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/ConstantFolding.h"
26 #include "llvm/Analysis/IndirectCallPromotionAnalysis.h"
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/Analysis/MemoryProfileInfo.h"
29 #include "llvm/Analysis/ProfileSummaryInfo.h"
30 #include "llvm/Analysis/StackSafetyAnalysis.h"
31 #include "llvm/Analysis/TypeMetadataUtils.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/Constant.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/GlobalAlias.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/GlobalVariable.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/Metadata.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/IR/ModuleSummaryIndex.h"
46 #include "llvm/IR/Use.h"
47 #include "llvm/IR/User.h"
48 #include "llvm/InitializePasses.h"
49 #include "llvm/Object/ModuleSymbolTable.h"
50 #include "llvm/Object/SymbolicFile.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/Casting.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/FileSystem.h"
55 #include <algorithm>
56 #include <cassert>
57 #include <cstdint>
58 #include <vector>
59
60 using namespace llvm;
61 using namespace llvm::memprof;
62
63 #define DEBUG_TYPE "module-summary-analysis"
64
65 // Option to force edges cold which will block importing when the
66 // -import-cold-multiplier is set to 0. Useful for debugging.
67 namespace llvm {
68 FunctionSummary::ForceSummaryHotnessType ForceSummaryEdgesCold =
69 FunctionSummary::FSHT_None;
70 } // namespace llvm
71
72 static cl::opt<FunctionSummary::ForceSummaryHotnessType, true> FSEC(
73 "force-summary-edges-cold", cl::Hidden, cl::location(ForceSummaryEdgesCold),
74 cl::desc("Force all edges in the function summary to cold"),
75 cl::values(clEnumValN(FunctionSummary::FSHT_None, "none", "None."),
76 clEnumValN(FunctionSummary::FSHT_AllNonCritical,
77 "all-non-critical", "All non-critical edges."),
78 clEnumValN(FunctionSummary::FSHT_All, "all", "All edges.")));
79
80 static cl::opt<std::string> ModuleSummaryDotFile(
81 "module-summary-dot-file", cl::Hidden, cl::value_desc("filename"),
82 cl::desc("File to emit dot graph of new summary into"));
83
84 extern cl::opt<bool> ScalePartialSampleProfileWorkingSetSize;
85
86 extern cl::opt<unsigned> MaxNumVTableAnnotations;
87
88 extern cl::opt<bool> MemProfReportHintedSizes;
89
90 // Walk through the operands of a given User via worklist iteration and populate
91 // the set of GlobalValue references encountered. Invoked either on an
92 // Instruction or a GlobalVariable (which walks its initializer).
93 // Return true if any of the operands contains blockaddress. This is important
94 // to know when computing summary for global var, because if global variable
95 // references basic block address we can't import it separately from function
96 // containing that basic block. For simplicity we currently don't import such
97 // global vars at all. When importing function we aren't interested if any
98 // instruction in it takes an address of any basic block, because instruction
99 // can only take an address of basic block located in the same function.
100 // Set `RefLocalLinkageIFunc` to true if the analyzed value references a
101 // local-linkage ifunc.
findRefEdges(ModuleSummaryIndex & Index,const User * CurUser,SetVector<ValueInfo,std::vector<ValueInfo>> & RefEdges,SmallPtrSet<const User *,8> & Visited,bool & RefLocalLinkageIFunc)102 static bool findRefEdges(ModuleSummaryIndex &Index, const User *CurUser,
103 SetVector<ValueInfo, std::vector<ValueInfo>> &RefEdges,
104 SmallPtrSet<const User *, 8> &Visited,
105 bool &RefLocalLinkageIFunc) {
106 bool HasBlockAddress = false;
107 SmallVector<const User *, 32> Worklist;
108 if (Visited.insert(CurUser).second)
109 Worklist.push_back(CurUser);
110
111 while (!Worklist.empty()) {
112 const User *U = Worklist.pop_back_val();
113 const auto *CB = dyn_cast<CallBase>(U);
114
115 for (const auto &OI : U->operands()) {
116 const User *Operand = dyn_cast<User>(OI);
117 if (!Operand)
118 continue;
119 if (isa<BlockAddress>(Operand)) {
120 HasBlockAddress = true;
121 continue;
122 }
123 if (auto *GV = dyn_cast<GlobalValue>(Operand)) {
124 // We have a reference to a global value. This should be added to
125 // the reference set unless it is a callee. Callees are handled
126 // specially by WriteFunction and are added to a separate list.
127 if (!(CB && CB->isCallee(&OI))) {
128 // If an ifunc has local linkage, do not add it into ref edges, and
129 // sets `RefLocalLinkageIFunc` to true. The referencer is not eligible
130 // for import. An ifunc doesn't have summary and ThinLTO cannot
131 // promote it; importing the referencer may cause linkage errors.
132 if (auto *GI = dyn_cast_if_present<GlobalIFunc>(GV);
133 GI && GI->hasLocalLinkage()) {
134 RefLocalLinkageIFunc = true;
135 continue;
136 }
137 RefEdges.insert(Index.getOrInsertValueInfo(GV));
138 }
139 continue;
140 }
141 if (Visited.insert(Operand).second)
142 Worklist.push_back(Operand);
143 }
144 }
145
146 const Instruction *I = dyn_cast<Instruction>(CurUser);
147 if (I) {
148 uint64_t TotalCount = 0;
149 // MaxNumVTableAnnotations is the maximum number of vtables annotated on
150 // the instruction.
151 auto ValueDataArray = getValueProfDataFromInst(
152 *I, IPVK_VTableTarget, MaxNumVTableAnnotations, TotalCount);
153
154 for (const auto &V : ValueDataArray)
155 RefEdges.insert(Index.getOrInsertValueInfo(/* VTableGUID = */
156 V.Value));
157 }
158 return HasBlockAddress;
159 }
160
getHotness(uint64_t ProfileCount,ProfileSummaryInfo * PSI)161 static CalleeInfo::HotnessType getHotness(uint64_t ProfileCount,
162 ProfileSummaryInfo *PSI) {
163 if (!PSI)
164 return CalleeInfo::HotnessType::Unknown;
165 if (PSI->isHotCount(ProfileCount))
166 return CalleeInfo::HotnessType::Hot;
167 if (PSI->isColdCount(ProfileCount))
168 return CalleeInfo::HotnessType::Cold;
169 return CalleeInfo::HotnessType::None;
170 }
171
isNonRenamableLocal(const GlobalValue & GV)172 static bool isNonRenamableLocal(const GlobalValue &GV) {
173 return GV.hasSection() && GV.hasLocalLinkage();
174 }
175
176 /// Determine whether this call has all constant integer arguments (excluding
177 /// "this") and summarize it to VCalls or ConstVCalls as appropriate.
addVCallToSet(DevirtCallSite Call,GlobalValue::GUID Guid,SetVector<FunctionSummary::VFuncId,std::vector<FunctionSummary::VFuncId>> & VCalls,SetVector<FunctionSummary::ConstVCall,std::vector<FunctionSummary::ConstVCall>> & ConstVCalls)178 static void addVCallToSet(
179 DevirtCallSite Call, GlobalValue::GUID Guid,
180 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
181 &VCalls,
182 SetVector<FunctionSummary::ConstVCall,
183 std::vector<FunctionSummary::ConstVCall>> &ConstVCalls) {
184 std::vector<uint64_t> Args;
185 // Start from the second argument to skip the "this" pointer.
186 for (auto &Arg : drop_begin(Call.CB.args())) {
187 auto *CI = dyn_cast<ConstantInt>(Arg);
188 if (!CI || CI->getBitWidth() > 64) {
189 VCalls.insert({Guid, Call.Offset});
190 return;
191 }
192 Args.push_back(CI->getZExtValue());
193 }
194 ConstVCalls.insert({{Guid, Call.Offset}, std::move(Args)});
195 }
196
197 /// If this intrinsic call requires that we add information to the function
198 /// summary, do so via the non-constant reference arguments.
addIntrinsicToSummary(const CallInst * CI,SetVector<GlobalValue::GUID,std::vector<GlobalValue::GUID>> & TypeTests,SetVector<FunctionSummary::VFuncId,std::vector<FunctionSummary::VFuncId>> & TypeTestAssumeVCalls,SetVector<FunctionSummary::VFuncId,std::vector<FunctionSummary::VFuncId>> & TypeCheckedLoadVCalls,SetVector<FunctionSummary::ConstVCall,std::vector<FunctionSummary::ConstVCall>> & TypeTestAssumeConstVCalls,SetVector<FunctionSummary::ConstVCall,std::vector<FunctionSummary::ConstVCall>> & TypeCheckedLoadConstVCalls,DominatorTree & DT)199 static void addIntrinsicToSummary(
200 const CallInst *CI,
201 SetVector<GlobalValue::GUID, std::vector<GlobalValue::GUID>> &TypeTests,
202 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
203 &TypeTestAssumeVCalls,
204 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
205 &TypeCheckedLoadVCalls,
206 SetVector<FunctionSummary::ConstVCall,
207 std::vector<FunctionSummary::ConstVCall>>
208 &TypeTestAssumeConstVCalls,
209 SetVector<FunctionSummary::ConstVCall,
210 std::vector<FunctionSummary::ConstVCall>>
211 &TypeCheckedLoadConstVCalls,
212 DominatorTree &DT) {
213 switch (CI->getCalledFunction()->getIntrinsicID()) {
214 case Intrinsic::type_test:
215 case Intrinsic::public_type_test: {
216 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(1));
217 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata());
218 if (!TypeId)
219 break;
220 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString());
221
222 // Produce a summary from type.test intrinsics. We only summarize type.test
223 // intrinsics that are used other than by an llvm.assume intrinsic.
224 // Intrinsics that are assumed are relevant only to the devirtualization
225 // pass, not the type test lowering pass.
226 bool HasNonAssumeUses = llvm::any_of(CI->uses(), [](const Use &CIU) {
227 return !isa<AssumeInst>(CIU.getUser());
228 });
229 if (HasNonAssumeUses)
230 TypeTests.insert(Guid);
231
232 SmallVector<DevirtCallSite, 4> DevirtCalls;
233 SmallVector<CallInst *, 4> Assumes;
234 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT);
235 for (auto &Call : DevirtCalls)
236 addVCallToSet(Call, Guid, TypeTestAssumeVCalls,
237 TypeTestAssumeConstVCalls);
238
239 break;
240 }
241
242 case Intrinsic::type_checked_load_relative:
243 case Intrinsic::type_checked_load: {
244 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(2));
245 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata());
246 if (!TypeId)
247 break;
248 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString());
249
250 SmallVector<DevirtCallSite, 4> DevirtCalls;
251 SmallVector<Instruction *, 4> LoadedPtrs;
252 SmallVector<Instruction *, 4> Preds;
253 bool HasNonCallUses = false;
254 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
255 HasNonCallUses, CI, DT);
256 // Any non-call uses of the result of llvm.type.checked.load will
257 // prevent us from optimizing away the llvm.type.test.
258 if (HasNonCallUses)
259 TypeTests.insert(Guid);
260 for (auto &Call : DevirtCalls)
261 addVCallToSet(Call, Guid, TypeCheckedLoadVCalls,
262 TypeCheckedLoadConstVCalls);
263
264 break;
265 }
266 default:
267 break;
268 }
269 }
270
isNonVolatileLoad(const Instruction * I)271 static bool isNonVolatileLoad(const Instruction *I) {
272 if (const auto *LI = dyn_cast<LoadInst>(I))
273 return !LI->isVolatile();
274
275 return false;
276 }
277
isNonVolatileStore(const Instruction * I)278 static bool isNonVolatileStore(const Instruction *I) {
279 if (const auto *SI = dyn_cast<StoreInst>(I))
280 return !SI->isVolatile();
281
282 return false;
283 }
284
285 // Returns true if the function definition must be unreachable.
286 //
287 // Note if this helper function returns true, `F` is guaranteed
288 // to be unreachable; if it returns false, `F` might still
289 // be unreachable but not covered by this helper function.
mustBeUnreachableFunction(const Function & F)290 static bool mustBeUnreachableFunction(const Function &F) {
291 // A function must be unreachable if its entry block ends with an
292 // 'unreachable'.
293 assert(!F.isDeclaration());
294 return isa<UnreachableInst>(F.getEntryBlock().getTerminator());
295 }
296
computeFunctionSummary(ModuleSummaryIndex & Index,const Module & M,const Function & F,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI,DominatorTree & DT,bool HasLocalsInUsedOrAsm,DenseSet<GlobalValue::GUID> & CantBePromoted,bool IsThinLTO,std::function<const StackSafetyInfo * (const Function & F)> GetSSICallback)297 static void computeFunctionSummary(
298 ModuleSummaryIndex &Index, const Module &M, const Function &F,
299 BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, DominatorTree &DT,
300 bool HasLocalsInUsedOrAsm, DenseSet<GlobalValue::GUID> &CantBePromoted,
301 bool IsThinLTO,
302 std::function<const StackSafetyInfo *(const Function &F)> GetSSICallback) {
303 // Summary not currently supported for anonymous functions, they should
304 // have been named.
305 assert(F.hasName());
306
307 unsigned NumInsts = 0;
308 // Map from callee ValueId to profile count. Used to accumulate profile
309 // counts for all static calls to a given callee.
310 MapVector<ValueInfo, CalleeInfo, DenseMap<ValueInfo, unsigned>,
311 std::vector<std::pair<ValueInfo, CalleeInfo>>>
312 CallGraphEdges;
313 SetVector<ValueInfo, std::vector<ValueInfo>> RefEdges, LoadRefEdges,
314 StoreRefEdges;
315 SetVector<GlobalValue::GUID, std::vector<GlobalValue::GUID>> TypeTests;
316 SetVector<FunctionSummary::VFuncId, std::vector<FunctionSummary::VFuncId>>
317 TypeTestAssumeVCalls, TypeCheckedLoadVCalls;
318 SetVector<FunctionSummary::ConstVCall,
319 std::vector<FunctionSummary::ConstVCall>>
320 TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls;
321 ICallPromotionAnalysis ICallAnalysis;
322 SmallPtrSet<const User *, 8> Visited;
323
324 // Add personality function, prefix data and prologue data to function's ref
325 // list.
326 bool HasLocalIFuncCallOrRef = false;
327 findRefEdges(Index, &F, RefEdges, Visited, HasLocalIFuncCallOrRef);
328 std::vector<const Instruction *> NonVolatileLoads;
329 std::vector<const Instruction *> NonVolatileStores;
330
331 std::vector<CallsiteInfo> Callsites;
332 std::vector<AllocInfo> Allocs;
333
334 #ifndef NDEBUG
335 DenseSet<const CallBase *> CallsThatMayHaveMemprofSummary;
336 #endif
337
338 bool HasInlineAsmMaybeReferencingInternal = false;
339 bool HasIndirBranchToBlockAddress = false;
340 bool HasUnknownCall = false;
341 bool MayThrow = false;
342 for (const BasicBlock &BB : F) {
343 // We don't allow inlining of function with indirect branch to blockaddress.
344 // If the blockaddress escapes the function, e.g., via a global variable,
345 // inlining may lead to an invalid cross-function reference. So we shouldn't
346 // import such function either.
347 if (BB.hasAddressTaken()) {
348 for (User *U : BlockAddress::get(const_cast<BasicBlock *>(&BB))->users())
349 if (!isa<CallBrInst>(*U)) {
350 HasIndirBranchToBlockAddress = true;
351 break;
352 }
353 }
354
355 for (const Instruction &I : BB) {
356 if (I.isDebugOrPseudoInst())
357 continue;
358 ++NumInsts;
359
360 // Regular LTO module doesn't participate in ThinLTO import,
361 // so no reference from it can be read/writeonly, since this
362 // would require importing variable as local copy
363 if (IsThinLTO) {
364 if (isNonVolatileLoad(&I)) {
365 // Postpone processing of non-volatile load instructions
366 // See comments below
367 Visited.insert(&I);
368 NonVolatileLoads.push_back(&I);
369 continue;
370 } else if (isNonVolatileStore(&I)) {
371 Visited.insert(&I);
372 NonVolatileStores.push_back(&I);
373 // All references from second operand of store (destination address)
374 // can be considered write-only if they're not referenced by any
375 // non-store instruction. References from first operand of store
376 // (stored value) can't be treated either as read- or as write-only
377 // so we add them to RefEdges as we do with all other instructions
378 // except non-volatile load.
379 Value *Stored = I.getOperand(0);
380 if (auto *GV = dyn_cast<GlobalValue>(Stored))
381 // findRefEdges will try to examine GV operands, so instead
382 // of calling it we should add GV to RefEdges directly.
383 RefEdges.insert(Index.getOrInsertValueInfo(GV));
384 else if (auto *U = dyn_cast<User>(Stored))
385 findRefEdges(Index, U, RefEdges, Visited, HasLocalIFuncCallOrRef);
386 continue;
387 }
388 }
389 findRefEdges(Index, &I, RefEdges, Visited, HasLocalIFuncCallOrRef);
390 const auto *CB = dyn_cast<CallBase>(&I);
391 if (!CB) {
392 if (I.mayThrow())
393 MayThrow = true;
394 continue;
395 }
396
397 const auto *CI = dyn_cast<CallInst>(&I);
398 // Since we don't know exactly which local values are referenced in inline
399 // assembly, conservatively mark the function as possibly referencing
400 // a local value from inline assembly to ensure we don't export a
401 // reference (which would require renaming and promotion of the
402 // referenced value).
403 if (HasLocalsInUsedOrAsm && CI && CI->isInlineAsm())
404 HasInlineAsmMaybeReferencingInternal = true;
405
406 auto *CalledValue = CB->getCalledOperand();
407 auto *CalledFunction = CB->getCalledFunction();
408 if (CalledValue && !CalledFunction) {
409 CalledValue = CalledValue->stripPointerCasts();
410 // Stripping pointer casts can reveal a called function.
411 CalledFunction = dyn_cast<Function>(CalledValue);
412 }
413 // Check if this is an alias to a function. If so, get the
414 // called aliasee for the checks below.
415 if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) {
416 assert(!CalledFunction && "Expected null called function in callsite for alias");
417 CalledFunction = dyn_cast<Function>(GA->getAliaseeObject());
418 }
419 // Check if this is a direct call to a known function or a known
420 // intrinsic, or an indirect call with profile data.
421 if (CalledFunction) {
422 if (CI && CalledFunction->isIntrinsic()) {
423 addIntrinsicToSummary(
424 CI, TypeTests, TypeTestAssumeVCalls, TypeCheckedLoadVCalls,
425 TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls, DT);
426 continue;
427 }
428 // We should have named any anonymous globals
429 assert(CalledFunction->hasName());
430 auto ScaledCount = PSI->getProfileCount(*CB, BFI);
431 auto Hotness = ScaledCount ? getHotness(*ScaledCount, PSI)
432 : CalleeInfo::HotnessType::Unknown;
433 if (ForceSummaryEdgesCold != FunctionSummary::FSHT_None)
434 Hotness = CalleeInfo::HotnessType::Cold;
435
436 // Use the original CalledValue, in case it was an alias. We want
437 // to record the call edge to the alias in that case. Eventually
438 // an alias summary will be created to associate the alias and
439 // aliasee.
440 auto &ValueInfo = CallGraphEdges[Index.getOrInsertValueInfo(
441 cast<GlobalValue>(CalledValue))];
442 ValueInfo.updateHotness(Hotness);
443 if (CB->isTailCall())
444 ValueInfo.setHasTailCall(true);
445 // Add the relative block frequency to CalleeInfo if there is no profile
446 // information.
447 if (BFI != nullptr && Hotness == CalleeInfo::HotnessType::Unknown) {
448 uint64_t BBFreq = BFI->getBlockFreq(&BB).getFrequency();
449 uint64_t EntryFreq = BFI->getEntryFreq().getFrequency();
450 ValueInfo.updateRelBlockFreq(BBFreq, EntryFreq);
451 }
452 } else {
453 HasUnknownCall = true;
454 // If F is imported, a local linkage ifunc (e.g. target_clones on a
455 // static function) called by F will be cloned. Since summaries don't
456 // track ifunc, we do not know implementation functions referenced by
457 // the ifunc resolver need to be promoted in the exporter, and we will
458 // get linker errors due to cloned declarations for implementation
459 // functions. As a simple fix, just mark F as not eligible for import.
460 // Non-local ifunc is not cloned and does not have the issue.
461 if (auto *GI = dyn_cast_if_present<GlobalIFunc>(CalledValue))
462 if (GI->hasLocalLinkage())
463 HasLocalIFuncCallOrRef = true;
464 // Skip inline assembly calls.
465 if (CI && CI->isInlineAsm())
466 continue;
467 // Skip direct calls.
468 if (!CalledValue || isa<Constant>(CalledValue))
469 continue;
470
471 // Check if the instruction has a callees metadata. If so, add callees
472 // to CallGraphEdges to reflect the references from the metadata, and
473 // to enable importing for subsequent indirect call promotion and
474 // inlining.
475 if (auto *MD = I.getMetadata(LLVMContext::MD_callees)) {
476 for (const auto &Op : MD->operands()) {
477 Function *Callee = mdconst::extract_or_null<Function>(Op);
478 if (Callee)
479 CallGraphEdges[Index.getOrInsertValueInfo(Callee)];
480 }
481 }
482
483 uint32_t NumCandidates;
484 uint64_t TotalCount;
485 auto CandidateProfileData =
486 ICallAnalysis.getPromotionCandidatesForInstruction(&I, TotalCount,
487 NumCandidates);
488 for (const auto &Candidate : CandidateProfileData)
489 CallGraphEdges[Index.getOrInsertValueInfo(Candidate.Value)]
490 .updateHotness(getHotness(Candidate.Count, PSI));
491 }
492
493 // Summarize memprof related metadata. This is only needed for ThinLTO.
494 if (!IsThinLTO)
495 continue;
496
497 // TODO: Skip indirect calls for now. Need to handle these better, likely
498 // by creating multiple Callsites, one per target, then speculatively
499 // devirtualize while applying clone info in the ThinLTO backends. This
500 // will also be important because we will have a different set of clone
501 // versions per target. This handling needs to match that in the ThinLTO
502 // backend so we handle things consistently for matching of callsite
503 // summaries to instructions.
504 if (!CalledFunction)
505 continue;
506
507 // Ensure we keep this analysis in sync with the handling in the ThinLTO
508 // backend (see MemProfContextDisambiguation::applyImport). Save this call
509 // so that we can skip it in checking the reverse case later.
510 assert(mayHaveMemprofSummary(CB));
511 #ifndef NDEBUG
512 CallsThatMayHaveMemprofSummary.insert(CB);
513 #endif
514
515 // Compute the list of stack ids first (so we can trim them from the stack
516 // ids on any MIBs).
517 CallStack<MDNode, MDNode::op_iterator> InstCallsite(
518 I.getMetadata(LLVMContext::MD_callsite));
519 auto *MemProfMD = I.getMetadata(LLVMContext::MD_memprof);
520 if (MemProfMD) {
521 std::vector<MIBInfo> MIBs;
522 std::vector<uint64_t> TotalSizes;
523 for (auto &MDOp : MemProfMD->operands()) {
524 auto *MIBMD = cast<const MDNode>(MDOp);
525 MDNode *StackNode = getMIBStackNode(MIBMD);
526 assert(StackNode);
527 SmallVector<unsigned> StackIdIndices;
528 CallStack<MDNode, MDNode::op_iterator> StackContext(StackNode);
529 // Collapse out any on the allocation call (inlining).
530 for (auto ContextIter =
531 StackContext.beginAfterSharedPrefix(InstCallsite);
532 ContextIter != StackContext.end(); ++ContextIter) {
533 unsigned StackIdIdx = Index.addOrGetStackIdIndex(*ContextIter);
534 // If this is a direct recursion, simply skip the duplicate
535 // entries. If this is mutual recursion, handling is left to
536 // the LTO link analysis client.
537 if (StackIdIndices.empty() || StackIdIndices.back() != StackIdIdx)
538 StackIdIndices.push_back(StackIdIdx);
539 }
540 MIBs.push_back(
541 MIBInfo(getMIBAllocType(MIBMD), std::move(StackIdIndices)));
542 if (MemProfReportHintedSizes) {
543 auto TotalSize = getMIBTotalSize(MIBMD);
544 assert(TotalSize);
545 TotalSizes.push_back(TotalSize);
546 }
547 }
548 Allocs.push_back(AllocInfo(std::move(MIBs)));
549 if (MemProfReportHintedSizes) {
550 assert(Allocs.back().MIBs.size() == TotalSizes.size());
551 Allocs.back().TotalSizes = std::move(TotalSizes);
552 }
553 } else if (!InstCallsite.empty()) {
554 SmallVector<unsigned> StackIdIndices;
555 for (auto StackId : InstCallsite)
556 StackIdIndices.push_back(Index.addOrGetStackIdIndex(StackId));
557 // Use the original CalledValue, in case it was an alias. We want
558 // to record the call edge to the alias in that case. Eventually
559 // an alias summary will be created to associate the alias and
560 // aliasee.
561 auto CalleeValueInfo =
562 Index.getOrInsertValueInfo(cast<GlobalValue>(CalledValue));
563 Callsites.push_back({CalleeValueInfo, StackIdIndices});
564 }
565 }
566 }
567
568 if (PSI->hasPartialSampleProfile() && ScalePartialSampleProfileWorkingSetSize)
569 Index.addBlockCount(F.size());
570
571 std::vector<ValueInfo> Refs;
572 if (IsThinLTO) {
573 auto AddRefEdges = [&](const std::vector<const Instruction *> &Instrs,
574 SetVector<ValueInfo, std::vector<ValueInfo>> &Edges,
575 SmallPtrSet<const User *, 8> &Cache) {
576 for (const auto *I : Instrs) {
577 Cache.erase(I);
578 findRefEdges(Index, I, Edges, Cache, HasLocalIFuncCallOrRef);
579 }
580 };
581
582 // By now we processed all instructions in a function, except
583 // non-volatile loads and non-volatile value stores. Let's find
584 // ref edges for both of instruction sets
585 AddRefEdges(NonVolatileLoads, LoadRefEdges, Visited);
586 // We can add some values to the Visited set when processing load
587 // instructions which are also used by stores in NonVolatileStores.
588 // For example this can happen if we have following code:
589 //
590 // store %Derived* @foo, %Derived** bitcast (%Base** @bar to %Derived**)
591 // %42 = load %Derived*, %Derived** bitcast (%Base** @bar to %Derived**)
592 //
593 // After processing loads we'll add bitcast to the Visited set, and if
594 // we use the same set while processing stores, we'll never see store
595 // to @bar and @bar will be mistakenly treated as readonly.
596 SmallPtrSet<const llvm::User *, 8> StoreCache;
597 AddRefEdges(NonVolatileStores, StoreRefEdges, StoreCache);
598
599 // If both load and store instruction reference the same variable
600 // we won't be able to optimize it. Add all such reference edges
601 // to RefEdges set.
602 for (const auto &VI : StoreRefEdges)
603 if (LoadRefEdges.remove(VI))
604 RefEdges.insert(VI);
605
606 unsigned RefCnt = RefEdges.size();
607 // All new reference edges inserted in two loops below are either
608 // read or write only. They will be grouped in the end of RefEdges
609 // vector, so we can use a single integer value to identify them.
610 for (const auto &VI : LoadRefEdges)
611 RefEdges.insert(VI);
612
613 unsigned FirstWORef = RefEdges.size();
614 for (const auto &VI : StoreRefEdges)
615 RefEdges.insert(VI);
616
617 Refs = RefEdges.takeVector();
618 for (; RefCnt < FirstWORef; ++RefCnt)
619 Refs[RefCnt].setReadOnly();
620
621 for (; RefCnt < Refs.size(); ++RefCnt)
622 Refs[RefCnt].setWriteOnly();
623 } else {
624 Refs = RefEdges.takeVector();
625 }
626 // Explicit add hot edges to enforce importing for designated GUIDs for
627 // sample PGO, to enable the same inlines as the profiled optimized binary.
628 for (auto &I : F.getImportGUIDs())
629 CallGraphEdges[Index.getOrInsertValueInfo(I)].updateHotness(
630 ForceSummaryEdgesCold == FunctionSummary::FSHT_All
631 ? CalleeInfo::HotnessType::Cold
632 : CalleeInfo::HotnessType::Critical);
633
634 #ifndef NDEBUG
635 // Make sure that all calls we decided could not have memprof summaries get a
636 // false value for mayHaveMemprofSummary, to ensure that this handling remains
637 // in sync with the ThinLTO backend handling.
638 if (IsThinLTO) {
639 for (const BasicBlock &BB : F) {
640 for (const Instruction &I : BB) {
641 const auto *CB = dyn_cast<CallBase>(&I);
642 if (!CB)
643 continue;
644 // We already checked these above.
645 if (CallsThatMayHaveMemprofSummary.count(CB))
646 continue;
647 assert(!mayHaveMemprofSummary(CB));
648 }
649 }
650 }
651 #endif
652
653 bool NonRenamableLocal = isNonRenamableLocal(F);
654 bool NotEligibleForImport =
655 NonRenamableLocal || HasInlineAsmMaybeReferencingInternal ||
656 HasIndirBranchToBlockAddress || HasLocalIFuncCallOrRef;
657 GlobalValueSummary::GVFlags Flags(
658 F.getLinkage(), F.getVisibility(), NotEligibleForImport,
659 /* Live = */ false, F.isDSOLocal(), F.canBeOmittedFromSymbolTable(),
660 GlobalValueSummary::ImportKind::Definition);
661 FunctionSummary::FFlags FunFlags{
662 F.doesNotAccessMemory(), F.onlyReadsMemory() && !F.doesNotAccessMemory(),
663 F.hasFnAttribute(Attribute::NoRecurse), F.returnDoesNotAlias(),
664 // FIXME: refactor this to use the same code that inliner is using.
665 // Don't try to import functions with noinline attribute.
666 F.getAttributes().hasFnAttr(Attribute::NoInline),
667 F.hasFnAttribute(Attribute::AlwaysInline),
668 F.hasFnAttribute(Attribute::NoUnwind), MayThrow, HasUnknownCall,
669 mustBeUnreachableFunction(F)};
670 std::vector<FunctionSummary::ParamAccess> ParamAccesses;
671 if (auto *SSI = GetSSICallback(F))
672 ParamAccesses = SSI->getParamAccesses(Index);
673 auto FuncSummary = std::make_unique<FunctionSummary>(
674 Flags, NumInsts, FunFlags, /*EntryCount=*/0, std::move(Refs),
675 CallGraphEdges.takeVector(), TypeTests.takeVector(),
676 TypeTestAssumeVCalls.takeVector(), TypeCheckedLoadVCalls.takeVector(),
677 TypeTestAssumeConstVCalls.takeVector(),
678 TypeCheckedLoadConstVCalls.takeVector(), std::move(ParamAccesses),
679 std::move(Callsites), std::move(Allocs));
680 if (NonRenamableLocal)
681 CantBePromoted.insert(F.getGUID());
682 Index.addGlobalValueSummary(F, std::move(FuncSummary));
683 }
684
685 /// Find function pointers referenced within the given vtable initializer
686 /// (or subset of an initializer) \p I. The starting offset of \p I within
687 /// the vtable initializer is \p StartingOffset. Any discovered function
688 /// pointers are added to \p VTableFuncs along with their cumulative offset
689 /// within the initializer.
findFuncPointers(const Constant * I,uint64_t StartingOffset,const Module & M,ModuleSummaryIndex & Index,VTableFuncList & VTableFuncs,const GlobalVariable & OrigGV)690 static void findFuncPointers(const Constant *I, uint64_t StartingOffset,
691 const Module &M, ModuleSummaryIndex &Index,
692 VTableFuncList &VTableFuncs,
693 const GlobalVariable &OrigGV) {
694 // First check if this is a function pointer.
695 if (I->getType()->isPointerTy()) {
696 auto C = I->stripPointerCasts();
697 auto A = dyn_cast<GlobalAlias>(C);
698 if (isa<Function>(C) || (A && isa<Function>(A->getAliasee()))) {
699 auto GV = dyn_cast<GlobalValue>(C);
700 assert(GV);
701 // We can disregard __cxa_pure_virtual as a possible call target, as
702 // calls to pure virtuals are UB.
703 if (GV && GV->getName() != "__cxa_pure_virtual")
704 VTableFuncs.push_back({Index.getOrInsertValueInfo(GV), StartingOffset});
705 return;
706 }
707 }
708
709 // Walk through the elements in the constant struct or array and recursively
710 // look for virtual function pointers.
711 const DataLayout &DL = M.getDataLayout();
712 if (auto *C = dyn_cast<ConstantStruct>(I)) {
713 StructType *STy = dyn_cast<StructType>(C->getType());
714 assert(STy);
715 const StructLayout *SL = DL.getStructLayout(C->getType());
716
717 for (auto EI : llvm::enumerate(STy->elements())) {
718 auto Offset = SL->getElementOffset(EI.index());
719 unsigned Op = SL->getElementContainingOffset(Offset);
720 findFuncPointers(cast<Constant>(I->getOperand(Op)),
721 StartingOffset + Offset, M, Index, VTableFuncs, OrigGV);
722 }
723 } else if (auto *C = dyn_cast<ConstantArray>(I)) {
724 ArrayType *ATy = C->getType();
725 Type *EltTy = ATy->getElementType();
726 uint64_t EltSize = DL.getTypeAllocSize(EltTy);
727 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) {
728 findFuncPointers(cast<Constant>(I->getOperand(i)),
729 StartingOffset + i * EltSize, M, Index, VTableFuncs,
730 OrigGV);
731 }
732 } else if (const auto *CE = dyn_cast<ConstantExpr>(I)) {
733 // For relative vtables, the next sub-component should be a trunc.
734 if (CE->getOpcode() != Instruction::Trunc ||
735 !(CE = dyn_cast<ConstantExpr>(CE->getOperand(0))))
736 return;
737
738 // If this constant can be reduced to the offset between a function and a
739 // global, then we know this is a valid virtual function if the RHS is the
740 // original vtable we're scanning through.
741 if (CE->getOpcode() == Instruction::Sub) {
742 GlobalValue *LHS, *RHS;
743 APSInt LHSOffset, RHSOffset;
744 if (IsConstantOffsetFromGlobal(CE->getOperand(0), LHS, LHSOffset, DL) &&
745 IsConstantOffsetFromGlobal(CE->getOperand(1), RHS, RHSOffset, DL) &&
746 RHS == &OrigGV &&
747
748 // For relative vtables, this component should point to the callable
749 // function without any offsets.
750 LHSOffset == 0 &&
751
752 // Also, the RHS should always point to somewhere within the vtable.
753 RHSOffset <=
754 static_cast<uint64_t>(DL.getTypeAllocSize(OrigGV.getInitializer()->getType()))) {
755 findFuncPointers(LHS, StartingOffset, M, Index, VTableFuncs, OrigGV);
756 }
757 }
758 }
759 }
760
761 // Identify the function pointers referenced by vtable definition \p V.
computeVTableFuncs(ModuleSummaryIndex & Index,const GlobalVariable & V,const Module & M,VTableFuncList & VTableFuncs)762 static void computeVTableFuncs(ModuleSummaryIndex &Index,
763 const GlobalVariable &V, const Module &M,
764 VTableFuncList &VTableFuncs) {
765 if (!V.isConstant())
766 return;
767
768 findFuncPointers(V.getInitializer(), /*StartingOffset=*/0, M, Index,
769 VTableFuncs, V);
770
771 #ifndef NDEBUG
772 // Validate that the VTableFuncs list is ordered by offset.
773 uint64_t PrevOffset = 0;
774 for (auto &P : VTableFuncs) {
775 // The findVFuncPointers traversal should have encountered the
776 // functions in offset order. We need to use ">=" since PrevOffset
777 // starts at 0.
778 assert(P.VTableOffset >= PrevOffset);
779 PrevOffset = P.VTableOffset;
780 }
781 #endif
782 }
783
784 /// Record vtable definition \p V for each type metadata it references.
785 static void
recordTypeIdCompatibleVtableReferences(ModuleSummaryIndex & Index,const GlobalVariable & V,SmallVectorImpl<MDNode * > & Types)786 recordTypeIdCompatibleVtableReferences(ModuleSummaryIndex &Index,
787 const GlobalVariable &V,
788 SmallVectorImpl<MDNode *> &Types) {
789 for (MDNode *Type : Types) {
790 auto TypeID = Type->getOperand(1).get();
791
792 uint64_t Offset =
793 cast<ConstantInt>(
794 cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
795 ->getZExtValue();
796
797 if (auto *TypeId = dyn_cast<MDString>(TypeID))
798 Index.getOrInsertTypeIdCompatibleVtableSummary(TypeId->getString())
799 .push_back({Offset, Index.getOrInsertValueInfo(&V)});
800 }
801 }
802
computeVariableSummary(ModuleSummaryIndex & Index,const GlobalVariable & V,DenseSet<GlobalValue::GUID> & CantBePromoted,const Module & M,SmallVectorImpl<MDNode * > & Types)803 static void computeVariableSummary(ModuleSummaryIndex &Index,
804 const GlobalVariable &V,
805 DenseSet<GlobalValue::GUID> &CantBePromoted,
806 const Module &M,
807 SmallVectorImpl<MDNode *> &Types) {
808 SetVector<ValueInfo, std::vector<ValueInfo>> RefEdges;
809 SmallPtrSet<const User *, 8> Visited;
810 bool RefLocalIFunc = false;
811 bool HasBlockAddress =
812 findRefEdges(Index, &V, RefEdges, Visited, RefLocalIFunc);
813 const bool NotEligibleForImport = (HasBlockAddress || RefLocalIFunc);
814 bool NonRenamableLocal = isNonRenamableLocal(V);
815 GlobalValueSummary::GVFlags Flags(
816 V.getLinkage(), V.getVisibility(), NonRenamableLocal,
817 /* Live = */ false, V.isDSOLocal(), V.canBeOmittedFromSymbolTable(),
818 GlobalValueSummary::Definition);
819
820 VTableFuncList VTableFuncs;
821 // If splitting is not enabled, then we compute the summary information
822 // necessary for index-based whole program devirtualization.
823 if (!Index.enableSplitLTOUnit()) {
824 Types.clear();
825 V.getMetadata(LLVMContext::MD_type, Types);
826 if (!Types.empty()) {
827 // Identify the function pointers referenced by this vtable definition.
828 computeVTableFuncs(Index, V, M, VTableFuncs);
829
830 // Record this vtable definition for each type metadata it references.
831 recordTypeIdCompatibleVtableReferences(Index, V, Types);
832 }
833 }
834
835 // Don't mark variables we won't be able to internalize as read/write-only.
836 bool CanBeInternalized =
837 !V.hasComdat() && !V.hasAppendingLinkage() && !V.isInterposable() &&
838 !V.hasAvailableExternallyLinkage() && !V.hasDLLExportStorageClass();
839 bool Constant = V.isConstant();
840 GlobalVarSummary::GVarFlags VarFlags(CanBeInternalized,
841 Constant ? false : CanBeInternalized,
842 Constant, V.getVCallVisibility());
843 auto GVarSummary = std::make_unique<GlobalVarSummary>(Flags, VarFlags,
844 RefEdges.takeVector());
845 if (NonRenamableLocal)
846 CantBePromoted.insert(V.getGUID());
847 if (NotEligibleForImport)
848 GVarSummary->setNotEligibleToImport();
849 if (!VTableFuncs.empty())
850 GVarSummary->setVTableFuncs(VTableFuncs);
851 Index.addGlobalValueSummary(V, std::move(GVarSummary));
852 }
853
computeAliasSummary(ModuleSummaryIndex & Index,const GlobalAlias & A,DenseSet<GlobalValue::GUID> & CantBePromoted)854 static void computeAliasSummary(ModuleSummaryIndex &Index, const GlobalAlias &A,
855 DenseSet<GlobalValue::GUID> &CantBePromoted) {
856 // Skip summary for indirect function aliases as summary for aliasee will not
857 // be emitted.
858 const GlobalObject *Aliasee = A.getAliaseeObject();
859 if (isa<GlobalIFunc>(Aliasee))
860 return;
861 bool NonRenamableLocal = isNonRenamableLocal(A);
862 GlobalValueSummary::GVFlags Flags(
863 A.getLinkage(), A.getVisibility(), NonRenamableLocal,
864 /* Live = */ false, A.isDSOLocal(), A.canBeOmittedFromSymbolTable(),
865 GlobalValueSummary::Definition);
866 auto AS = std::make_unique<AliasSummary>(Flags);
867 auto AliaseeVI = Index.getValueInfo(Aliasee->getGUID());
868 assert(AliaseeVI && "Alias expects aliasee summary to be available");
869 assert(AliaseeVI.getSummaryList().size() == 1 &&
870 "Expected a single entry per aliasee in per-module index");
871 AS->setAliasee(AliaseeVI, AliaseeVI.getSummaryList()[0].get());
872 if (NonRenamableLocal)
873 CantBePromoted.insert(A.getGUID());
874 Index.addGlobalValueSummary(A, std::move(AS));
875 }
876
877 // Set LiveRoot flag on entries matching the given value name.
setLiveRoot(ModuleSummaryIndex & Index,StringRef Name)878 static void setLiveRoot(ModuleSummaryIndex &Index, StringRef Name) {
879 if (ValueInfo VI = Index.getValueInfo(GlobalValue::getGUID(Name)))
880 for (const auto &Summary : VI.getSummaryList())
881 Summary->setLive(true);
882 }
883
buildModuleSummaryIndex(const Module & M,std::function<BlockFrequencyInfo * (const Function & F)> GetBFICallback,ProfileSummaryInfo * PSI,std::function<const StackSafetyInfo * (const Function & F)> GetSSICallback)884 ModuleSummaryIndex llvm::buildModuleSummaryIndex(
885 const Module &M,
886 std::function<BlockFrequencyInfo *(const Function &F)> GetBFICallback,
887 ProfileSummaryInfo *PSI,
888 std::function<const StackSafetyInfo *(const Function &F)> GetSSICallback) {
889 assert(PSI);
890 bool EnableSplitLTOUnit = false;
891 bool UnifiedLTO = false;
892 if (auto *MD = mdconst::extract_or_null<ConstantInt>(
893 M.getModuleFlag("EnableSplitLTOUnit")))
894 EnableSplitLTOUnit = MD->getZExtValue();
895 if (auto *MD =
896 mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("UnifiedLTO")))
897 UnifiedLTO = MD->getZExtValue();
898 ModuleSummaryIndex Index(/*HaveGVs=*/true, EnableSplitLTOUnit, UnifiedLTO);
899
900 // Identify the local values in the llvm.used and llvm.compiler.used sets,
901 // which should not be exported as they would then require renaming and
902 // promotion, but we may have opaque uses e.g. in inline asm. We collect them
903 // here because we use this information to mark functions containing inline
904 // assembly calls as not importable.
905 SmallPtrSet<GlobalValue *, 4> LocalsUsed;
906 SmallVector<GlobalValue *, 4> Used;
907 // First collect those in the llvm.used set.
908 collectUsedGlobalVariables(M, Used, /*CompilerUsed=*/false);
909 // Next collect those in the llvm.compiler.used set.
910 collectUsedGlobalVariables(M, Used, /*CompilerUsed=*/true);
911 DenseSet<GlobalValue::GUID> CantBePromoted;
912 for (auto *V : Used) {
913 if (V->hasLocalLinkage()) {
914 LocalsUsed.insert(V);
915 CantBePromoted.insert(V->getGUID());
916 }
917 }
918
919 bool HasLocalInlineAsmSymbol = false;
920 if (!M.getModuleInlineAsm().empty()) {
921 // Collect the local values defined by module level asm, and set up
922 // summaries for these symbols so that they can be marked as NoRename,
923 // to prevent export of any use of them in regular IR that would require
924 // renaming within the module level asm. Note we don't need to create a
925 // summary for weak or global defs, as they don't need to be flagged as
926 // NoRename, and defs in module level asm can't be imported anyway.
927 // Also, any values used but not defined within module level asm should
928 // be listed on the llvm.used or llvm.compiler.used global and marked as
929 // referenced from there.
930 ModuleSymbolTable::CollectAsmSymbols(
931 M, [&](StringRef Name, object::BasicSymbolRef::Flags Flags) {
932 // Symbols not marked as Weak or Global are local definitions.
933 if (Flags & (object::BasicSymbolRef::SF_Weak |
934 object::BasicSymbolRef::SF_Global))
935 return;
936 HasLocalInlineAsmSymbol = true;
937 GlobalValue *GV = M.getNamedValue(Name);
938 if (!GV)
939 return;
940 assert(GV->isDeclaration() && "Def in module asm already has definition");
941 GlobalValueSummary::GVFlags GVFlags(
942 GlobalValue::InternalLinkage, GlobalValue::DefaultVisibility,
943 /* NotEligibleToImport = */ true,
944 /* Live = */ true,
945 /* Local */ GV->isDSOLocal(), GV->canBeOmittedFromSymbolTable(),
946 GlobalValueSummary::Definition);
947 CantBePromoted.insert(GV->getGUID());
948 // Create the appropriate summary type.
949 if (Function *F = dyn_cast<Function>(GV)) {
950 std::unique_ptr<FunctionSummary> Summary =
951 std::make_unique<FunctionSummary>(
952 GVFlags, /*InstCount=*/0,
953 FunctionSummary::FFlags{
954 F->hasFnAttribute(Attribute::ReadNone),
955 F->hasFnAttribute(Attribute::ReadOnly),
956 F->hasFnAttribute(Attribute::NoRecurse),
957 F->returnDoesNotAlias(),
958 /* NoInline = */ false,
959 F->hasFnAttribute(Attribute::AlwaysInline),
960 F->hasFnAttribute(Attribute::NoUnwind),
961 /* MayThrow */ true,
962 /* HasUnknownCall */ true,
963 /* MustBeUnreachable */ false},
964 /*EntryCount=*/0, ArrayRef<ValueInfo>{},
965 ArrayRef<FunctionSummary::EdgeTy>{},
966 ArrayRef<GlobalValue::GUID>{},
967 ArrayRef<FunctionSummary::VFuncId>{},
968 ArrayRef<FunctionSummary::VFuncId>{},
969 ArrayRef<FunctionSummary::ConstVCall>{},
970 ArrayRef<FunctionSummary::ConstVCall>{},
971 ArrayRef<FunctionSummary::ParamAccess>{},
972 ArrayRef<CallsiteInfo>{}, ArrayRef<AllocInfo>{});
973 Index.addGlobalValueSummary(*GV, std::move(Summary));
974 } else {
975 std::unique_ptr<GlobalVarSummary> Summary =
976 std::make_unique<GlobalVarSummary>(
977 GVFlags,
978 GlobalVarSummary::GVarFlags(
979 false, false, cast<GlobalVariable>(GV)->isConstant(),
980 GlobalObject::VCallVisibilityPublic),
981 ArrayRef<ValueInfo>{});
982 Index.addGlobalValueSummary(*GV, std::move(Summary));
983 }
984 });
985 }
986
987 bool IsThinLTO = true;
988 if (auto *MD =
989 mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("ThinLTO")))
990 IsThinLTO = MD->getZExtValue();
991
992 // Compute summaries for all functions defined in module, and save in the
993 // index.
994 for (const auto &F : M) {
995 if (F.isDeclaration())
996 continue;
997
998 DominatorTree DT(const_cast<Function &>(F));
999 BlockFrequencyInfo *BFI = nullptr;
1000 std::unique_ptr<BlockFrequencyInfo> BFIPtr;
1001 if (GetBFICallback)
1002 BFI = GetBFICallback(F);
1003 else if (F.hasProfileData()) {
1004 LoopInfo LI{DT};
1005 BranchProbabilityInfo BPI{F, LI};
1006 BFIPtr = std::make_unique<BlockFrequencyInfo>(F, BPI, LI);
1007 BFI = BFIPtr.get();
1008 }
1009
1010 computeFunctionSummary(Index, M, F, BFI, PSI, DT,
1011 !LocalsUsed.empty() || HasLocalInlineAsmSymbol,
1012 CantBePromoted, IsThinLTO, GetSSICallback);
1013 }
1014
1015 // Compute summaries for all variables defined in module, and save in the
1016 // index.
1017 SmallVector<MDNode *, 2> Types;
1018 for (const GlobalVariable &G : M.globals()) {
1019 if (G.isDeclaration())
1020 continue;
1021 computeVariableSummary(Index, G, CantBePromoted, M, Types);
1022 }
1023
1024 // Compute summaries for all aliases defined in module, and save in the
1025 // index.
1026 for (const GlobalAlias &A : M.aliases())
1027 computeAliasSummary(Index, A, CantBePromoted);
1028
1029 // Iterate through ifuncs, set their resolvers all alive.
1030 for (const GlobalIFunc &I : M.ifuncs()) {
1031 I.applyAlongResolverPath([&Index](const GlobalValue &GV) {
1032 Index.getGlobalValueSummary(GV)->setLive(true);
1033 });
1034 }
1035
1036 for (auto *V : LocalsUsed) {
1037 auto *Summary = Index.getGlobalValueSummary(*V);
1038 assert(Summary && "Missing summary for global value");
1039 Summary->setNotEligibleToImport();
1040 }
1041
1042 // The linker doesn't know about these LLVM produced values, so we need
1043 // to flag them as live in the index to ensure index-based dead value
1044 // analysis treats them as live roots of the analysis.
1045 setLiveRoot(Index, "llvm.used");
1046 setLiveRoot(Index, "llvm.compiler.used");
1047 setLiveRoot(Index, "llvm.global_ctors");
1048 setLiveRoot(Index, "llvm.global_dtors");
1049 setLiveRoot(Index, "llvm.global.annotations");
1050
1051 for (auto &GlobalList : Index) {
1052 // Ignore entries for references that are undefined in the current module.
1053 if (GlobalList.second.SummaryList.empty())
1054 continue;
1055
1056 assert(GlobalList.second.SummaryList.size() == 1 &&
1057 "Expected module's index to have one summary per GUID");
1058 auto &Summary = GlobalList.second.SummaryList[0];
1059 if (!IsThinLTO) {
1060 Summary->setNotEligibleToImport();
1061 continue;
1062 }
1063
1064 bool AllRefsCanBeExternallyReferenced =
1065 llvm::all_of(Summary->refs(), [&](const ValueInfo &VI) {
1066 return !CantBePromoted.count(VI.getGUID());
1067 });
1068 if (!AllRefsCanBeExternallyReferenced) {
1069 Summary->setNotEligibleToImport();
1070 continue;
1071 }
1072
1073 if (auto *FuncSummary = dyn_cast<FunctionSummary>(Summary.get())) {
1074 bool AllCallsCanBeExternallyReferenced = llvm::all_of(
1075 FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) {
1076 return !CantBePromoted.count(Edge.first.getGUID());
1077 });
1078 if (!AllCallsCanBeExternallyReferenced)
1079 Summary->setNotEligibleToImport();
1080 }
1081 }
1082
1083 if (!ModuleSummaryDotFile.empty()) {
1084 std::error_code EC;
1085 raw_fd_ostream OSDot(ModuleSummaryDotFile, EC, sys::fs::OpenFlags::OF_Text);
1086 if (EC)
1087 report_fatal_error(Twine("Failed to open dot file ") +
1088 ModuleSummaryDotFile + ": " + EC.message() + "\n");
1089 Index.exportToDot(OSDot, {});
1090 }
1091
1092 return Index;
1093 }
1094
1095 AnalysisKey ModuleSummaryIndexAnalysis::Key;
1096
1097 ModuleSummaryIndex
run(Module & M,ModuleAnalysisManager & AM)1098 ModuleSummaryIndexAnalysis::run(Module &M, ModuleAnalysisManager &AM) {
1099 ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M);
1100 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1101 bool NeedSSI = needsParamAccessSummary(M);
1102 return buildModuleSummaryIndex(
1103 M,
1104 [&FAM](const Function &F) {
1105 return &FAM.getResult<BlockFrequencyAnalysis>(
1106 *const_cast<Function *>(&F));
1107 },
1108 &PSI,
1109 [&FAM, NeedSSI](const Function &F) -> const StackSafetyInfo * {
1110 return NeedSSI ? &FAM.getResult<StackSafetyAnalysis>(
1111 const_cast<Function &>(F))
1112 : nullptr;
1113 });
1114 }
1115
1116 char ModuleSummaryIndexWrapperPass::ID = 0;
1117
1118 INITIALIZE_PASS_BEGIN(ModuleSummaryIndexWrapperPass, "module-summary-analysis",
1119 "Module Summary Analysis", false, true)
INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)1120 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
1121 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
1122 INITIALIZE_PASS_DEPENDENCY(StackSafetyInfoWrapperPass)
1123 INITIALIZE_PASS_END(ModuleSummaryIndexWrapperPass, "module-summary-analysis",
1124 "Module Summary Analysis", false, true)
1125
1126 ModulePass *llvm::createModuleSummaryIndexWrapperPass() {
1127 return new ModuleSummaryIndexWrapperPass();
1128 }
1129
ModuleSummaryIndexWrapperPass()1130 ModuleSummaryIndexWrapperPass::ModuleSummaryIndexWrapperPass()
1131 : ModulePass(ID) {
1132 initializeModuleSummaryIndexWrapperPassPass(*PassRegistry::getPassRegistry());
1133 }
1134
runOnModule(Module & M)1135 bool ModuleSummaryIndexWrapperPass::runOnModule(Module &M) {
1136 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
1137 bool NeedSSI = needsParamAccessSummary(M);
1138 Index.emplace(buildModuleSummaryIndex(
1139 M,
1140 [this](const Function &F) {
1141 return &(this->getAnalysis<BlockFrequencyInfoWrapperPass>(
1142 *const_cast<Function *>(&F))
1143 .getBFI());
1144 },
1145 PSI,
1146 [&](const Function &F) -> const StackSafetyInfo * {
1147 return NeedSSI ? &getAnalysis<StackSafetyInfoWrapperPass>(
1148 const_cast<Function &>(F))
1149 .getResult()
1150 : nullptr;
1151 }));
1152 return false;
1153 }
1154
doFinalization(Module & M)1155 bool ModuleSummaryIndexWrapperPass::doFinalization(Module &M) {
1156 Index.reset();
1157 return false;
1158 }
1159
getAnalysisUsage(AnalysisUsage & AU) const1160 void ModuleSummaryIndexWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1161 AU.setPreservesAll();
1162 AU.addRequired<BlockFrequencyInfoWrapperPass>();
1163 AU.addRequired<ProfileSummaryInfoWrapperPass>();
1164 AU.addRequired<StackSafetyInfoWrapperPass>();
1165 }
1166
1167 char ImmutableModuleSummaryIndexWrapperPass::ID = 0;
1168
ImmutableModuleSummaryIndexWrapperPass(const ModuleSummaryIndex * Index)1169 ImmutableModuleSummaryIndexWrapperPass::ImmutableModuleSummaryIndexWrapperPass(
1170 const ModuleSummaryIndex *Index)
1171 : ImmutablePass(ID), Index(Index) {
1172 initializeImmutableModuleSummaryIndexWrapperPassPass(
1173 *PassRegistry::getPassRegistry());
1174 }
1175
getAnalysisUsage(AnalysisUsage & AU) const1176 void ImmutableModuleSummaryIndexWrapperPass::getAnalysisUsage(
1177 AnalysisUsage &AU) const {
1178 AU.setPreservesAll();
1179 }
1180
createImmutableModuleSummaryIndexWrapperPass(const ModuleSummaryIndex * Index)1181 ImmutablePass *llvm::createImmutableModuleSummaryIndexWrapperPass(
1182 const ModuleSummaryIndex *Index) {
1183 return new ImmutableModuleSummaryIndexWrapperPass(Index);
1184 }
1185
1186 INITIALIZE_PASS(ImmutableModuleSummaryIndexWrapperPass, "module-summary-info",
1187 "Module summary info", false, true)
1188
mayHaveMemprofSummary(const CallBase * CB)1189 bool llvm::mayHaveMemprofSummary(const CallBase *CB) {
1190 if (!CB)
1191 return false;
1192 if (CB->isDebugOrPseudoInst())
1193 return false;
1194 auto *CI = dyn_cast<CallInst>(CB);
1195 auto *CalledValue = CB->getCalledOperand();
1196 auto *CalledFunction = CB->getCalledFunction();
1197 if (CalledValue && !CalledFunction) {
1198 CalledValue = CalledValue->stripPointerCasts();
1199 // Stripping pointer casts can reveal a called function.
1200 CalledFunction = dyn_cast<Function>(CalledValue);
1201 }
1202 // Check if this is an alias to a function. If so, get the
1203 // called aliasee for the checks below.
1204 if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) {
1205 assert(!CalledFunction &&
1206 "Expected null called function in callsite for alias");
1207 CalledFunction = dyn_cast<Function>(GA->getAliaseeObject());
1208 }
1209 // Check if this is a direct call to a known function or a known
1210 // intrinsic, or an indirect call with profile data.
1211 if (CalledFunction) {
1212 if (CI && CalledFunction->isIntrinsic())
1213 return false;
1214 } else {
1215 // TODO: For now skip indirect calls. See comments in
1216 // computeFunctionSummary for what is needed to handle this.
1217 return false;
1218 }
1219 return true;
1220 }
1221