             }
1639           }
1640         } else if (isAsynchronousEHPersonality(Personality)) {
1641           // SEH is even less tolerant, there may not be any sort of exceptional
1642           // funclet in the callee.
1643           for (const BasicBlock &CalledBB : *CalledFunc) {
1644             if (CalledBB.isEHPad())
1645               return "SEH in cleanup funclet";
1646           }
1647         }
1648       }
1649     }
1650   }
1651 
1652   // Determine if we are dealing with a call in an EHPad which does not unwind
1653   // to caller.
1654   bool EHPadForCallUnwindsLocally = false;
1655   if (CallSiteEHPad && CS.isCall()) {
1656     UnwindDestMemoTy FuncletUnwindMap;
1657     Value *CallSiteUnwindDestToken =
1658         getUnwindDestToken(CallSiteEHPad, FuncletUnwindMap);
1659 
1660     EHPadForCallUnwindsLocally =
1661         CallSiteUnwindDestToken &&
1662         !isa<ConstantTokenNone>(CallSiteUnwindDestToken);
1663   }
1664 
1665   // Get an iterator to the last basic block in the function, which will have
1666   // the new function inlined after it.
1667   Function::iterator LastBlock = --Caller->end();
1668 
1669   // Make sure to capture all of the return instructions from the cloned
1670   // function.
1671   SmallVector<ReturnInst*, 8> Returns;
1672   ClonedCodeInfo InlinedFunctionInfo;
1673   Function::iterator FirstNewBlock;
1674 
1675   { // Scope to destroy VMap after cloning.
1676     ValueToValueMapTy VMap;
1677     // Keep a list of pair (dst, src) to emit byval initializations.
1678     SmallVector<std::pair<Value*, Value*>, 4> ByValInit;
1679 
1680     auto &DL = Caller->getParent()->getDataLayout();
1681 
1682     // Calculate the vector of arguments to pass into the function cloner, which
1683     // matches up the formal to the actual argument values.
1684     CallSite::arg_iterator AI = CS.arg_begin();
1685     unsigned ArgNo = 0;
1686     for (Function::arg_iterator I = CalledFunc->arg_begin(),
1687          E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
1688       Value *ActualArg = *AI;
1689 
1690       // When byval arguments actually inlined, we need to make the copy implied
1691       // by them explicit.  However, we don't do this if the callee is readonly
1692       // or readnone, because the copy would be unneeded: the callee doesn't
1693       // modify the struct.
1694       if (CS.isByValArgument(ArgNo)) {
1695         ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI,
1696                                         CalledFunc->getParamAlignment(ArgNo));
1697         if (ActualArg != *AI)
1698           ByValInit.push_back(std::make_pair(ActualArg, (Value*) *AI));
1699       }
1700 
1701       VMap[&*I] = ActualArg;
1702     }
1703 
1704     // Add alignment assumptions if necessary. We do this before the inlined
1705     // instructions are actually cloned into the caller so that we can easily
1706     // check what will be known at the start of the inlined code.
1707     AddAlignmentAssumptions(CS, IFI);
1708 
1709     // We want the inliner to prune the code as it copies.  We would LOVE to
1710     // have no dead or constant instructions leftover after inlining occurs
1711     // (which can happen, e.g., because an argument was constant), but we'll be
1712     // happy with whatever the cloner can do.
1713     CloneAndPruneFunctionInto(Caller, CalledFunc, VMap,
1714                               /*ModuleLevelChanges=*/false, Returns, ".i",
1715                               &InlinedFunctionInfo, TheCall);
1716     // Remember the first block that is newly cloned over.
1717     FirstNewBlock = LastBlock; ++FirstNewBlock;
1718 
1719     if (IFI.CallerBFI != nullptr && IFI.CalleeBFI != nullptr)
1720       // Update the BFI of blocks cloned into the caller.
1721       updateCallerBFI(OrigBB, VMap, IFI.CallerBFI, IFI.CalleeBFI,
1722                       CalledFunc->front());
1723 
1724     updateCallProfile(CalledFunc, VMap, CalledFunc->getEntryCount(), TheCall,
1725                       IFI.PSI, IFI.CallerBFI);
1726 
1727     // Inject byval arguments initialization.
1728     for (std::pair<Value*, Value*> &Init : ByValInit)
1729       HandleByValArgumentInit(Init.first, Init.second, Caller->getParent(),
1730                               &*FirstNewBlock, IFI);
1731 
1732     Optional<OperandBundleUse> ParentDeopt =
1733         CS.getOperandBundle(LLVMContext::OB_deopt);
1734     if (ParentDeopt) {
1735       SmallVector<OperandBundleDef, 2> OpDefs;
1736 
1737       for (auto &VH : InlinedFunctionInfo.OperandBundleCallSites) {
1738         Instruction *I = dyn_cast_or_null<Instruction>(VH);
1739         if (!I) continue;  // instruction was DCE'd or RAUW'ed to undef
1740 
1741         OpDefs.clear();
1742 
1743         CallSite ICS(I);
1744         OpDefs.reserve(ICS.getNumOperandBundles());
1745 
1746         for (unsigned i = 0, e = ICS.getNumOperandBundles(); i < e; ++i) {
1747           auto ChildOB = ICS.getOperandBundleAt(i);
1748           if (ChildOB.getTagID() != LLVMContext::OB_deopt) {
1749             // If the inlined call has other operand bundles, let them be
1750             OpDefs.emplace_back(ChildOB);
1751             continue;
1752           }
1753 
1754           // It may be useful to separate this logic (of handling operand
1755           // bundles) out to a separate "policy" component if this gets crowded.
1756           // Prepend the parent's deoptimization continuation to the newly
1757           // inlined call's deoptimization continuation.
1758           std::vector<Value *> MergedDeoptArgs;
1759           MergedDeoptArgs.reserve(ParentDeopt->Inputs.size() +
1760                                   ChildOB.Inputs.size());
1761 
1762           MergedDeoptArgs.insert(MergedDeoptArgs.end(),
1763                                  ParentDeopt->Inputs.begin(),
1764                                  ParentDeopt->Inputs.end());
1765           MergedDeoptArgs.insert(MergedDeoptArgs.end(), ChildOB.Inputs.begin(),
1766                                  ChildOB.Inputs.end());
1767 
1768           OpDefs.emplace_back("deopt", std::move(MergedDeoptArgs));
1769         }
1770 
1771         Instruction *NewI = nullptr;
1772         if (isa<CallInst>(I))
1773           NewI = CallInst::Create(cast<CallInst>(I), OpDefs, I);
1774         else if (isa<CallBrInst>(I))
1775           NewI = CallBrInst::Create(cast<CallBrInst>(I), OpDefs, I);
1776         else
1777           NewI = InvokeInst::Create(cast<InvokeInst>(I), OpDefs, I);
1778 
1779         // Note: the RAUW does the appropriate fixup in VMap, so we need to do
1780         // this even if the call returns void.
1781         I->replaceAllUsesWith(NewI);
1782 
1783         VH = nullptr;
1784         I->eraseFromParent();
1785       }
1786     }
1787 
1788     // Update the callgraph if requested.
1789     if (IFI.CG)
1790       UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI);
1791 
1792     // For 'nodebug' functions, the associated DISubprogram is always null.
1793     // Conservatively avoid propagating the callsite debug location to
1794     // instructions inlined from a function whose DISubprogram is not null.
1795     fixupLineNumbers(Caller, FirstNewBlock, TheCall,
1796                      CalledFunc->getSubprogram() != nullptr);
1797 
1798     // Clone existing noalias metadata if necessary.
1799     CloneAliasScopeMetadata(CS, VMap);
1800 
1801     // Add noalias metadata if necessary.
1802     AddAliasScopeMetadata(CS, VMap, DL, CalleeAAR);
1803 
1804     // Propagate llvm.mem.parallel_loop_access if necessary.
1805     PropagateParallelLoopAccessMetadata(CS, VMap);
1806 
1807     // Register any cloned assumptions.
1808     if (IFI.GetAssumptionCache)
1809       for (BasicBlock &NewBlock :
1810            make_range(FirstNewBlock->getIterator(), Caller->end()))
1811         for (Instruction &I : NewBlock) {
1812           if (auto *II = dyn_cast<IntrinsicInst>(&I))
1813             if (II->getIntrinsicID() == Intrinsic::assume)
1814               (*IFI.GetAssumptionCache)(*Caller).registerAssumption(II);
1815         }
1816   }
1817 
1818   // If there are any alloca instructions in the block that used to be the entry
1819   // block for the callee, move them to the entry block of the caller.  First
1820   // calculate which instruction they should be inserted before.  We insert the
1821   // instructions at the end of the current alloca list.
1822   {
1823     BasicBlock::iterator InsertPoint = Caller->begin()->begin();
1824     for (BasicBlock::iterator I = FirstNewBlock->begin(),
1825          E = FirstNewBlock->end(); I != E; ) {
1826       AllocaInst *AI = dyn_cast<AllocaInst>(I++);
1827       if (!AI) continue;
1828 
1829       // If the alloca is now dead, remove it.  This often occurs due to code
1830       // specialization.
1831       if (AI->use_empty()) {
1832         AI->eraseFromParent();
1833         continue;
1834       }
1835 
1836       if (!allocaWouldBeStaticInEntry(AI))
1837         continue;
1838 
1839       // Keep track of the static allocas that we inline into the caller.
1840       IFI.StaticAllocas.push_back(AI);
1841 
1842       // Scan for the block of allocas that we can move over, and move them
1843       // all at once.
1844       while (isa<AllocaInst>(I) &&
1845              allocaWouldBeStaticInEntry(cast<AllocaInst>(I))) {
1846         IFI.StaticAllocas.push_back(cast<AllocaInst>(I));
1847         ++I;
1848       }
1849 
1850       // Transfer all of the allocas over in a block.  Using splice means
1851       // that the instructions aren't removed from the symbol table, then
1852       // reinserted.
1853       Caller->getEntryBlock().getInstList().splice(
1854           InsertPoint, FirstNewBlock->getInstList(), AI->getIterator(), I);
1855     }
1856     // Move any dbg.declares describing the allocas into the entry basic block.
1857     DIBuilder DIB(*Caller->getParent());
1858     for (auto &AI : IFI.StaticAllocas)
1859       replaceDbgDeclareForAlloca(AI, AI, DIB, DIExpression::ApplyOffset, 0);
1860   }
1861 
1862   SmallVector<Value*,4> VarArgsToForward;
1863   SmallVector<AttributeSet, 4> VarArgsAttrs;
1864   for (unsigned i = CalledFunc->getFunctionType()->getNumParams();
1865        i < CS.getNumArgOperands(); i++) {
1866     VarArgsToForward.push_back(CS.getArgOperand(i));
1867     VarArgsAttrs.push_back(CS.getAttributes().getParamAttributes(i));
1868   }
1869 
1870   bool InlinedMustTailCalls = false, InlinedDeoptimizeCalls = false;
1871   if (InlinedFunctionInfo.ContainsCalls) {
1872     CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None;
1873     if (CallInst *CI = dyn_cast<CallInst>(TheCall))
1874       CallSiteTailKind = CI->getTailCallKind();
1875 
1876     // For inlining purposes, the "notail" marker is the same as no marker.
1877     if (CallSiteTailKind == CallInst::TCK_NoTail)
1878       CallSiteTailKind = CallInst::TCK_None;
1879 
1880     for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E;
1881          ++BB) {
1882       for (auto II = BB->begin(); II != BB->end();) {
1883         Instruction &I = *II++;
1884         CallInst *CI = dyn_cast<CallInst>(&I);
1885         if (!CI)
1886           continue;
1887 
1888         // Forward varargs from inlined call site to calls to the
1889         // ForwardVarArgsTo function, if requested, and to musttail calls.
1890         if (!VarArgsToForward.empty() &&
1891             ((ForwardVarArgsTo &&
1892               CI->getCalledFunction() == ForwardVarArgsTo) ||
1893              CI->isMustTailCall())) {
1894           // Collect attributes for non-vararg parameters.
1895           AttributeList Attrs = CI->getAttributes();
1896           SmallVector<AttributeSet, 8> ArgAttrs;
1897           if (!Attrs.isEmpty() || !VarArgsAttrs.empty()) {
1898             for (unsigned ArgNo = 0;
1899                  ArgNo < CI->getFunctionType()->getNumParams(); ++ArgNo)
1900               ArgAttrs.push_back(Attrs.getParamAttributes(ArgNo));
1901           }
1902 
1903           // Add VarArg attributes.
1904           ArgAttrs.append(VarArgsAttrs.begin(), VarArgsAttrs.end());
1905           Attrs = AttributeList::get(CI->getContext(), Attrs.getFnAttributes(),
1906                                      Attrs.getRetAttributes(), ArgAttrs);
1907           // Add VarArgs to existing parameters.
1908           SmallVector<Value *, 6> Params(CI->arg_operands());
1909           Params.append(VarArgsToForward.begin(), VarArgsToForward.end());
1910           CallInst *NewCI = CallInst::Create(
1911               CI->getFunctionType(), CI->getCalledOperand(), Params, "", CI);
1912           NewCI->setDebugLoc(CI->getDebugLoc());
1913           NewCI->setAttributes(Attrs);
1914           NewCI->setCallingConv(CI->getCallingConv());
1915           CI->replaceAllUsesWith(NewCI);
1916           CI->eraseFromParent();
1917           CI = NewCI;
1918         }
1919 
1920         if (Function *F = CI->getCalledFunction())
1921           InlinedDeoptimizeCalls |=
1922               F->getIntrinsicID() == Intrinsic::experimental_deoptimize;
1923 
1924         // We need to reduce the strength of any inlined tail calls.  For
1925         // musttail, we have to avoid introducing potential unbounded stack
1926         // growth.  For example, if functions 'f' and 'g' are mutually recursive
1927         // with musttail, we can inline 'g' into 'f' so long as we preserve
1928         // musttail on the cloned call to 'f'.  If either the inlined call site
1929         // or the cloned call site is *not* musttail, the program already has
1930         // one frame of stack growth, so it's safe to remove musttail.  Here is
1931         // a table of example transformations:
1932         //
1933         //    f -> musttail g -> musttail f  ==>  f -> musttail f
1934         //    f -> musttail g ->     tail f  ==>  f ->     tail f
1935         //    f ->          g -> musttail f  ==>  f ->          f
1936         //    f ->          g ->     tail f  ==>  f ->          f
1937         //
1938         // Inlined notail calls should remain notail calls.
1939         CallInst::TailCallKind ChildTCK = CI->getTailCallKind();
1940         if (ChildTCK != CallInst::TCK_NoTail)
1941           ChildTCK = std::min(CallSiteTailKind, ChildTCK);
1942         CI->setTailCallKind(ChildTCK);
1943         InlinedMustTailCalls |= CI->isMustTailCall();
1944 
1945         // Calls inlined through a 'nounwind' call site should be marked
1946         // 'nounwind'.
1947         if (MarkNoUnwind)
1948           CI->setDoesNotThrow();
1949       }
1950     }
1951   }
1952 
1953   // Leave lifetime markers for the static alloca's, scoping them to the
1954   // function we just inlined.
1955   if (InsertLifetime && !IFI.StaticAllocas.empty()) {
1956     IRBuilder<> builder(&FirstNewBlock->front());
1957     for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) {
1958       AllocaInst *AI = IFI.StaticAllocas[ai];
1959       // Don't mark swifterror allocas. They can't have bitcast uses.
1960       if (AI->isSwiftError())
1961         continue;
1962 
1963       // If the alloca is already scoped to something smaller than the whole
1964       // function then there's no need to add redundant, less accurate markers.
1965       if (hasLifetimeMarkers(AI))
1966         continue;
1967 
1968       // Try to determine the size of the allocation.
1969       ConstantInt *AllocaSize = nullptr;
1970       if (ConstantInt *AIArraySize =
1971           dyn_cast<ConstantInt>(AI->getArraySize())) {
1972         auto &DL = Caller->getParent()->getDataLayout();
1973         Type *AllocaType = AI->getAllocatedType();
1974         uint64_t AllocaTypeSize = DL.getTypeAllocSize(AllocaType);
1975         uint64_t AllocaArraySize = AIArraySize->getLimitedValue();
1976 
1977         // Don't add markers for zero-sized allocas.
1978         if (AllocaArraySize == 0)
1979           continue;
1980 
1981         // Check that array size doesn't saturate uint64_t and doesn't
1982         // overflow when it's multiplied by type size.
1983         if (AllocaArraySize != std::numeric_limits<uint64_t>::max() &&
1984             std::numeric_limits<uint64_t>::max() / AllocaArraySize >=
1985                 AllocaTypeSize) {
1986           AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()),
1987                                         AllocaArraySize * AllocaTypeSize);
1988         }
1989       }
1990 
1991       builder.CreateLifetimeStart(AI, AllocaSize);
1992       for (ReturnInst *RI : Returns) {
1993         // Don't insert llvm.lifetime.end calls between a musttail or deoptimize
1994         // call and a return.  The return kills all local allocas.
1995         if (InlinedMustTailCalls &&
1996             RI->getParent()->getTerminatingMustTailCall())
1997           continue;
1998         if (InlinedDeoptimizeCalls &&
1999             RI->getParent()->getTerminatingDeoptimizeCall())
2000           continue;
2001         IRBuilder<>(RI).CreateLifetimeEnd(AI, AllocaSize);
2002       }
2003     }
2004   }
2005 
2006   // If the inlined code contained dynamic alloca instructions, wrap the inlined
2007   // code with llvm.stacksave/llvm.stackrestore intrinsics.
2008   if (InlinedFunctionInfo.ContainsDynamicAllocas) {
2009     Module *M = Caller->getParent();
2010     // Get the two intrinsics we care about.
2011     Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
2012     Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore);
2013 
2014     // Insert the llvm.stacksave.
2015     CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin())
2016                              .CreateCall(StackSave, {}, "savedstack");
2017 
2018     // Insert a call to llvm.stackrestore before any return instructions in the
2019     // inlined function.
2020     for (ReturnInst *RI : Returns) {
2021       // Don't insert llvm.stackrestore calls between a musttail or deoptimize
2022       // call and a return.  The return will restore the stack pointer.
2023       if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall())
2024         continue;
2025       if (InlinedDeoptimizeCalls && RI->getParent()->getTerminatingDeoptimizeCall())
2026         continue;
2027       IRBuilder<>(RI).CreateCall(StackRestore, SavedPtr);
2028     }
2029   }
2030 
2031   // If we are inlining for an invoke instruction, we must make sure to rewrite
2032   // any call instructions into invoke instructions.  This is sensitive to which
2033   // funclet pads were top-level in the inlinee, so must be done before
2034   // rewriting the "parent pad" links.
2035   if (auto *II = dyn_cast<InvokeInst>(TheCall)) {
2036     BasicBlock *UnwindDest = II->getUnwindDest();
2037     Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI();
2038     if (isa<LandingPadInst>(FirstNonPHI)) {
2039       HandleInlinedLandingPad(II, &*FirstNewBlock, InlinedFunctionInfo);
2040     } else {
2041       HandleInlinedEHPad(II, &*FirstNewBlock, InlinedFunctionInfo);
2042     }
2043   }
2044 
2045   // Update the lexical scopes of the new funclets and callsites.
2046   // Anything that had 'none' as its parent is now nested inside the callsite's
2047   // EHPad.
2048 
2049   if (CallSiteEHPad) {
2050     for (Function::iterator BB = FirstNewBlock->getIterator(),
2051                             E = Caller->end();
2052          BB != E; ++BB) {
2053       // Add bundle operands to any top-level call sites.
2054       SmallVector<OperandBundleDef, 1> OpBundles;
2055       for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;) {
2056         Instruction *I = &*BBI++;
2057         CallSite CS(I);
2058         if (!CS)
2059           continue;
2060 
2061         // Skip call sites which are nounwind intrinsics.
2062         auto *CalledFn =
2063             dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts());
2064         if (CalledFn && CalledFn->isIntrinsic() && CS.doesNotThrow())
2065           continue;
2066 
2067         // Skip call sites which already have a "funclet" bundle.
2068         if (CS.getOperandBundle(LLVMContext::OB_funclet))
2069           continue;
2070 
2071         CS.getOperandBundlesAsDefs(OpBundles);
2072         OpBundles.emplace_back("funclet", CallSiteEHPad);
2073 
2074         Instruction *NewInst;
2075         if (CS.isCall())
2076           NewInst = CallInst::Create(cast<CallInst>(I), OpBundles, I);
2077         else if (CS.isCallBr())
2078           NewInst = CallBrInst::Create(cast<CallBrInst>(I), OpBundles, I);
2079         else
2080           NewInst = InvokeInst::Create(cast<InvokeInst>(I), OpBundles, I);
2081         NewInst->takeName(I);
2082         I->replaceAllUsesWith(NewInst);
2083         I->eraseFromParent();
2084 
2085         OpBundles.clear();
2086       }
2087 
2088       // It is problematic if the inlinee has a cleanupret which unwinds to
2089       // caller and we inline it into a call site which doesn't unwind but into
2090       // an EH pad that does.  Such an edge must be dynamically unreachable.
2091       // As such, we replace the cleanupret with unreachable.
2092       if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(BB->getTerminator()))
2093         if (CleanupRet->unwindsToCaller() && EHPadForCallUnwindsLocally)
2094           changeToUnreachable(CleanupRet, /*UseLLVMTrap=*/false);
2095 
2096       Instruction *I = BB->getFirstNonPHI();
2097       if (!I->isEHPad())
2098         continue;
2099 
2100       if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) {
2101         if (isa<ConstantTokenNone>(CatchSwitch->getParentPad()))
2102           CatchSwitch->setParentPad(CallSiteEHPad);
2103       } else {
2104         auto *FPI = cast<FuncletPadInst>(I);
2105         if (isa<ConstantTokenNone>(FPI->getParentPad()))
2106           FPI->setParentPad(CallSiteEHPad);
2107       }
2108     }
2109   }
2110 
2111   if (InlinedDeoptimizeCalls) {
2112     // We need to at least remove the deoptimizing returns from the Return set,
2113     // so that the control flow from those returns does not get merged into the
2114     // caller (but terminate it instead).  If the caller's return type does not
2115     // match the callee's return type, we also need to change the return type of
2116     // the intrinsic.
2117     if (Caller->getReturnType() == TheCall->getType()) {
2118       auto NewEnd = llvm::remove_if(Returns, [](ReturnInst *RI) {
2119         return RI->getParent()->getTerminatingDeoptimizeCall() != nullptr;
2120       });
2121       Returns.erase(NewEnd, Returns.end());
2122     } else {
2123       SmallVector<ReturnInst *, 8> NormalReturns;
2124       Function *NewDeoptIntrinsic = Intrinsic::getDeclaration(
2125           Caller->getParent(), Intrinsic::experimental_deoptimize,
2126           {Caller->getReturnType()});
2127 
2128       for (ReturnInst *RI : Returns) {
2129         CallInst *DeoptCall = RI->getParent()->getTerminatingDeoptimizeCall();
2130         if (!DeoptCall) {
2131           NormalReturns.push_back(RI);
2132           continue;
2133         }
2134 
2135         // The calling convention on the deoptimize call itself may be bogus,
2136         // since the code we're inlining may have undefined behavior (and may
2137         // never actually execute at runtime); but all
2138         // @llvm.experimental.deoptimize declarations have to have the same
2139         // calling convention in a well-formed module.
2140         auto CallingConv = DeoptCall->getCalledFunction()->getCallingConv();
2141         NewDeoptIntrinsic->setCallingConv(CallingConv);
2142         auto *CurBB = RI->getParent();
2143         RI->eraseFromParent();
2144 
2145         SmallVector<Value *, 4> CallArgs(DeoptCall->arg_begin(),
2146                                          DeoptCall->arg_end());
2147 
2148         SmallVector<OperandBundleDef, 1> OpBundles;
2149         DeoptCall->getOperandBundlesAsDefs(OpBundles);
2150         DeoptCall->eraseFromParent();
2151         assert(!OpBundles.empty() &&
2152                "Expected at least the deopt operand bundle");
2153 
2154         IRBuilder<> Builder(CurBB);
2155         CallInst *NewDeoptCall =
2156             Builder.CreateCall(NewDeoptIntrinsic, CallArgs, OpBundles);
2157         NewDeoptCall->setCallingConv(CallingConv);
2158         if (NewDeoptCall->getType()->isVoidTy())
2159           Builder.CreateRetVoid();
2160         else
2161           Builder.CreateRet(NewDeoptCall);
2162       }
2163 
2164       // Leave behind the normal returns so we can merge control flow.
2165       std::swap(Returns, NormalReturns);
2166     }
2167   }
2168 
2169   // Handle any inlined musttail call sites.  In order for a new call site to be
2170   // musttail, the source of the clone and the inlined call site must have been
2171   // musttail.  Therefore it's safe to return without merging control into the
2172   // phi below.
2173   if (InlinedMustTailCalls) {
2174     // Check if we need to bitcast the result of any musttail calls.
2175     Type *NewRetTy = Caller->getReturnType();
2176     bool NeedBitCast = !TheCall->use_empty() && TheCall->getType() != NewRetTy;
2177 
2178     // Handle the returns preceded by musttail calls separately.
2179     SmallVector<ReturnInst *, 8> NormalReturns;
2180     for (ReturnInst *RI : Returns) {
2181       CallInst *ReturnedMustTail =
2182           RI->getParent()->getTerminatingMustTailCall();
2183       if (!ReturnedMustTail) {
2184         NormalReturns.push_back(RI);
2185         continue;
2186       }
2187       if (!NeedBitCast)
2188         continue;
2189 
2190       // Delete the old return and any preceding bitcast.
2191       BasicBlock *CurBB = RI->getParent();
2192       auto *OldCast = dyn_cast_or_null<BitCastInst>(RI->getReturnValue());
2193       RI->eraseFromParent();
2194       if (OldCast)
2195         OldCast->eraseFromParent();
2196 
2197       // Insert a new bitcast and return with the right type.
2198       IRBuilder<> Builder(CurBB);
2199       Builder.CreateRet(Builder.CreateBitCast(ReturnedMustTail, NewRetTy));
2200     }
2201 
2202     // Leave behind the normal returns so we can merge control flow.
2203     std::swap(Returns, NormalReturns);
2204   }
2205 
2206   // Now that all of the transforms on the inlined code have taken place but
2207   // before we splice the inlined code into the CFG and lose track of which
2208   // blocks were actually inlined, collect the call sites. We only do this if
2209   // call graph updates weren't requested, as those provide value handle based
2210   // tracking of inlined call sites instead.
2211   if (InlinedFunctionInfo.ContainsCalls && !IFI.CG) {
2212     // Otherwise just collect the raw call sites that were inlined.
2213     for (BasicBlock &NewBB :
2214          make_range(FirstNewBlock->getIterator(), Caller->end()))
2215       for (Instruction &I : NewBB)
2216         if (auto CS = CallSite(&I))
2217           IFI.InlinedCallSites.push_back(CS);
2218   }
2219 
2220   // If we cloned in _exactly one_ basic block, and if that block ends in a
2221   // return instruction, we splice the body of the inlined callee directly into
2222   // the calling basic block.
2223   if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
2224     // Move all of the instructions right before the call.
2225     OrigBB->getInstList().splice(TheCall->getIterator(),
2226                                  FirstNewBlock->getInstList(),
2227                                  FirstNewBlock->begin(), FirstNewBlock->end());
2228     // Remove the cloned basic block.
2229     Caller->getBasicBlockList().pop_back();
2230 
2231     // If the call site was an invoke instruction, add a branch to the normal
2232     // destination.
2233     if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
2234       BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
2235       NewBr->setDebugLoc(Returns[0]->getDebugLoc());
2236     }
2237 
2238     // If the return instruction returned a value, replace uses of the call with
2239     // uses of the returned value.
2240     if (!TheCall->use_empty()) {
2241       ReturnInst *R = Returns[0];
2242       if (TheCall == R->getReturnValue())
2243         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
2244       else
2245         TheCall->replaceAllUsesWith(R->getReturnValue());
2246     }
2247     // Since we are now done with the Call/Invoke, we can delete it.
2248     TheCall->eraseFromParent();
2249 
2250     // Since we are now done with the return instruction, delete it also.
2251     Returns[0]->eraseFromParent();
2252 
2253     // We are now done with the inlining.
2254     return true;
2255   }
2256 
2257   // Otherwise, we have the normal case, of more than one block to inline or
2258   // multiple return sites.
2259 
2260   // We want to clone the entire callee function into the hole between the
2261   // "starter" and "ender" blocks.  How we accomplish this depends on whether
2262   // this is an invoke instruction or a call instruction.
2263   BasicBlock *AfterCallBB;
2264   BranchInst *CreatedBranchToNormalDest = nullptr;
2265   if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
2266 
2267     // Add an unconditional branch to make this look like the CallInst case...
2268     CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall);
2269 
2270     // Split the basic block.  This guarantees that no PHI nodes will have to be
2271     // updated due to new incoming edges, and make the invoke case more
2272     // symmetric to the call case.
2273     AfterCallBB =
2274         OrigBB->splitBasicBlock(CreatedBranchToNormalDest->getIterator(),
2275                                 CalledFunc->getName() + ".exit");
2276 
2277   } else {  // It's a call
2278     // If this is a call instruction, we need to split the basic block that
2279     // the call lives in.
2280     //
2281     AfterCallBB = OrigBB->splitBasicBlock(TheCall->getIterator(),
2282                                           CalledFunc->getName() + ".exit");
2283   }
2284 
2285   if (IFI.CallerBFI) {
2286     // Copy original BB's block frequency to AfterCallBB
2287     IFI.CallerBFI->setBlockFreq(
2288         AfterCallBB, IFI.CallerBFI->getBlockFreq(OrigBB).getFrequency());
2289   }
2290 
2291   // Change the branch that used to go to AfterCallBB to branch to the first
2292   // basic block of the inlined function.
2293   //
2294   Instruction *Br = OrigBB->getTerminator();
2295   assert(Br && Br->getOpcode() == Instruction::Br &&
2296          "splitBasicBlock broken!");
2297   Br->setOperand(0, &*FirstNewBlock);
2298 
2299   // Now that the function is correct, make it a little bit nicer.  In
2300   // particular, move the basic blocks inserted from the end of the function
2301   // into the space made by splitting the source basic block.
2302   Caller->getBasicBlockList().splice(AfterCallBB->getIterator(),
2303                                      Caller->getBasicBlockList(), FirstNewBlock,
2304                                      Caller->end());
2305 
2306   // Handle all of the return instructions that we just cloned in, and eliminate
2307   // any users of the original call/invoke instruction.
2308   Type *RTy = CalledFunc->getReturnType();
2309 
2310   PHINode *PHI = nullptr;
2311   if (Returns.size() > 1) {
2312     // The PHI node should go at the front of the new basic block to merge all
2313     // possible incoming values.
2314     if (!TheCall->use_empty()) {
2315       PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(),
2316                             &AfterCallBB->front());
2317       // Anything that used the result of the function call should now use the
2318       // PHI node as their operand.
2319       TheCall->replaceAllUsesWith(PHI);
2320     }
2321 
2322     // Loop over all of the return instructions adding entries to the PHI node
2323     // as appropriate.
2324     if (PHI) {
2325       for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
2326         ReturnInst *RI = Returns[i];
2327         assert(RI->getReturnValue()->getType() == PHI->getType() &&
2328                "Ret value not consistent in function!");
2329         PHI->addIncoming(RI->getReturnValue(), RI->getParent());
2330       }
2331     }
2332 
2333     // Add a branch to the merge points and remove return instructions.
2334     DebugLoc Loc;
2335     for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
2336       ReturnInst *RI = Returns[i];
2337       BranchInst* BI = BranchInst::Create(AfterCallBB, RI);
2338       Loc = RI->getDebugLoc();
2339       BI->setDebugLoc(Loc);
2340       RI->eraseFromParent();
2341     }
2342     // We need to set the debug location to *somewhere* inside the
2343     // inlined function. The line number may be nonsensical, but the
2344     // instruction will at least be associated with the right
2345     // function.
2346     if (CreatedBranchToNormalDest)
2347       CreatedBranchToNormalDest->setDebugLoc(Loc);
2348   } else if (!Returns.empty()) {
2349     // Otherwise, if there is exactly one return value, just replace anything
2350     // using the return value of the call with the computed value.
2351     if (!TheCall->use_empty()) {
2352       if (TheCall == Returns[0]->getReturnValue())
2353         TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
2354       else
2355         TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
2356     }
2357 
2358     // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
2359     BasicBlock *ReturnBB = Returns[0]->getParent();
2360     ReturnBB->replaceAllUsesWith(AfterCallBB);
2361 
2362     // Splice the code from the return block into the block that it will return
2363     // to, which contains the code that was after the call.
2364     AfterCallBB->getInstList().splice(AfterCallBB->begin(),
2365                                       ReturnBB->getInstList());
2366 
2367     if (CreatedBranchToNormalDest)
2368       CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc());
2369 
2370     // Delete the return instruction now and empty ReturnBB now.
2371     Returns[0]->eraseFromParent();
2372     ReturnBB->eraseFromParent();
2373   } else if (!TheCall->use_empty()) {
2374     // No returns, but something is using the return value of the call.  Just
2375     // nuke the result.
2376     TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
2377   }
2378 
2379   // Since we are now done with the Call/Invoke, we can delete it.
2380   TheCall->eraseFromParent();
2381 
2382   // If we inlined any musttail calls and the original return is now
2383   // unreachable, delete it.  It can only contain a bitcast and ret.
2384   if (InlinedMustTailCalls && pred_begin(AfterCallBB) == pred_end(AfterCallBB))
2385     AfterCallBB->eraseFromParent();
2386 
2387   // We should always be able to fold the entry block of the function into the
2388   // single predecessor of the block...
2389   assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
2390   BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
2391 
2392   // Splice the code entry block into calling block, right before the
2393   // unconditional branch.
2394   CalleeEntry->replaceAllUsesWith(OrigBB);  // Update PHI nodes
2395   OrigBB->getInstList().splice(Br->getIterator(), CalleeEntry->getInstList());
2396 
2397   // Remove the unconditional branch.
2398   OrigBB->getInstList().erase(Br);
2399 
2400   // Now we can remove the CalleeEntry block, which is now empty.
2401   Caller->getBasicBlockList().erase(CalleeEntry);
2402 
2403   // If we inserted a phi node, check to see if it has a single value (e.g. all
2404   // the entries are the same or undef).  If so, remove the PHI so it doesn't
2405   // block other optimizations.
2406   if (PHI) {
2407     AssumptionCache *AC =
2408         IFI.GetAssumptionCache ? &(*IFI.GetAssumptionCache)(*Caller) : nullptr;
2409     auto &DL = Caller->getParent()->getDataLayout();
2410     if (Value *V = SimplifyInstruction(PHI, {DL, nullptr, nullptr, AC})) {
2411       PHI->replaceAllUsesWith(V);
2412       PHI->eraseFromParent();
2413     }
2414   }
2415 
2416   return true;
2417 }
2418