xref: /llvm-project/llvm/lib/Analysis/ModuleSummaryAnalysis.cpp (revision dda73336ad22bd0b5ecda17040c50fb10fcbe5fb)
1 //===- ModuleSummaryAnalysis.cpp - Module summary index builder -----------===//
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 pass builds a ModuleSummaryIndex object for the module, to be written
10 // to bitcode or LLVM assembly.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Analysis/ModuleSummaryAnalysis.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/MapVector.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/Analysis/BlockFrequencyInfo.h"
24 #include "llvm/Analysis/BranchProbabilityInfo.h"
25 #include "llvm/Analysis/IndirectCallPromotionAnalysis.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/MemoryProfileInfo.h"
28 #include "llvm/Analysis/ProfileSummaryInfo.h"
29 #include "llvm/Analysis/StackSafetyAnalysis.h"
30 #include "llvm/Analysis/TypeMetadataUtils.h"
31 #include "llvm/IR/Attributes.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/Constant.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/GlobalAlias.h"
38 #include "llvm/IR/GlobalValue.h"
39 #include "llvm/IR/GlobalVariable.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/Metadata.h"
43 #include "llvm/IR/Module.h"
44 #include "llvm/IR/ModuleSummaryIndex.h"
45 #include "llvm/IR/Use.h"
46 #include "llvm/IR/User.h"
47 #include "llvm/InitializePasses.h"
48 #include "llvm/Object/ModuleSymbolTable.h"
49 #include "llvm/Object/SymbolicFile.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Support/Casting.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/FileSystem.h"
54 #include <algorithm>
55 #include <cassert>
56 #include <cstdint>
57 #include <vector>
58 
59 using namespace llvm;
60 using namespace llvm::memprof;
61 
62 #define DEBUG_TYPE "module-summary-analysis"
63 
64 // Option to force edges cold which will block importing when the
65 // -import-cold-multiplier is set to 0. Useful for debugging.
66 namespace llvm {
67 FunctionSummary::ForceSummaryHotnessType ForceSummaryEdgesCold =
68     FunctionSummary::FSHT_None;
69 } // namespace llvm
70 
71 static cl::opt<FunctionSummary::ForceSummaryHotnessType, true> FSEC(
72     "force-summary-edges-cold", cl::Hidden, cl::location(ForceSummaryEdgesCold),
73     cl::desc("Force all edges in the function summary to cold"),
74     cl::values(clEnumValN(FunctionSummary::FSHT_None, "none", "None."),
75                clEnumValN(FunctionSummary::FSHT_AllNonCritical,
76                           "all-non-critical", "All non-critical edges."),
77                clEnumValN(FunctionSummary::FSHT_All, "all", "All edges.")));
78 
79 static cl::opt<std::string> ModuleSummaryDotFile(
80     "module-summary-dot-file", cl::Hidden, cl::value_desc("filename"),
81     cl::desc("File to emit dot graph of new summary into"));
82 
83 extern cl::opt<bool> ScalePartialSampleProfileWorkingSetSize;
84 
85 extern cl::opt<unsigned> MaxNumVTableAnnotations;
86 
87 // Walk through the operands of a given User via worklist iteration and populate
88 // the set of GlobalValue references encountered. Invoked either on an
89 // Instruction or a GlobalVariable (which walks its initializer).
90 // Return true if any of the operands contains blockaddress. This is important
91 // to know when computing summary for global var, because if global variable
92 // references basic block address we can't import it separately from function
93 // containing that basic block. For simplicity we currently don't import such
94 // global vars at all. When importing function we aren't interested if any
95 // instruction in it takes an address of any basic block, because instruction
96 // can only take an address of basic block located in the same function.
97 static bool findRefEdges(ModuleSummaryIndex &Index, const User *CurUser,
98                          SetVector<ValueInfo, std::vector<ValueInfo>> &RefEdges,
99                          SmallPtrSet<const User *, 8> &Visited) {
100   bool HasBlockAddress = false;
101   SmallVector<const User *, 32> Worklist;
102   if (Visited.insert(CurUser).second)
103     Worklist.push_back(CurUser);
104 
105   while (!Worklist.empty()) {
106     const User *U = Worklist.pop_back_val();
107     const auto *CB = dyn_cast<CallBase>(U);
108 
109     for (const auto &OI : U->operands()) {
110       const User *Operand = dyn_cast<User>(OI);
111       if (!Operand)
112         continue;
113       if (isa<BlockAddress>(Operand)) {
114         HasBlockAddress = true;
115         continue;
116       }
117       if (auto *GV = dyn_cast<GlobalValue>(Operand)) {
118         // We have a reference to a global value. This should be added to
119         // the reference set unless it is a callee. Callees are handled
120         // specially by WriteFunction and are added to a separate list.
121         if (!(CB && CB->isCallee(&OI)))
122           RefEdges.insert(Index.getOrInsertValueInfo(GV));
123         continue;
124       }
125       if (Visited.insert(Operand).second)
126         Worklist.push_back(Operand);
127     }
128   }
129 
130   const Instruction *I = dyn_cast<Instruction>(CurUser);
131   if (I) {
132     uint32_t ActualNumValueData = 0;
133     uint64_t TotalCount = 0;
134     // MaxNumVTableAnnotations is the maximum number of vtables annotated on
135     // the instruction.
136     auto ValueDataArray =
137         getValueProfDataFromInst(*I, IPVK_VTableTarget, MaxNumVTableAnnotations,
138                                  ActualNumValueData, TotalCount);
139 
140     if (ValueDataArray.get()) {
141       for (uint32_t j = 0; j < ActualNumValueData; j++) {
142         RefEdges.insert(Index.getOrInsertValueInfo(/* VTableGUID = */
143                                                    ValueDataArray[j].Value));
144       }
145     }
146   }
147   return HasBlockAddress;
148 }
149 
150 static CalleeInfo::HotnessType getHotness(uint64_t ProfileCount,
151                                           ProfileSummaryInfo *PSI) {
152   if (!PSI)
153     return CalleeInfo::HotnessType::Unknown;
154   if (PSI->isHotCount(ProfileCount))
155     return CalleeInfo::HotnessType::Hot;
156   if (PSI->isColdCount(ProfileCount))
157     return CalleeInfo::HotnessType::Cold;
158   return CalleeInfo::HotnessType::None;
159 }
160 
161 static bool isNonRenamableLocal(const GlobalValue &GV) {
162   return GV.hasSection() && GV.hasLocalLinkage();
163 }
164 
165 /// Determine whether this call has all constant integer arguments (excluding
166 /// "this") and summarize it to VCalls or ConstVCalls as appropriate.
167 static void addVCallToSet(
168     DevirtCallSite Call, GlobalValue::GUID Guid,
169     SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
170         &VCalls,
171     SetVector<FunctionSummary::ConstVCall,
172               std::vector<FunctionSummary::ConstVCall>> &ConstVCalls) {
173   std::vector<uint64_t> Args;
174   // Start from the second argument to skip the "this" pointer.
175   for (auto &Arg : drop_begin(Call.CB.args())) {
176     auto *CI = dyn_cast<ConstantInt>(Arg);
177     if (!CI || CI->getBitWidth() > 64) {
178       VCalls.insert({Guid, Call.Offset});
179       return;
180     }
181     Args.push_back(CI->getZExtValue());
182   }
183   ConstVCalls.insert({{Guid, Call.Offset}, std::move(Args)});
184 }
185 
186 /// If this intrinsic call requires that we add information to the function
187 /// summary, do so via the non-constant reference arguments.
188 static void addIntrinsicToSummary(
189     const CallInst *CI,
190     SetVector<GlobalValue::GUID, std::vector<GlobalValue::GUID>> &TypeTests,
191     SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
192         &TypeTestAssumeVCalls,
193     SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
194         &TypeCheckedLoadVCalls,
195     SetVector<FunctionSummary::ConstVCall,
196               std::vector<FunctionSummary::ConstVCall>>
197         &TypeTestAssumeConstVCalls,
198     SetVector<FunctionSummary::ConstVCall,
199               std::vector<FunctionSummary::ConstVCall>>
200         &TypeCheckedLoadConstVCalls,
201     DominatorTree &DT) {
202   switch (CI->getCalledFunction()->getIntrinsicID()) {
203   case Intrinsic::type_test:
204   case Intrinsic::public_type_test: {
205     auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(1));
206     auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata());
207     if (!TypeId)
208       break;
209     GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString());
210 
211     // Produce a summary from type.test intrinsics. We only summarize type.test
212     // intrinsics that are used other than by an llvm.assume intrinsic.
213     // Intrinsics that are assumed are relevant only to the devirtualization
214     // pass, not the type test lowering pass.
215     bool HasNonAssumeUses = llvm::any_of(CI->uses(), [](const Use &CIU) {
216       return !isa<AssumeInst>(CIU.getUser());
217     });
218     if (HasNonAssumeUses)
219       TypeTests.insert(Guid);
220 
221     SmallVector<DevirtCallSite, 4> DevirtCalls;
222     SmallVector<CallInst *, 4> Assumes;
223     findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT);
224     for (auto &Call : DevirtCalls)
225       addVCallToSet(Call, Guid, TypeTestAssumeVCalls,
226                     TypeTestAssumeConstVCalls);
227 
228     break;
229   }
230 
231   case Intrinsic::type_checked_load_relative:
232   case Intrinsic::type_checked_load: {
233     auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(2));
234     auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata());
235     if (!TypeId)
236       break;
237     GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString());
238 
239     SmallVector<DevirtCallSite, 4> DevirtCalls;
240     SmallVector<Instruction *, 4> LoadedPtrs;
241     SmallVector<Instruction *, 4> Preds;
242     bool HasNonCallUses = false;
243     findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
244                                                HasNonCallUses, CI, DT);
245     // Any non-call uses of the result of llvm.type.checked.load will
246     // prevent us from optimizing away the llvm.type.test.
247     if (HasNonCallUses)
248       TypeTests.insert(Guid);
249     for (auto &Call : DevirtCalls)
250       addVCallToSet(Call, Guid, TypeCheckedLoadVCalls,
251                     TypeCheckedLoadConstVCalls);
252 
253     break;
254   }
255   default:
256     break;
257   }
258 }
259 
260 static bool isNonVolatileLoad(const Instruction *I) {
261   if (const auto *LI = dyn_cast<LoadInst>(I))
262     return !LI->isVolatile();
263 
264   return false;
265 }
266 
267 static bool isNonVolatileStore(const Instruction *I) {
268   if (const auto *SI = dyn_cast<StoreInst>(I))
269     return !SI->isVolatile();
270 
271   return false;
272 }
273 
274 // Returns true if the function definition must be unreachable.
275 //
276 // Note if this helper function returns true, `F` is guaranteed
277 // to be unreachable; if it returns false, `F` might still
278 // be unreachable but not covered by this helper function.
279 static bool mustBeUnreachableFunction(const Function &F) {
280   // A function must be unreachable if its entry block ends with an
281   // 'unreachable'.
282   assert(!F.isDeclaration());
283   return isa<UnreachableInst>(F.getEntryBlock().getTerminator());
284 }
285 
286 static void computeFunctionSummary(
287     ModuleSummaryIndex &Index, const Module &M, const Function &F,
288     BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, DominatorTree &DT,
289     bool HasLocalsInUsedOrAsm, DenseSet<GlobalValue::GUID> &CantBePromoted,
290     bool IsThinLTO,
291     std::function<const StackSafetyInfo *(const Function &F)> GetSSICallback) {
292   // Summary not currently supported for anonymous functions, they should
293   // have been named.
294   assert(F.hasName());
295 
296   unsigned NumInsts = 0;
297   // Map from callee ValueId to profile count. Used to accumulate profile
298   // counts for all static calls to a given callee.
299   MapVector<ValueInfo, CalleeInfo, DenseMap<ValueInfo, unsigned>,
300             std::vector<std::pair<ValueInfo, CalleeInfo>>>
301       CallGraphEdges;
302   SetVector<ValueInfo, std::vector<ValueInfo>> RefEdges, LoadRefEdges,
303       StoreRefEdges;
304   SetVector<GlobalValue::GUID, std::vector<GlobalValue::GUID>> TypeTests;
305   SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
306       TypeTestAssumeVCalls, TypeCheckedLoadVCalls;
307   SetVector<FunctionSummary::ConstVCall,
308             std::vector<FunctionSummary::ConstVCall>>
309       TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls;
310   ICallPromotionAnalysis ICallAnalysis;
311   SmallPtrSet<const User *, 8> Visited;
312 
313   // Add personality function, prefix data and prologue data to function's ref
314   // list.
315   findRefEdges(Index, &F, RefEdges, Visited);
316   std::vector<const Instruction *> NonVolatileLoads;
317   std::vector<const Instruction *> NonVolatileStores;
318 
319   std::vector<CallsiteInfo> Callsites;
320   std::vector<AllocInfo> Allocs;
321 
322 #ifndef NDEBUG
323   DenseSet<const CallBase *> CallsThatMayHaveMemprofSummary;
324 #endif
325 
326   bool HasInlineAsmMaybeReferencingInternal = false;
327   bool HasIndirBranchToBlockAddress = false;
328   bool HasIFuncCall = false;
329   bool HasUnknownCall = false;
330   bool MayThrow = false;
331   for (const BasicBlock &BB : F) {
332     // We don't allow inlining of function with indirect branch to blockaddress.
333     // If the blockaddress escapes the function, e.g., via a global variable,
334     // inlining may lead to an invalid cross-function reference. So we shouldn't
335     // import such function either.
336     if (BB.hasAddressTaken()) {
337       for (User *U : BlockAddress::get(const_cast<BasicBlock *>(&BB))->users())
338         if (!isa<CallBrInst>(*U)) {
339           HasIndirBranchToBlockAddress = true;
340           break;
341         }
342     }
343 
344     for (const Instruction &I : BB) {
345       if (I.isDebugOrPseudoInst())
346         continue;
347       ++NumInsts;
348 
349       // Regular LTO module doesn't participate in ThinLTO import,
350       // so no reference from it can be read/writeonly, since this
351       // would require importing variable as local copy
352       if (IsThinLTO) {
353         if (isNonVolatileLoad(&I)) {
354           // Postpone processing of non-volatile load instructions
355           // See comments below
356           Visited.insert(&I);
357           NonVolatileLoads.push_back(&I);
358           continue;
359         } else if (isNonVolatileStore(&I)) {
360           Visited.insert(&I);
361           NonVolatileStores.push_back(&I);
362           // All references from second operand of store (destination address)
363           // can be considered write-only if they're not referenced by any
364           // non-store instruction. References from first operand of store
365           // (stored value) can't be treated either as read- or as write-only
366           // so we add them to RefEdges as we do with all other instructions
367           // except non-volatile load.
368           Value *Stored = I.getOperand(0);
369           if (auto *GV = dyn_cast<GlobalValue>(Stored))
370             // findRefEdges will try to examine GV operands, so instead
371             // of calling it we should add GV to RefEdges directly.
372             RefEdges.insert(Index.getOrInsertValueInfo(GV));
373           else if (auto *U = dyn_cast<User>(Stored))
374             findRefEdges(Index, U, RefEdges, Visited);
375           continue;
376         }
377       }
378       findRefEdges(Index, &I, RefEdges, Visited);
379       const auto *CB = dyn_cast<CallBase>(&I);
380       if (!CB) {
381         if (I.mayThrow())
382           MayThrow = true;
383         continue;
384       }
385 
386       const auto *CI = dyn_cast<CallInst>(&I);
387       // Since we don't know exactly which local values are referenced in inline
388       // assembly, conservatively mark the function as possibly referencing
389       // a local value from inline assembly to ensure we don't export a
390       // reference (which would require renaming and promotion of the
391       // referenced value).
392       if (HasLocalsInUsedOrAsm && CI && CI->isInlineAsm())
393         HasInlineAsmMaybeReferencingInternal = true;
394 
395       auto *CalledValue = CB->getCalledOperand();
396       auto *CalledFunction = CB->getCalledFunction();
397       if (CalledValue && !CalledFunction) {
398         CalledValue = CalledValue->stripPointerCasts();
399         // Stripping pointer casts can reveal a called function.
400         CalledFunction = dyn_cast<Function>(CalledValue);
401       }
402       // Check if this is an alias to a function. If so, get the
403       // called aliasee for the checks below.
404       if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) {
405         assert(!CalledFunction && "Expected null called function in callsite for alias");
406         CalledFunction = dyn_cast<Function>(GA->getAliaseeObject());
407       }
408       // Check if this is a direct call to a known function or a known
409       // intrinsic, or an indirect call with profile data.
410       if (CalledFunction) {
411         if (CI && CalledFunction->isIntrinsic()) {
412           addIntrinsicToSummary(
413               CI, TypeTests, TypeTestAssumeVCalls, TypeCheckedLoadVCalls,
414               TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls, DT);
415           continue;
416         }
417         // We should have named any anonymous globals
418         assert(CalledFunction->hasName());
419         auto ScaledCount = PSI->getProfileCount(*CB, BFI);
420         auto Hotness = ScaledCount ? getHotness(*ScaledCount, PSI)
421                                    : CalleeInfo::HotnessType::Unknown;
422         if (ForceSummaryEdgesCold != FunctionSummary::FSHT_None)
423           Hotness = CalleeInfo::HotnessType::Cold;
424 
425         // Use the original CalledValue, in case it was an alias. We want
426         // to record the call edge to the alias in that case. Eventually
427         // an alias summary will be created to associate the alias and
428         // aliasee.
429         auto &ValueInfo = CallGraphEdges[Index.getOrInsertValueInfo(
430             cast<GlobalValue>(CalledValue))];
431         ValueInfo.updateHotness(Hotness);
432         if (CB->isTailCall())
433           ValueInfo.setHasTailCall(true);
434         // Add the relative block frequency to CalleeInfo if there is no profile
435         // information.
436         if (BFI != nullptr && Hotness == CalleeInfo::HotnessType::Unknown) {
437           uint64_t BBFreq = BFI->getBlockFreq(&BB).getFrequency();
438           uint64_t EntryFreq = BFI->getEntryFreq().getFrequency();
439           ValueInfo.updateRelBlockFreq(BBFreq, EntryFreq);
440         }
441       } else {
442         HasUnknownCall = true;
443         // If F is imported, a local linkage ifunc (e.g. target_clones on a
444         // static function) called by F will be cloned. Since summaries don't
445         // track ifunc, we do not know implementation functions referenced by
446         // the ifunc resolver need to be promoted in the exporter, and we will
447         // get linker errors due to cloned declarations for implementation
448         // functions. As a simple fix, just mark F as not eligible for import.
449         // Non-local ifunc is not cloned and does not have the issue.
450         if (auto *GI = dyn_cast_if_present<GlobalIFunc>(CalledValue))
451           if (GI->hasLocalLinkage())
452             HasIFuncCall = true;
453         // Skip inline assembly calls.
454         if (CI && CI->isInlineAsm())
455           continue;
456         // Skip direct calls.
457         if (!CalledValue || isa<Constant>(CalledValue))
458           continue;
459 
460         // Check if the instruction has a callees metadata. If so, add callees
461         // to CallGraphEdges to reflect the references from the metadata, and
462         // to enable importing for subsequent indirect call promotion and
463         // inlining.
464         if (auto *MD = I.getMetadata(LLVMContext::MD_callees)) {
465           for (const auto &Op : MD->operands()) {
466             Function *Callee = mdconst::extract_or_null<Function>(Op);
467             if (Callee)
468               CallGraphEdges[Index.getOrInsertValueInfo(Callee)];
469           }
470         }
471 
472         uint32_t NumVals, NumCandidates;
473         uint64_t TotalCount;
474         auto CandidateProfileData =
475             ICallAnalysis.getPromotionCandidatesForInstruction(
476                 &I, NumVals, TotalCount, NumCandidates);
477         for (const auto &Candidate : CandidateProfileData)
478           CallGraphEdges[Index.getOrInsertValueInfo(Candidate.Value)]
479               .updateHotness(getHotness(Candidate.Count, PSI));
480       }
481 
482       // Summarize memprof related metadata. This is only needed for ThinLTO.
483       if (!IsThinLTO)
484         continue;
485 
486       // TODO: Skip indirect calls for now. Need to handle these better, likely
487       // by creating multiple Callsites, one per target, then speculatively
488       // devirtualize while applying clone info in the ThinLTO backends. This
489       // will also be important because we will have a different set of clone
490       // versions per target. This handling needs to match that in the ThinLTO
491       // backend so we handle things consistently for matching of callsite
492       // summaries to instructions.
493       if (!CalledFunction)
494         continue;
495 
496       // Ensure we keep this analysis in sync with the handling in the ThinLTO
497       // backend (see MemProfContextDisambiguation::applyImport). Save this call
498       // so that we can skip it in checking the reverse case later.
499       assert(mayHaveMemprofSummary(CB));
500 #ifndef NDEBUG
501       CallsThatMayHaveMemprofSummary.insert(CB);
502 #endif
503 
504       // Compute the list of stack ids first (so we can trim them from the stack
505       // ids on any MIBs).
506       CallStack<MDNode, MDNode::op_iterator> InstCallsite(
507           I.getMetadata(LLVMContext::MD_callsite));
508       auto *MemProfMD = I.getMetadata(LLVMContext::MD_memprof);
509       if (MemProfMD) {
510         std::vector<MIBInfo> MIBs;
511         for (auto &MDOp : MemProfMD->operands()) {
512           auto *MIBMD = cast<const MDNode>(MDOp);
513           MDNode *StackNode = getMIBStackNode(MIBMD);
514           assert(StackNode);
515           SmallVector<unsigned> StackIdIndices;
516           CallStack<MDNode, MDNode::op_iterator> StackContext(StackNode);
517           // Collapse out any on the allocation call (inlining).
518           for (auto ContextIter =
519                    StackContext.beginAfterSharedPrefix(InstCallsite);
520                ContextIter != StackContext.end(); ++ContextIter) {
521             unsigned StackIdIdx = Index.addOrGetStackIdIndex(*ContextIter);
522             // If this is a direct recursion, simply skip the duplicate
523             // entries. If this is mutual recursion, handling is left to
524             // the LTO link analysis client.
525             if (StackIdIndices.empty() || StackIdIndices.back() != StackIdIdx)
526               StackIdIndices.push_back(StackIdIdx);
527           }
528           MIBs.push_back(
529               MIBInfo(getMIBAllocType(MIBMD), std::move(StackIdIndices)));
530         }
531         Allocs.push_back(AllocInfo(std::move(MIBs)));
532       } else if (!InstCallsite.empty()) {
533         SmallVector<unsigned> StackIdIndices;
534         for (auto StackId : InstCallsite)
535           StackIdIndices.push_back(Index.addOrGetStackIdIndex(StackId));
536         // Use the original CalledValue, in case it was an alias. We want
537         // to record the call edge to the alias in that case. Eventually
538         // an alias summary will be created to associate the alias and
539         // aliasee.
540         auto CalleeValueInfo =
541             Index.getOrInsertValueInfo(cast<GlobalValue>(CalledValue));
542         Callsites.push_back({CalleeValueInfo, StackIdIndices});
543       }
544     }
545   }
546 
547   if (PSI->hasPartialSampleProfile() && ScalePartialSampleProfileWorkingSetSize)
548     Index.addBlockCount(F.size());
549 
550   std::vector<ValueInfo> Refs;
551   if (IsThinLTO) {
552     auto AddRefEdges = [&](const std::vector<const Instruction *> &Instrs,
553                            SetVector<ValueInfo, std::vector<ValueInfo>> &Edges,
554                            SmallPtrSet<const User *, 8> &Cache) {
555       for (const auto *I : Instrs) {
556         Cache.erase(I);
557         findRefEdges(Index, I, Edges, Cache);
558       }
559     };
560 
561     // By now we processed all instructions in a function, except
562     // non-volatile loads and non-volatile value stores. Let's find
563     // ref edges for both of instruction sets
564     AddRefEdges(NonVolatileLoads, LoadRefEdges, Visited);
565     // We can add some values to the Visited set when processing load
566     // instructions which are also used by stores in NonVolatileStores.
567     // For example this can happen if we have following code:
568     //
569     // store %Derived* @foo, %Derived** bitcast (%Base** @bar to %Derived**)
570     // %42 = load %Derived*, %Derived** bitcast (%Base** @bar to %Derived**)
571     //
572     // After processing loads we'll add bitcast to the Visited set, and if
573     // we use the same set while processing stores, we'll never see store
574     // to @bar and @bar will be mistakenly treated as readonly.
575     SmallPtrSet<const llvm::User *, 8> StoreCache;
576     AddRefEdges(NonVolatileStores, StoreRefEdges, StoreCache);
577 
578     // If both load and store instruction reference the same variable
579     // we won't be able to optimize it. Add all such reference edges
580     // to RefEdges set.
581     for (const auto &VI : StoreRefEdges)
582       if (LoadRefEdges.remove(VI))
583         RefEdges.insert(VI);
584 
585     unsigned RefCnt = RefEdges.size();
586     // All new reference edges inserted in two loops below are either
587     // read or write only. They will be grouped in the end of RefEdges
588     // vector, so we can use a single integer value to identify them.
589     for (const auto &VI : LoadRefEdges)
590       RefEdges.insert(VI);
591 
592     unsigned FirstWORef = RefEdges.size();
593     for (const auto &VI : StoreRefEdges)
594       RefEdges.insert(VI);
595 
596     Refs = RefEdges.takeVector();
597     for (; RefCnt < FirstWORef; ++RefCnt)
598       Refs[RefCnt].setReadOnly();
599 
600     for (; RefCnt < Refs.size(); ++RefCnt)
601       Refs[RefCnt].setWriteOnly();
602   } else {
603     Refs = RefEdges.takeVector();
604   }
605   // Explicit add hot edges to enforce importing for designated GUIDs for
606   // sample PGO, to enable the same inlines as the profiled optimized binary.
607   for (auto &I : F.getImportGUIDs())
608     CallGraphEdges[Index.getOrInsertValueInfo(I)].updateHotness(
609         ForceSummaryEdgesCold == FunctionSummary::FSHT_All
610             ? CalleeInfo::HotnessType::Cold
611             : CalleeInfo::HotnessType::Critical);
612 
613 #ifndef NDEBUG
614   // Make sure that all calls we decided could not have memprof summaries get a
615   // false value for mayHaveMemprofSummary, to ensure that this handling remains
616   // in sync with the ThinLTO backend handling.
617   if (IsThinLTO) {
618     for (const BasicBlock &BB : F) {
619       for (const Instruction &I : BB) {
620         const auto *CB = dyn_cast<CallBase>(&I);
621         if (!CB)
622           continue;
623         // We already checked these above.
624         if (CallsThatMayHaveMemprofSummary.count(CB))
625           continue;
626         assert(!mayHaveMemprofSummary(CB));
627       }
628     }
629   }
630 #endif
631 
632   bool NonRenamableLocal = isNonRenamableLocal(F);
633   bool NotEligibleForImport = NonRenamableLocal ||
634                               HasInlineAsmMaybeReferencingInternal ||
635                               HasIndirBranchToBlockAddress || HasIFuncCall;
636   GlobalValueSummary::GVFlags Flags(
637       F.getLinkage(), F.getVisibility(), NotEligibleForImport,
638       /* Live = */ false, F.isDSOLocal(), F.canBeOmittedFromSymbolTable(),
639       GlobalValueSummary::ImportKind::Definition);
640   FunctionSummary::FFlags FunFlags{
641       F.doesNotAccessMemory(), F.onlyReadsMemory() && !F.doesNotAccessMemory(),
642       F.hasFnAttribute(Attribute::NoRecurse), F.returnDoesNotAlias(),
643       // FIXME: refactor this to use the same code that inliner is using.
644       // Don't try to import functions with noinline attribute.
645       F.getAttributes().hasFnAttr(Attribute::NoInline),
646       F.hasFnAttribute(Attribute::AlwaysInline),
647       F.hasFnAttribute(Attribute::NoUnwind), MayThrow, HasUnknownCall,
648       mustBeUnreachableFunction(F)};
649   std::vector<FunctionSummary::ParamAccess> ParamAccesses;
650   if (auto *SSI = GetSSICallback(F))
651     ParamAccesses = SSI->getParamAccesses(Index);
652   auto FuncSummary = std::make_unique<FunctionSummary>(
653       Flags, NumInsts, FunFlags, /*EntryCount=*/0, std::move(Refs),
654       CallGraphEdges.takeVector(), TypeTests.takeVector(),
655       TypeTestAssumeVCalls.takeVector(), TypeCheckedLoadVCalls.takeVector(),
656       TypeTestAssumeConstVCalls.takeVector(),
657       TypeCheckedLoadConstVCalls.takeVector(), std::move(ParamAccesses),
658       std::move(Callsites), std::move(Allocs));
659   if (NonRenamableLocal)
660     CantBePromoted.insert(F.getGUID());
661   Index.addGlobalValueSummary(F, std::move(FuncSummary));
662 }
663 
664 /// Find function pointers referenced within the given vtable initializer
665 /// (or subset of an initializer) \p I. The starting offset of \p I within
666 /// the vtable initializer is \p StartingOffset. Any discovered function
667 /// pointers are added to \p VTableFuncs along with their cumulative offset
668 /// within the initializer.
669 static void findFuncPointers(const Constant *I, uint64_t StartingOffset,
670                              const Module &M, ModuleSummaryIndex &Index,
671                              VTableFuncList &VTableFuncs) {
672   // First check if this is a function pointer.
673   if (I->getType()->isPointerTy()) {
674     auto C = I->stripPointerCasts();
675     auto A = dyn_cast<GlobalAlias>(C);
676     if (isa<Function>(C) || (A && isa<Function>(A->getAliasee()))) {
677       auto GV = dyn_cast<GlobalValue>(C);
678       assert(GV);
679       // We can disregard __cxa_pure_virtual as a possible call target, as
680       // calls to pure virtuals are UB.
681       if (GV && GV->getName() != "__cxa_pure_virtual")
682         VTableFuncs.push_back({Index.getOrInsertValueInfo(GV), StartingOffset});
683       return;
684     }
685   }
686 
687   // Walk through the elements in the constant struct or array and recursively
688   // look for virtual function pointers.
689   const DataLayout &DL = M.getDataLayout();
690   if (auto *C = dyn_cast<ConstantStruct>(I)) {
691     StructType *STy = dyn_cast<StructType>(C->getType());
692     assert(STy);
693     const StructLayout *SL = DL.getStructLayout(C->getType());
694 
695     for (auto EI : llvm::enumerate(STy->elements())) {
696       auto Offset = SL->getElementOffset(EI.index());
697       unsigned Op = SL->getElementContainingOffset(Offset);
698       findFuncPointers(cast<Constant>(I->getOperand(Op)),
699                        StartingOffset + Offset, M, Index, VTableFuncs);
700     }
701   } else if (auto *C = dyn_cast<ConstantArray>(I)) {
702     ArrayType *ATy = C->getType();
703     Type *EltTy = ATy->getElementType();
704     uint64_t EltSize = DL.getTypeAllocSize(EltTy);
705     for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) {
706       findFuncPointers(cast<Constant>(I->getOperand(i)),
707                        StartingOffset + i * EltSize, M, Index, VTableFuncs);
708     }
709   }
710 }
711 
712 // Identify the function pointers referenced by vtable definition \p V.
713 static void computeVTableFuncs(ModuleSummaryIndex &Index,
714                                const GlobalVariable &V, const Module &M,
715                                VTableFuncList &VTableFuncs) {
716   if (!V.isConstant())
717     return;
718 
719   findFuncPointers(V.getInitializer(), /*StartingOffset=*/0, M, Index,
720                    VTableFuncs);
721 
722 #ifndef NDEBUG
723   // Validate that the VTableFuncs list is ordered by offset.
724   uint64_t PrevOffset = 0;
725   for (auto &P : VTableFuncs) {
726     // The findVFuncPointers traversal should have encountered the
727     // functions in offset order. We need to use ">=" since PrevOffset
728     // starts at 0.
729     assert(P.VTableOffset >= PrevOffset);
730     PrevOffset = P.VTableOffset;
731   }
732 #endif
733 }
734 
735 /// Record vtable definition \p V for each type metadata it references.
736 static void
737 recordTypeIdCompatibleVtableReferences(ModuleSummaryIndex &Index,
738                                        const GlobalVariable &V,
739                                        SmallVectorImpl<MDNode *> &Types) {
740   for (MDNode *Type : Types) {
741     auto TypeID = Type->getOperand(1).get();
742 
743     uint64_t Offset =
744         cast<ConstantInt>(
745             cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
746             ->getZExtValue();
747 
748     if (auto *TypeId = dyn_cast<MDString>(TypeID))
749       Index.getOrInsertTypeIdCompatibleVtableSummary(TypeId->getString())
750           .push_back({Offset, Index.getOrInsertValueInfo(&V)});
751   }
752 }
753 
754 static void computeVariableSummary(ModuleSummaryIndex &Index,
755                                    const GlobalVariable &V,
756                                    DenseSet<GlobalValue::GUID> &CantBePromoted,
757                                    const Module &M,
758                                    SmallVectorImpl<MDNode *> &Types) {
759   SetVector<ValueInfo, std::vector<ValueInfo>> RefEdges;
760   SmallPtrSet<const User *, 8> Visited;
761   bool HasBlockAddress = findRefEdges(Index, &V, RefEdges, Visited);
762   bool NonRenamableLocal = isNonRenamableLocal(V);
763   GlobalValueSummary::GVFlags Flags(
764       V.getLinkage(), V.getVisibility(), NonRenamableLocal,
765       /* Live = */ false, V.isDSOLocal(), V.canBeOmittedFromSymbolTable(),
766       GlobalValueSummary::Definition);
767 
768   VTableFuncList VTableFuncs;
769   // If splitting is not enabled, then we compute the summary information
770   // necessary for index-based whole program devirtualization.
771   if (!Index.enableSplitLTOUnit()) {
772     Types.clear();
773     V.getMetadata(LLVMContext::MD_type, Types);
774     if (!Types.empty()) {
775       // Identify the function pointers referenced by this vtable definition.
776       computeVTableFuncs(Index, V, M, VTableFuncs);
777 
778       // Record this vtable definition for each type metadata it references.
779       recordTypeIdCompatibleVtableReferences(Index, V, Types);
780     }
781   }
782 
783   // Don't mark variables we won't be able to internalize as read/write-only.
784   bool CanBeInternalized =
785       !V.hasComdat() && !V.hasAppendingLinkage() && !V.isInterposable() &&
786       !V.hasAvailableExternallyLinkage() && !V.hasDLLExportStorageClass();
787   bool Constant = V.isConstant();
788   GlobalVarSummary::GVarFlags VarFlags(CanBeInternalized,
789                                        Constant ? false : CanBeInternalized,
790                                        Constant, V.getVCallVisibility());
791   auto GVarSummary = std::make_unique<GlobalVarSummary>(Flags, VarFlags,
792                                                          RefEdges.takeVector());
793   if (NonRenamableLocal)
794     CantBePromoted.insert(V.getGUID());
795   if (HasBlockAddress)
796     GVarSummary->setNotEligibleToImport();
797   if (!VTableFuncs.empty())
798     GVarSummary->setVTableFuncs(VTableFuncs);
799   Index.addGlobalValueSummary(V, std::move(GVarSummary));
800 }
801 
802 static void computeAliasSummary(ModuleSummaryIndex &Index, const GlobalAlias &A,
803                                 DenseSet<GlobalValue::GUID> &CantBePromoted) {
804   // Skip summary for indirect function aliases as summary for aliasee will not
805   // be emitted.
806   const GlobalObject *Aliasee = A.getAliaseeObject();
807   if (isa<GlobalIFunc>(Aliasee))
808     return;
809   bool NonRenamableLocal = isNonRenamableLocal(A);
810   GlobalValueSummary::GVFlags Flags(
811       A.getLinkage(), A.getVisibility(), NonRenamableLocal,
812       /* Live = */ false, A.isDSOLocal(), A.canBeOmittedFromSymbolTable(),
813       GlobalValueSummary::Definition);
814   auto AS = std::make_unique<AliasSummary>(Flags);
815   auto AliaseeVI = Index.getValueInfo(Aliasee->getGUID());
816   assert(AliaseeVI && "Alias expects aliasee summary to be available");
817   assert(AliaseeVI.getSummaryList().size() == 1 &&
818          "Expected a single entry per aliasee in per-module index");
819   AS->setAliasee(AliaseeVI, AliaseeVI.getSummaryList()[0].get());
820   if (NonRenamableLocal)
821     CantBePromoted.insert(A.getGUID());
822   Index.addGlobalValueSummary(A, std::move(AS));
823 }
824 
825 // Set LiveRoot flag on entries matching the given value name.
826 static void setLiveRoot(ModuleSummaryIndex &Index, StringRef Name) {
827   if (ValueInfo VI = Index.getValueInfo(GlobalValue::getGUID(Name)))
828     for (const auto &Summary : VI.getSummaryList())
829       Summary->setLive(true);
830 }
831 
832 ModuleSummaryIndex llvm::buildModuleSummaryIndex(
833     const Module &M,
834     std::function<BlockFrequencyInfo *(const Function &F)> GetBFICallback,
835     ProfileSummaryInfo *PSI,
836     std::function<const StackSafetyInfo *(const Function &F)> GetSSICallback) {
837   assert(PSI);
838   bool EnableSplitLTOUnit = false;
839   bool UnifiedLTO = false;
840   if (auto *MD = mdconst::extract_or_null<ConstantInt>(
841           M.getModuleFlag("EnableSplitLTOUnit")))
842     EnableSplitLTOUnit = MD->getZExtValue();
843   if (auto *MD =
844           mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("UnifiedLTO")))
845     UnifiedLTO = MD->getZExtValue();
846   ModuleSummaryIndex Index(/*HaveGVs=*/true, EnableSplitLTOUnit, UnifiedLTO);
847 
848   // Identify the local values in the llvm.used and llvm.compiler.used sets,
849   // which should not be exported as they would then require renaming and
850   // promotion, but we may have opaque uses e.g. in inline asm. We collect them
851   // here because we use this information to mark functions containing inline
852   // assembly calls as not importable.
853   SmallPtrSet<GlobalValue *, 4> LocalsUsed;
854   SmallVector<GlobalValue *, 4> Used;
855   // First collect those in the llvm.used set.
856   collectUsedGlobalVariables(M, Used, /*CompilerUsed=*/false);
857   // Next collect those in the llvm.compiler.used set.
858   collectUsedGlobalVariables(M, Used, /*CompilerUsed=*/true);
859   DenseSet<GlobalValue::GUID> CantBePromoted;
860   for (auto *V : Used) {
861     if (V->hasLocalLinkage()) {
862       LocalsUsed.insert(V);
863       CantBePromoted.insert(V->getGUID());
864     }
865   }
866 
867   bool HasLocalInlineAsmSymbol = false;
868   if (!M.getModuleInlineAsm().empty()) {
869     // Collect the local values defined by module level asm, and set up
870     // summaries for these symbols so that they can be marked as NoRename,
871     // to prevent export of any use of them in regular IR that would require
872     // renaming within the module level asm. Note we don't need to create a
873     // summary for weak or global defs, as they don't need to be flagged as
874     // NoRename, and defs in module level asm can't be imported anyway.
875     // Also, any values used but not defined within module level asm should
876     // be listed on the llvm.used or llvm.compiler.used global and marked as
877     // referenced from there.
878     ModuleSymbolTable::CollectAsmSymbols(
879         M, [&](StringRef Name, object::BasicSymbolRef::Flags Flags) {
880           // Symbols not marked as Weak or Global are local definitions.
881           if (Flags & (object::BasicSymbolRef::SF_Weak |
882                        object::BasicSymbolRef::SF_Global))
883             return;
884           HasLocalInlineAsmSymbol = true;
885           GlobalValue *GV = M.getNamedValue(Name);
886           if (!GV)
887             return;
888           assert(GV->isDeclaration() && "Def in module asm already has definition");
889           GlobalValueSummary::GVFlags GVFlags(
890               GlobalValue::InternalLinkage, GlobalValue::DefaultVisibility,
891               /* NotEligibleToImport = */ true,
892               /* Live = */ true,
893               /* Local */ GV->isDSOLocal(), GV->canBeOmittedFromSymbolTable(),
894               GlobalValueSummary::Definition);
895           CantBePromoted.insert(GV->getGUID());
896           // Create the appropriate summary type.
897           if (Function *F = dyn_cast<Function>(GV)) {
898             std::unique_ptr<FunctionSummary> Summary =
899                 std::make_unique<FunctionSummary>(
900                     GVFlags, /*InstCount=*/0,
901                     FunctionSummary::FFlags{
902                         F->hasFnAttribute(Attribute::ReadNone),
903                         F->hasFnAttribute(Attribute::ReadOnly),
904                         F->hasFnAttribute(Attribute::NoRecurse),
905                         F->returnDoesNotAlias(),
906                         /* NoInline = */ false,
907                         F->hasFnAttribute(Attribute::AlwaysInline),
908                         F->hasFnAttribute(Attribute::NoUnwind),
909                         /* MayThrow */ true,
910                         /* HasUnknownCall */ true,
911                         /* MustBeUnreachable */ false},
912                     /*EntryCount=*/0, ArrayRef<ValueInfo>{},
913                     ArrayRef<FunctionSummary::EdgeTy>{},
914                     ArrayRef<GlobalValue::GUID>{},
915                     ArrayRef<FunctionSummary::VFuncId>{},
916                     ArrayRef<FunctionSummary::VFuncId>{},
917                     ArrayRef<FunctionSummary::ConstVCall>{},
918                     ArrayRef<FunctionSummary::ConstVCall>{},
919                     ArrayRef<FunctionSummary::ParamAccess>{},
920                     ArrayRef<CallsiteInfo>{}, ArrayRef<AllocInfo>{});
921             Index.addGlobalValueSummary(*GV, std::move(Summary));
922           } else {
923             std::unique_ptr<GlobalVarSummary> Summary =
924                 std::make_unique<GlobalVarSummary>(
925                     GVFlags,
926                     GlobalVarSummary::GVarFlags(
927                         false, false, cast<GlobalVariable>(GV)->isConstant(),
928                         GlobalObject::VCallVisibilityPublic),
929                     ArrayRef<ValueInfo>{});
930             Index.addGlobalValueSummary(*GV, std::move(Summary));
931           }
932         });
933   }
934 
935   bool IsThinLTO = true;
936   if (auto *MD =
937           mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("ThinLTO")))
938     IsThinLTO = MD->getZExtValue();
939 
940   // Compute summaries for all functions defined in module, and save in the
941   // index.
942   for (const auto &F : M) {
943     if (F.isDeclaration())
944       continue;
945 
946     DominatorTree DT(const_cast<Function &>(F));
947     BlockFrequencyInfo *BFI = nullptr;
948     std::unique_ptr<BlockFrequencyInfo> BFIPtr;
949     if (GetBFICallback)
950       BFI = GetBFICallback(F);
951     else if (F.hasProfileData()) {
952       LoopInfo LI{DT};
953       BranchProbabilityInfo BPI{F, LI};
954       BFIPtr = std::make_unique<BlockFrequencyInfo>(F, BPI, LI);
955       BFI = BFIPtr.get();
956     }
957 
958     computeFunctionSummary(Index, M, F, BFI, PSI, DT,
959                            !LocalsUsed.empty() || HasLocalInlineAsmSymbol,
960                            CantBePromoted, IsThinLTO, GetSSICallback);
961   }
962 
963   // Compute summaries for all variables defined in module, and save in the
964   // index.
965   SmallVector<MDNode *, 2> Types;
966   for (const GlobalVariable &G : M.globals()) {
967     if (G.isDeclaration())
968       continue;
969     computeVariableSummary(Index, G, CantBePromoted, M, Types);
970   }
971 
972   // Compute summaries for all aliases defined in module, and save in the
973   // index.
974   for (const GlobalAlias &A : M.aliases())
975     computeAliasSummary(Index, A, CantBePromoted);
976 
977   // Iterate through ifuncs, set their resolvers all alive.
978   for (const GlobalIFunc &I : M.ifuncs()) {
979     I.applyAlongResolverPath([&Index](const GlobalValue &GV) {
980       Index.getGlobalValueSummary(GV)->setLive(true);
981     });
982   }
983 
984   for (auto *V : LocalsUsed) {
985     auto *Summary = Index.getGlobalValueSummary(*V);
986     assert(Summary && "Missing summary for global value");
987     Summary->setNotEligibleToImport();
988   }
989 
990   // The linker doesn't know about these LLVM produced values, so we need
991   // to flag them as live in the index to ensure index-based dead value
992   // analysis treats them as live roots of the analysis.
993   setLiveRoot(Index, "llvm.used");
994   setLiveRoot(Index, "llvm.compiler.used");
995   setLiveRoot(Index, "llvm.global_ctors");
996   setLiveRoot(Index, "llvm.global_dtors");
997   setLiveRoot(Index, "llvm.global.annotations");
998 
999   for (auto &GlobalList : Index) {
1000     // Ignore entries for references that are undefined in the current module.
1001     if (GlobalList.second.SummaryList.empty())
1002       continue;
1003 
1004     assert(GlobalList.second.SummaryList.size() == 1 &&
1005            "Expected module's index to have one summary per GUID");
1006     auto &Summary = GlobalList.second.SummaryList[0];
1007     if (!IsThinLTO) {
1008       Summary->setNotEligibleToImport();
1009       continue;
1010     }
1011 
1012     bool AllRefsCanBeExternallyReferenced =
1013         llvm::all_of(Summary->refs(), [&](const ValueInfo &VI) {
1014           return !CantBePromoted.count(VI.getGUID());
1015         });
1016     if (!AllRefsCanBeExternallyReferenced) {
1017       Summary->setNotEligibleToImport();
1018       continue;
1019     }
1020 
1021     if (auto *FuncSummary = dyn_cast<FunctionSummary>(Summary.get())) {
1022       bool AllCallsCanBeExternallyReferenced = llvm::all_of(
1023           FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) {
1024             return !CantBePromoted.count(Edge.first.getGUID());
1025           });
1026       if (!AllCallsCanBeExternallyReferenced)
1027         Summary->setNotEligibleToImport();
1028     }
1029   }
1030 
1031   if (!ModuleSummaryDotFile.empty()) {
1032     std::error_code EC;
1033     raw_fd_ostream OSDot(ModuleSummaryDotFile, EC, sys::fs::OpenFlags::OF_None);
1034     if (EC)
1035       report_fatal_error(Twine("Failed to open dot file ") +
1036                          ModuleSummaryDotFile + ": " + EC.message() + "\n");
1037     Index.exportToDot(OSDot, {});
1038   }
1039 
1040   return Index;
1041 }
1042 
1043 AnalysisKey ModuleSummaryIndexAnalysis::Key;
1044 
1045 ModuleSummaryIndex
1046 ModuleSummaryIndexAnalysis::run(Module &M, ModuleAnalysisManager &AM) {
1047   ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M);
1048   auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1049   bool NeedSSI = needsParamAccessSummary(M);
1050   return buildModuleSummaryIndex(
1051       M,
1052       [&FAM](const Function &F) {
1053         return &FAM.getResult<BlockFrequencyAnalysis>(
1054             *const_cast<Function *>(&F));
1055       },
1056       &PSI,
1057       [&FAM, NeedSSI](const Function &F) -> const StackSafetyInfo * {
1058         return NeedSSI ? &FAM.getResult<StackSafetyAnalysis>(
1059                              const_cast<Function &>(F))
1060                        : nullptr;
1061       });
1062 }
1063 
1064 char ModuleSummaryIndexWrapperPass::ID = 0;
1065 
1066 INITIALIZE_PASS_BEGIN(ModuleSummaryIndexWrapperPass, "module-summary-analysis",
1067                       "Module Summary Analysis", false, true)
1068 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
1069 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
1070 INITIALIZE_PASS_DEPENDENCY(StackSafetyInfoWrapperPass)
1071 INITIALIZE_PASS_END(ModuleSummaryIndexWrapperPass, "module-summary-analysis",
1072                     "Module Summary Analysis", false, true)
1073 
1074 ModulePass *llvm::createModuleSummaryIndexWrapperPass() {
1075   return new ModuleSummaryIndexWrapperPass();
1076 }
1077 
1078 ModuleSummaryIndexWrapperPass::ModuleSummaryIndexWrapperPass()
1079     : ModulePass(ID) {
1080   initializeModuleSummaryIndexWrapperPassPass(*PassRegistry::getPassRegistry());
1081 }
1082 
1083 bool ModuleSummaryIndexWrapperPass::runOnModule(Module &M) {
1084   auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
1085   bool NeedSSI = needsParamAccessSummary(M);
1086   Index.emplace(buildModuleSummaryIndex(
1087       M,
1088       [this](const Function &F) {
1089         return &(this->getAnalysis<BlockFrequencyInfoWrapperPass>(
1090                          *const_cast<Function *>(&F))
1091                      .getBFI());
1092       },
1093       PSI,
1094       [&](const Function &F) -> const StackSafetyInfo * {
1095         return NeedSSI ? &getAnalysis<StackSafetyInfoWrapperPass>(
1096                               const_cast<Function &>(F))
1097                               .getResult()
1098                        : nullptr;
1099       }));
1100   return false;
1101 }
1102 
1103 bool ModuleSummaryIndexWrapperPass::doFinalization(Module &M) {
1104   Index.reset();
1105   return false;
1106 }
1107 
1108 void ModuleSummaryIndexWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1109   AU.setPreservesAll();
1110   AU.addRequired<BlockFrequencyInfoWrapperPass>();
1111   AU.addRequired<ProfileSummaryInfoWrapperPass>();
1112   AU.addRequired<StackSafetyInfoWrapperPass>();
1113 }
1114 
1115 char ImmutableModuleSummaryIndexWrapperPass::ID = 0;
1116 
1117 ImmutableModuleSummaryIndexWrapperPass::ImmutableModuleSummaryIndexWrapperPass(
1118     const ModuleSummaryIndex *Index)
1119     : ImmutablePass(ID), Index(Index) {
1120   initializeImmutableModuleSummaryIndexWrapperPassPass(
1121       *PassRegistry::getPassRegistry());
1122 }
1123 
1124 void ImmutableModuleSummaryIndexWrapperPass::getAnalysisUsage(
1125     AnalysisUsage &AU) const {
1126   AU.setPreservesAll();
1127 }
1128 
1129 ImmutablePass *llvm::createImmutableModuleSummaryIndexWrapperPass(
1130     const ModuleSummaryIndex *Index) {
1131   return new ImmutableModuleSummaryIndexWrapperPass(Index);
1132 }
1133 
1134 INITIALIZE_PASS(ImmutableModuleSummaryIndexWrapperPass, "module-summary-info",
1135                 "Module summary info", false, true)
1136 
1137 bool llvm::mayHaveMemprofSummary(const CallBase *CB) {
1138   if (!CB)
1139     return false;
1140   if (CB->isDebugOrPseudoInst())
1141     return false;
1142   auto *CI = dyn_cast<CallInst>(CB);
1143   auto *CalledValue = CB->getCalledOperand();
1144   auto *CalledFunction = CB->getCalledFunction();
1145   if (CalledValue && !CalledFunction) {
1146     CalledValue = CalledValue->stripPointerCasts();
1147     // Stripping pointer casts can reveal a called function.
1148     CalledFunction = dyn_cast<Function>(CalledValue);
1149   }
1150   // Check if this is an alias to a function. If so, get the
1151   // called aliasee for the checks below.
1152   if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) {
1153     assert(!CalledFunction &&
1154            "Expected null called function in callsite for alias");
1155     CalledFunction = dyn_cast<Function>(GA->getAliaseeObject());
1156   }
1157   // Check if this is a direct call to a known function or a known
1158   // intrinsic, or an indirect call with profile data.
1159   if (CalledFunction) {
1160     if (CI && CalledFunction->isIntrinsic())
1161       return false;
1162   } else {
1163     // TODO: For now skip indirect calls. See comments in
1164     // computeFunctionSummary for what is needed to handle this.
1165     return false;
1166   }
1167   return true;
1168 }
1169