1 //===-- IndirectCallPromotion.cpp - Optimizations based on value profiling ===//
2 //
3 //                      The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the transformation that promotes indirect calls to
11 // conditional direct calls when the indirect-call value profile metadata is
12 // available.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/ADT/StringRef.h"
19 #include "llvm/ADT/Twine.h"
20 #include "llvm/Analysis/BlockFrequencyInfo.h"
21 #include "llvm/Analysis/GlobalsModRef.h"
22 #include "llvm/Analysis/IndirectCallPromotionAnalysis.h"
23 #include "llvm/Analysis/IndirectCallSiteVisitor.h"
24 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
25 #include "llvm/Analysis/ProfileSummaryInfo.h"
26 #include "llvm/IR/BasicBlock.h"
27 #include "llvm/IR/CallSite.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/DiagnosticInfo.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/InstrTypes.h"
33 #include "llvm/IR/Instruction.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/MDBuilder.h"
37 #include "llvm/IR/PassManager.h"
38 #include "llvm/IR/Type.h"
39 #include "llvm/Pass.h"
40 #include "llvm/PassRegistry.h"
41 #include "llvm/PassSupport.h"
42 #include "llvm/ProfileData/InstrProf.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Transforms/Instrumentation.h"
49 #include "llvm/Transforms/PGOInstrumentation.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include <cassert>
52 #include <cstdint>
53 #include <vector>
54 
55 using namespace llvm;
56 
57 #define DEBUG_TYPE "pgo-icall-prom"
58 
59 STATISTIC(NumOfPGOICallPromotion, "Number of indirect call promotions.");
60 STATISTIC(NumOfPGOICallsites, "Number of indirect call candidate sites.");
61 
62 // Command line option to disable indirect-call promotion with the default as
63 // false. This is for debug purpose.
64 static cl::opt<bool> DisableICP("disable-icp", cl::init(false), cl::Hidden,
65                                 cl::desc("Disable indirect call promotion"));
66 
67 // Set the cutoff value for the promotion. If the value is other than 0, we
68 // stop the transformation once the total number of promotions equals the cutoff
69 // value.
70 // For debug use only.
71 static cl::opt<unsigned>
72     ICPCutOff("icp-cutoff", cl::init(0), cl::Hidden, cl::ZeroOrMore,
73               cl::desc("Max number of promotions for this compilation"));
74 
75 // If ICPCSSkip is non zero, the first ICPCSSkip callsites will be skipped.
76 // For debug use only.
77 static cl::opt<unsigned>
78     ICPCSSkip("icp-csskip", cl::init(0), cl::Hidden, cl::ZeroOrMore,
79               cl::desc("Skip Callsite up to this number for this compilation"));
80 
81 // Set if the pass is called in LTO optimization. The difference for LTO mode
82 // is the pass won't prefix the source module name to the internal linkage
83 // symbols.
84 static cl::opt<bool> ICPLTOMode("icp-lto", cl::init(false), cl::Hidden,
85                                 cl::desc("Run indirect-call promotion in LTO "
86                                          "mode"));
87 
88 // Set if the pass is called in SamplePGO mode. The difference for SamplePGO
89 // mode is it will add prof metadatato the created direct call.
90 static cl::opt<bool>
91     ICPSamplePGOMode("icp-samplepgo", cl::init(false), cl::Hidden,
92                      cl::desc("Run indirect-call promotion in SamplePGO mode"));
93 
94 // If the option is set to true, only call instructions will be considered for
95 // transformation -- invoke instructions will be ignored.
96 static cl::opt<bool>
97     ICPCallOnly("icp-call-only", cl::init(false), cl::Hidden,
98                 cl::desc("Run indirect-call promotion for call instructions "
99                          "only"));
100 
101 // If the option is set to true, only invoke instructions will be considered for
102 // transformation -- call instructions will be ignored.
103 static cl::opt<bool> ICPInvokeOnly("icp-invoke-only", cl::init(false),
104                                    cl::Hidden,
105                                    cl::desc("Run indirect-call promotion for "
106                                             "invoke instruction only"));
107 
108 // Dump the function level IR if the transformation happened in this
109 // function. For debug use only.
110 static cl::opt<bool>
111     ICPDUMPAFTER("icp-dumpafter", cl::init(false), cl::Hidden,
112                  cl::desc("Dump IR after transformation happens"));
113 
114 namespace {
115 class PGOIndirectCallPromotionLegacyPass : public ModulePass {
116 public:
117   static char ID;
118 
119   PGOIndirectCallPromotionLegacyPass(bool InLTO = false, bool SamplePGO = false)
120       : ModulePass(ID), InLTO(InLTO), SamplePGO(SamplePGO) {
121     initializePGOIndirectCallPromotionLegacyPassPass(
122         *PassRegistry::getPassRegistry());
123   }
124 
125   void getAnalysisUsage(AnalysisUsage &AU) const override {
126     AU.addRequired<ProfileSummaryInfoWrapperPass>();
127   }
128 
129   StringRef getPassName() const override { return "PGOIndirectCallPromotion"; }
130 
131 private:
132   bool runOnModule(Module &M) override;
133 
134   // If this pass is called in LTO. We need to special handling the PGOFuncName
135   // for the static variables due to LTO's internalization.
136   bool InLTO;
137 
138   // If this pass is called in SamplePGO. We need to add the prof metadata to
139   // the promoted direct call.
140   bool SamplePGO;
141 };
142 } // end anonymous namespace
143 
144 char PGOIndirectCallPromotionLegacyPass::ID = 0;
145 INITIALIZE_PASS_BEGIN(PGOIndirectCallPromotionLegacyPass, "pgo-icall-prom",
146                       "Use PGO instrumentation profile to promote indirect "
147                       "calls to direct calls.",
148                       false, false)
149 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
150 INITIALIZE_PASS_END(PGOIndirectCallPromotionLegacyPass, "pgo-icall-prom",
151                     "Use PGO instrumentation profile to promote indirect "
152                     "calls to direct calls.",
153                     false, false)
154 
155 ModulePass *llvm::createPGOIndirectCallPromotionLegacyPass(bool InLTO,
156                                                            bool SamplePGO) {
157   return new PGOIndirectCallPromotionLegacyPass(InLTO, SamplePGO);
158 }
159 
160 namespace {
161 // The class for main data structure to promote indirect calls to conditional
162 // direct calls.
163 class ICallPromotionFunc {
164 private:
165   Function &F;
166   Module *M;
167 
168   // Symtab that maps indirect call profile values to function names and
169   // defines.
170   InstrProfSymtab *Symtab;
171 
172   bool SamplePGO;
173 
174   OptimizationRemarkEmitter &ORE;
175 
176   // A struct that records the direct target and it's call count.
177   struct PromotionCandidate {
178     Function *TargetFunction;
179     uint64_t Count;
180     PromotionCandidate(Function *F, uint64_t C) : TargetFunction(F), Count(C) {}
181   };
182 
183   // Check if the indirect-call call site should be promoted. Return the number
184   // of promotions. Inst is the candidate indirect call, ValueDataRef
185   // contains the array of value profile data for profiled targets,
186   // TotalCount is the total profiled count of call executions, and
187   // NumCandidates is the number of candidate entries in ValueDataRef.
188   std::vector<PromotionCandidate> getPromotionCandidatesForCallSite(
189       Instruction *Inst, const ArrayRef<InstrProfValueData> &ValueDataRef,
190       uint64_t TotalCount, uint32_t NumCandidates);
191 
192   // Promote a list of targets for one indirect-call callsite. Return
193   // the number of promotions.
194   uint32_t tryToPromote(Instruction *Inst,
195                         const std::vector<PromotionCandidate> &Candidates,
196                         uint64_t &TotalCount);
197 
198   // Noncopyable
199   ICallPromotionFunc(const ICallPromotionFunc &other) = delete;
200   ICallPromotionFunc &operator=(const ICallPromotionFunc &other) = delete;
201 
202 public:
203   ICallPromotionFunc(Function &Func, Module *Modu, InstrProfSymtab *Symtab,
204                      bool SamplePGO, OptimizationRemarkEmitter &ORE)
205       : F(Func), M(Modu), Symtab(Symtab), SamplePGO(SamplePGO), ORE(ORE) {}
206 
207   bool processFunction(ProfileSummaryInfo *PSI);
208 };
209 } // end anonymous namespace
210 
211 bool llvm::isLegalToPromote(Instruction *Inst, Function *F,
212                             const char **Reason) {
213   // Check the return type.
214   Type *CallRetType = Inst->getType();
215   if (!CallRetType->isVoidTy()) {
216     Type *FuncRetType = F->getReturnType();
217     if (FuncRetType != CallRetType &&
218         !CastInst::isBitCastable(FuncRetType, CallRetType)) {
219       if (Reason)
220         *Reason = "Return type mismatch";
221       return false;
222     }
223   }
224 
225   // Check if the arguments are compatible with the parameters
226   FunctionType *DirectCalleeType = F->getFunctionType();
227   unsigned ParamNum = DirectCalleeType->getFunctionNumParams();
228   CallSite CS(Inst);
229   unsigned ArgNum = CS.arg_size();
230 
231   if (ParamNum != ArgNum && !DirectCalleeType->isVarArg()) {
232     if (Reason)
233       *Reason = "The number of arguments mismatch";
234     return false;
235   }
236 
237   for (unsigned I = 0; I < ParamNum; ++I) {
238     Type *PTy = DirectCalleeType->getFunctionParamType(I);
239     Type *ATy = CS.getArgument(I)->getType();
240     if (PTy == ATy)
241       continue;
242     if (!CastInst::castIsValid(Instruction::BitCast, CS.getArgument(I), PTy)) {
243       if (Reason)
244         *Reason = "Argument type mismatch";
245       return false;
246     }
247   }
248 
249   DEBUG(dbgs() << " #" << NumOfPGOICallPromotion << " Promote the icall to "
250                << F->getName() << "\n");
251   return true;
252 }
253 
254 // Indirect-call promotion heuristic. The direct targets are sorted based on
255 // the count. Stop at the first target that is not promoted.
256 std::vector<ICallPromotionFunc::PromotionCandidate>
257 ICallPromotionFunc::getPromotionCandidatesForCallSite(
258     Instruction *Inst, const ArrayRef<InstrProfValueData> &ValueDataRef,
259     uint64_t TotalCount, uint32_t NumCandidates) {
260   std::vector<PromotionCandidate> Ret;
261 
262   DEBUG(dbgs() << " \nWork on callsite #" << NumOfPGOICallsites << *Inst
263                << " Num_targets: " << ValueDataRef.size()
264                << " Num_candidates: " << NumCandidates << "\n");
265   NumOfPGOICallsites++;
266   if (ICPCSSkip != 0 && NumOfPGOICallsites <= ICPCSSkip) {
267     DEBUG(dbgs() << " Skip: User options.\n");
268     return Ret;
269   }
270 
271   for (uint32_t I = 0; I < NumCandidates; I++) {
272     uint64_t Count = ValueDataRef[I].Count;
273     assert(Count <= TotalCount);
274     uint64_t Target = ValueDataRef[I].Value;
275     DEBUG(dbgs() << " Candidate " << I << " Count=" << Count
276                  << "  Target_func: " << Target << "\n");
277 
278     if (ICPInvokeOnly && dyn_cast<CallInst>(Inst)) {
279       DEBUG(dbgs() << " Not promote: User options.\n");
280       ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "UserOptions", Inst)
281                << " Not promote: User options");
282       break;
283     }
284     if (ICPCallOnly && dyn_cast<InvokeInst>(Inst)) {
285       DEBUG(dbgs() << " Not promote: User option.\n");
286       ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "UserOptions", Inst)
287                << " Not promote: User options");
288       break;
289     }
290     if (ICPCutOff != 0 && NumOfPGOICallPromotion >= ICPCutOff) {
291       DEBUG(dbgs() << " Not promote: Cutoff reached.\n");
292       ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "CutOffReached", Inst)
293                << " Not promote: Cutoff reached");
294       break;
295     }
296 
297     Function *TargetFunction = Symtab->getFunction(Target);
298     if (TargetFunction == nullptr) {
299       DEBUG(dbgs() << " Not promote: Cannot find the target\n");
300       ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "UnableToFindTarget", Inst)
301                << "Cannot promote indirect call: target not found");
302       break;
303     }
304 
305     const char *Reason = nullptr;
306     if (!isLegalToPromote(Inst, TargetFunction, &Reason)) {
307       using namespace ore;
308       ORE.emit(OptimizationRemarkMissed(DEBUG_TYPE, "UnableToPromote", Inst)
309                << "Cannot promote indirect call to "
310                << NV("TargetFunction", TargetFunction) << " with count of "
311                << NV("Count", Count) << ": " << Reason);
312       break;
313     }
314 
315     Ret.push_back(PromotionCandidate(TargetFunction, Count));
316     TotalCount -= Count;
317   }
318   return Ret;
319 }
320 
321 // Create a diamond structure for If_Then_Else. Also update the profile
322 // count. Do the fix-up for the invoke instruction.
323 static void createIfThenElse(Instruction *Inst, Function *DirectCallee,
324                              uint64_t Count, uint64_t TotalCount,
325                              BasicBlock **DirectCallBB,
326                              BasicBlock **IndirectCallBB,
327                              BasicBlock **MergeBB) {
328   CallSite CS(Inst);
329   Value *OrigCallee = CS.getCalledValue();
330 
331   IRBuilder<> BBBuilder(Inst);
332   LLVMContext &Ctx = Inst->getContext();
333   Value *BCI1 =
334       BBBuilder.CreateBitCast(OrigCallee, Type::getInt8PtrTy(Ctx), "");
335   Value *BCI2 =
336       BBBuilder.CreateBitCast(DirectCallee, Type::getInt8PtrTy(Ctx), "");
337   Value *PtrCmp = BBBuilder.CreateICmpEQ(BCI1, BCI2, "");
338 
339   uint64_t ElseCount = TotalCount - Count;
340   uint64_t MaxCount = (Count >= ElseCount ? Count : ElseCount);
341   uint64_t Scale = calculateCountScale(MaxCount);
342   MDBuilder MDB(Inst->getContext());
343   MDNode *BranchWeights = MDB.createBranchWeights(
344       scaleBranchCount(Count, Scale), scaleBranchCount(ElseCount, Scale));
345   TerminatorInst *ThenTerm, *ElseTerm;
346   SplitBlockAndInsertIfThenElse(PtrCmp, Inst, &ThenTerm, &ElseTerm,
347                                 BranchWeights);
348   *DirectCallBB = ThenTerm->getParent();
349   (*DirectCallBB)->setName("if.true.direct_targ");
350   *IndirectCallBB = ElseTerm->getParent();
351   (*IndirectCallBB)->setName("if.false.orig_indirect");
352   *MergeBB = Inst->getParent();
353   (*MergeBB)->setName("if.end.icp");
354 
355   // Special handing of Invoke instructions.
356   InvokeInst *II = dyn_cast<InvokeInst>(Inst);
357   if (!II)
358     return;
359 
360   // We don't need branch instructions for invoke.
361   ThenTerm->eraseFromParent();
362   ElseTerm->eraseFromParent();
363 
364   // Add jump from Merge BB to the NormalDest. This is needed for the newly
365   // created direct invoke stmt -- as its NormalDst will be fixed up to MergeBB.
366   BranchInst::Create(II->getNormalDest(), *MergeBB);
367 }
368 
369 // Find the PHI in BB that have the CallResult as the operand.
370 static bool getCallRetPHINode(BasicBlock *BB, Instruction *Inst) {
371   BasicBlock *From = Inst->getParent();
372   for (auto &I : *BB) {
373     PHINode *PHI = dyn_cast<PHINode>(&I);
374     if (!PHI)
375       continue;
376     int IX = PHI->getBasicBlockIndex(From);
377     if (IX == -1)
378       continue;
379     Value *V = PHI->getIncomingValue(IX);
380     if (dyn_cast<Instruction>(V) == Inst)
381       return true;
382   }
383   return false;
384 }
385 
386 // This method fixes up PHI nodes in BB where BB is the UnwindDest of an
387 // invoke instruction. In BB, there may be PHIs with incoming block being
388 // OrigBB (the MergeBB after if-then-else splitting). After moving the invoke
389 // instructions to its own BB, OrigBB is no longer the predecessor block of BB.
390 // Instead two new predecessors are added: IndirectCallBB and DirectCallBB,
391 // so the PHI node's incoming BBs need to be fixed up accordingly.
392 static void fixupPHINodeForUnwind(Instruction *Inst, BasicBlock *BB,
393                                   BasicBlock *OrigBB,
394                                   BasicBlock *IndirectCallBB,
395                                   BasicBlock *DirectCallBB) {
396   for (auto &I : *BB) {
397     PHINode *PHI = dyn_cast<PHINode>(&I);
398     if (!PHI)
399       continue;
400     int IX = PHI->getBasicBlockIndex(OrigBB);
401     if (IX == -1)
402       continue;
403     Value *V = PHI->getIncomingValue(IX);
404     PHI->addIncoming(V, IndirectCallBB);
405     PHI->setIncomingBlock(IX, DirectCallBB);
406   }
407 }
408 
409 // This method fixes up PHI nodes in BB where BB is the NormalDest of an
410 // invoke instruction. In BB, there may be PHIs with incoming block being
411 // OrigBB (the MergeBB after if-then-else splitting). After moving the invoke
412 // instructions to its own BB, a new incoming edge will be added to the original
413 // NormalDstBB from the IndirectCallBB.
414 static void fixupPHINodeForNormalDest(Instruction *Inst, BasicBlock *BB,
415                                       BasicBlock *OrigBB,
416                                       BasicBlock *IndirectCallBB,
417                                       Instruction *NewInst) {
418   for (auto &I : *BB) {
419     PHINode *PHI = dyn_cast<PHINode>(&I);
420     if (!PHI)
421       continue;
422     int IX = PHI->getBasicBlockIndex(OrigBB);
423     if (IX == -1)
424       continue;
425     Value *V = PHI->getIncomingValue(IX);
426     if (dyn_cast<Instruction>(V) == Inst) {
427       PHI->setIncomingBlock(IX, IndirectCallBB);
428       PHI->addIncoming(NewInst, OrigBB);
429       continue;
430     }
431     PHI->addIncoming(V, IndirectCallBB);
432   }
433 }
434 
435 // Add a bitcast instruction to the direct-call return value if needed.
436 static Instruction *insertCallRetCast(const Instruction *Inst,
437                                       Instruction *DirectCallInst,
438                                       Function *DirectCallee) {
439   if (Inst->getType()->isVoidTy())
440     return DirectCallInst;
441 
442   Type *CallRetType = Inst->getType();
443   Type *FuncRetType = DirectCallee->getReturnType();
444   if (FuncRetType == CallRetType)
445     return DirectCallInst;
446 
447   BasicBlock *InsertionBB;
448   if (CallInst *CI = dyn_cast<CallInst>(DirectCallInst))
449     InsertionBB = CI->getParent();
450   else
451     InsertionBB = (dyn_cast<InvokeInst>(DirectCallInst))->getNormalDest();
452 
453   return (new BitCastInst(DirectCallInst, CallRetType, "",
454                           InsertionBB->getTerminator()));
455 }
456 
457 // Create a DirectCall instruction in the DirectCallBB.
458 // Parameter Inst is the indirect-call (invoke) instruction.
459 // DirectCallee is the decl of the direct-call (invoke) target.
460 // DirecallBB is the BB that the direct-call (invoke) instruction is inserted.
461 // MergeBB is the bottom BB of the if-then-else-diamond after the
462 // transformation. For invoke instruction, the edges from DirectCallBB and
463 // IndirectCallBB to MergeBB are removed before this call (during
464 // createIfThenElse).
465 static Instruction *createDirectCallInst(const Instruction *Inst,
466                                          Function *DirectCallee,
467                                          BasicBlock *DirectCallBB,
468                                          BasicBlock *MergeBB) {
469   Instruction *NewInst = Inst->clone();
470   if (CallInst *CI = dyn_cast<CallInst>(NewInst)) {
471     CI->setCalledFunction(DirectCallee);
472     CI->mutateFunctionType(DirectCallee->getFunctionType());
473   } else {
474     // Must be an invoke instruction. Direct invoke's normal destination is
475     // fixed up to MergeBB. MergeBB is the place where return cast is inserted.
476     // Also since IndirectCallBB does not have an edge to MergeBB, there is no
477     // need to insert new PHIs into MergeBB.
478     InvokeInst *II = dyn_cast<InvokeInst>(NewInst);
479     assert(II);
480     II->setCalledFunction(DirectCallee);
481     II->mutateFunctionType(DirectCallee->getFunctionType());
482     II->setNormalDest(MergeBB);
483   }
484 
485   DirectCallBB->getInstList().insert(DirectCallBB->getFirstInsertionPt(),
486                                      NewInst);
487 
488   // Clear the value profile data.
489   NewInst->setMetadata(LLVMContext::MD_prof, nullptr);
490   CallSite NewCS(NewInst);
491   FunctionType *DirectCalleeType = DirectCallee->getFunctionType();
492   unsigned ParamNum = DirectCalleeType->getFunctionNumParams();
493   for (unsigned I = 0; I < ParamNum; ++I) {
494     Type *ATy = NewCS.getArgument(I)->getType();
495     Type *PTy = DirectCalleeType->getParamType(I);
496     if (ATy != PTy) {
497       BitCastInst *BI = new BitCastInst(NewCS.getArgument(I), PTy, "", NewInst);
498       NewCS.setArgument(I, BI);
499     }
500   }
501 
502   return insertCallRetCast(Inst, NewInst, DirectCallee);
503 }
504 
505 // Create a PHI to unify the return values of calls.
506 static void insertCallRetPHI(Instruction *Inst, Instruction *CallResult,
507                              Function *DirectCallee) {
508   if (Inst->getType()->isVoidTy())
509     return;
510 
511   if (Inst->use_empty())
512     return;
513 
514   BasicBlock *RetValBB = CallResult->getParent();
515 
516   BasicBlock *PHIBB;
517   if (InvokeInst *II = dyn_cast<InvokeInst>(CallResult))
518     RetValBB = II->getNormalDest();
519 
520   PHIBB = RetValBB->getSingleSuccessor();
521   if (getCallRetPHINode(PHIBB, Inst))
522     return;
523 
524   PHINode *CallRetPHI = PHINode::Create(Inst->getType(), 0);
525   PHIBB->getInstList().push_front(CallRetPHI);
526   Inst->replaceAllUsesWith(CallRetPHI);
527   CallRetPHI->addIncoming(Inst, Inst->getParent());
528   CallRetPHI->addIncoming(CallResult, RetValBB);
529 }
530 
531 // This function does the actual indirect-call promotion transformation:
532 // For an indirect-call like:
533 //     Ret = (*Foo)(Args);
534 // It transforms to:
535 //     if (Foo == DirectCallee)
536 //        Ret1 = DirectCallee(Args);
537 //     else
538 //        Ret2 = (*Foo)(Args);
539 //     Ret = phi(Ret1, Ret2);
540 // It adds type casts for the args do not match the parameters and the return
541 // value. Branch weights metadata also updated.
542 // If \p AttachProfToDirectCall is true, a prof metadata is attached to the
543 // new direct call to contain \p Count. This is used by SamplePGO inliner to
544 // check callsite hotness.
545 // Returns the promoted direct call instruction.
546 Instruction *llvm::promoteIndirectCall(Instruction *Inst,
547                                        Function *DirectCallee, uint64_t Count,
548                                        uint64_t TotalCount,
549                                        bool AttachProfToDirectCall,
550                                        OptimizationRemarkEmitter *ORE) {
551   assert(DirectCallee != nullptr);
552   BasicBlock *BB = Inst->getParent();
553   // Just to suppress the non-debug build warning.
554   (void)BB;
555   DEBUG(dbgs() << "\n\n== Basic Block Before ==\n");
556   DEBUG(dbgs() << *BB << "\n");
557 
558   BasicBlock *DirectCallBB, *IndirectCallBB, *MergeBB;
559   createIfThenElse(Inst, DirectCallee, Count, TotalCount, &DirectCallBB,
560                    &IndirectCallBB, &MergeBB);
561 
562   Instruction *NewInst =
563       createDirectCallInst(Inst, DirectCallee, DirectCallBB, MergeBB);
564 
565   if (AttachProfToDirectCall) {
566     SmallVector<uint32_t, 1> Weights;
567     Weights.push_back(Count);
568     MDBuilder MDB(NewInst->getContext());
569     if (Instruction *DI = dyn_cast<Instruction>(NewInst->stripPointerCasts()))
570       DI->setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
571   }
572 
573   // Move Inst from MergeBB to IndirectCallBB.
574   Inst->removeFromParent();
575   IndirectCallBB->getInstList().insert(IndirectCallBB->getFirstInsertionPt(),
576                                        Inst);
577 
578   if (InvokeInst *II = dyn_cast<InvokeInst>(Inst)) {
579     // At this point, the original indirect invoke instruction has the original
580     // UnwindDest and NormalDest. For the direct invoke instruction, the
581     // NormalDest points to MergeBB, and MergeBB jumps to the original
582     // NormalDest. MergeBB might have a new bitcast instruction for the return
583     // value. The PHIs are with the original NormalDest. Since we now have two
584     // incoming edges to NormalDest and UnwindDest, we have to do some fixups.
585     //
586     // UnwindDest will not use the return value. So pass nullptr here.
587     fixupPHINodeForUnwind(Inst, II->getUnwindDest(), MergeBB, IndirectCallBB,
588                           DirectCallBB);
589     // We don't need to update the operand from NormalDest for DirectCallBB.
590     // Pass nullptr here.
591     fixupPHINodeForNormalDest(Inst, II->getNormalDest(), MergeBB,
592                               IndirectCallBB, NewInst);
593   }
594 
595   insertCallRetPHI(Inst, NewInst, DirectCallee);
596 
597   DEBUG(dbgs() << "\n== Basic Blocks After ==\n");
598   DEBUG(dbgs() << *BB << *DirectCallBB << *IndirectCallBB << *MergeBB << "\n");
599 
600   using namespace ore;
601   if (ORE)
602     ORE->emit(OptimizationRemark(DEBUG_TYPE, "Promoted", Inst)
603               << "Promote indirect call to " << NV("DirectCallee", DirectCallee)
604               << " with count " << NV("Count", Count) << " out of "
605               << NV("TotalCount", TotalCount));
606   return NewInst;
607 }
608 
609 // Promote indirect-call to conditional direct-call for one callsite.
610 uint32_t ICallPromotionFunc::tryToPromote(
611     Instruction *Inst, const std::vector<PromotionCandidate> &Candidates,
612     uint64_t &TotalCount) {
613   uint32_t NumPromoted = 0;
614 
615   for (auto &C : Candidates) {
616     uint64_t Count = C.Count;
617     promoteIndirectCall(Inst, C.TargetFunction, Count, TotalCount, SamplePGO,
618                         &ORE);
619     assert(TotalCount >= Count);
620     TotalCount -= Count;
621     NumOfPGOICallPromotion++;
622     NumPromoted++;
623   }
624   return NumPromoted;
625 }
626 
627 // Traverse all the indirect-call callsite and get the value profile
628 // annotation to perform indirect-call promotion.
629 bool ICallPromotionFunc::processFunction(ProfileSummaryInfo *PSI) {
630   bool Changed = false;
631   ICallPromotionAnalysis ICallAnalysis;
632   for (auto &I : findIndirectCallSites(F)) {
633     uint32_t NumVals, NumCandidates;
634     uint64_t TotalCount;
635     auto ICallProfDataRef = ICallAnalysis.getPromotionCandidatesForInstruction(
636         I, NumVals, TotalCount, NumCandidates);
637     if (!NumCandidates ||
638         (PSI && PSI->hasProfileSummary() && !PSI->isHotCount(TotalCount)))
639       continue;
640     auto PromotionCandidates = getPromotionCandidatesForCallSite(
641         I, ICallProfDataRef, TotalCount, NumCandidates);
642     uint32_t NumPromoted = tryToPromote(I, PromotionCandidates, TotalCount);
643     if (NumPromoted == 0)
644       continue;
645 
646     Changed = true;
647     // Adjust the MD.prof metadata. First delete the old one.
648     I->setMetadata(LLVMContext::MD_prof, nullptr);
649     // If all promoted, we don't need the MD.prof metadata.
650     if (TotalCount == 0 || NumPromoted == NumVals)
651       continue;
652     // Otherwise we need update with the un-promoted records back.
653     annotateValueSite(*M, *I, ICallProfDataRef.slice(NumPromoted), TotalCount,
654                       IPVK_IndirectCallTarget, NumCandidates);
655   }
656   return Changed;
657 }
658 
659 // A wrapper function that does the actual work.
660 static bool promoteIndirectCalls(Module &M, ProfileSummaryInfo *PSI,
661                                  bool InLTO, bool SamplePGO,
662                                  ModuleAnalysisManager *AM = nullptr) {
663   if (DisableICP)
664     return false;
665   InstrProfSymtab Symtab;
666   if (Error E = Symtab.create(M, InLTO)) {
667     std::string SymtabFailure = toString(std::move(E));
668     DEBUG(dbgs() << "Failed to create symtab: " << SymtabFailure << "\n");
669     (void)SymtabFailure;
670     return false;
671   }
672   bool Changed = false;
673   for (auto &F : M) {
674     if (F.isDeclaration())
675       continue;
676     if (F.hasFnAttribute(Attribute::OptimizeNone))
677       continue;
678 
679     std::unique_ptr<OptimizationRemarkEmitter> OwnedORE;
680     OptimizationRemarkEmitter *ORE;
681     if (AM) {
682       auto &FAM =
683           AM->getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
684       ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
685     } else {
686       OwnedORE = make_unique<OptimizationRemarkEmitter>(&F);
687       ORE = OwnedORE.get();
688     }
689 
690     ICallPromotionFunc ICallPromotion(F, &M, &Symtab, SamplePGO, *ORE);
691     bool FuncChanged = ICallPromotion.processFunction(PSI);
692     if (ICPDUMPAFTER && FuncChanged) {
693       DEBUG(dbgs() << "\n== IR Dump After =="; F.print(dbgs()));
694       DEBUG(dbgs() << "\n");
695     }
696     Changed |= FuncChanged;
697     if (ICPCutOff != 0 && NumOfPGOICallPromotion >= ICPCutOff) {
698       DEBUG(dbgs() << " Stop: Cutoff reached.\n");
699       break;
700     }
701   }
702   return Changed;
703 }
704 
705 bool PGOIndirectCallPromotionLegacyPass::runOnModule(Module &M) {
706   ProfileSummaryInfo *PSI =
707       getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
708 
709   // Command-line option has the priority for InLTO.
710   return promoteIndirectCalls(M, PSI, InLTO | ICPLTOMode,
711                               SamplePGO | ICPSamplePGOMode);
712 }
713 
714 PreservedAnalyses PGOIndirectCallPromotion::run(Module &M,
715                                                 ModuleAnalysisManager &AM) {
716   ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(M);
717 
718   if (!promoteIndirectCalls(M, PSI, InLTO | ICPLTOMode,
719                             SamplePGO | ICPSamplePGOMode, &AM))
720     return PreservedAnalyses::all();
721 
722   return PreservedAnalyses::none();
723 }
724