1 //===- ModuleSummaryAnalysis.cpp - Module summary index builder -----------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass builds a ModuleSummaryIndex object for the module, to be written
11 // to bitcode or LLVM assembly.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/Analysis/ModuleSummaryAnalysis.h"
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/MapVector.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/Analysis/BlockFrequencyInfo.h"
25 #include "llvm/Analysis/BranchProbabilityInfo.h"
26 #include "llvm/Analysis/IndirectCallPromotionAnalysis.h"
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/Analysis/ProfileSummaryInfo.h"
29 #include "llvm/Analysis/TypeMetadataUtils.h"
30 #include "llvm/IR/Attributes.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/CallSite.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/Intrinsics.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/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 <algorithm>
54 #include <cassert>
55 #include <cstdint>
56 #include <vector>
57
58 using namespace llvm;
59
60 #define DEBUG_TYPE "module-summary-analysis"
61
62 // Option to force edges cold which will block importing when the
63 // -import-cold-multiplier is set to 0. Useful for debugging.
64 FunctionSummary::ForceSummaryHotnessType ForceSummaryEdgesCold =
65 FunctionSummary::FSHT_None;
66 cl::opt<FunctionSummary::ForceSummaryHotnessType, true> FSEC(
67 "force-summary-edges-cold", cl::Hidden, cl::location(ForceSummaryEdgesCold),
68 cl::desc("Force all edges in the function summary to cold"),
69 cl::values(clEnumValN(FunctionSummary::FSHT_None, "none", "None."),
70 clEnumValN(FunctionSummary::FSHT_AllNonCritical,
71 "all-non-critical", "All non-critical edges."),
72 clEnumValN(FunctionSummary::FSHT_All, "all", "All edges.")));
73
74 // Walk through the operands of a given User via worklist iteration and populate
75 // the set of GlobalValue references encountered. Invoked either on an
76 // Instruction or a GlobalVariable (which walks its initializer).
77 // Return true if any of the operands contains blockaddress. This is important
78 // to know when computing summary for global var, because if global variable
79 // references basic block address we can't import it separately from function
80 // containing that basic block. For simplicity we currently don't import such
81 // global vars at all. When importing function we aren't interested if any
82 // instruction in it takes an address of any basic block, because instruction
83 // can only take an address of basic block located in the same function.
findRefEdges(ModuleSummaryIndex & Index,const User * CurUser,SetVector<ValueInfo> & RefEdges,SmallPtrSet<const User *,8> & Visited)84 static bool findRefEdges(ModuleSummaryIndex &Index, const User *CurUser,
85 SetVector<ValueInfo> &RefEdges,
86 SmallPtrSet<const User *, 8> &Visited) {
87 bool HasBlockAddress = false;
88 SmallVector<const User *, 32> Worklist;
89 Worklist.push_back(CurUser);
90
91 while (!Worklist.empty()) {
92 const User *U = Worklist.pop_back_val();
93
94 if (!Visited.insert(U).second)
95 continue;
96
97 ImmutableCallSite CS(U);
98
99 for (const auto &OI : U->operands()) {
100 const User *Operand = dyn_cast<User>(OI);
101 if (!Operand)
102 continue;
103 if (isa<BlockAddress>(Operand)) {
104 HasBlockAddress = true;
105 continue;
106 }
107 if (auto *GV = dyn_cast<GlobalValue>(Operand)) {
108 // We have a reference to a global value. This should be added to
109 // the reference set unless it is a callee. Callees are handled
110 // specially by WriteFunction and are added to a separate list.
111 if (!(CS && CS.isCallee(&OI)))
112 RefEdges.insert(Index.getOrInsertValueInfo(GV));
113 continue;
114 }
115 Worklist.push_back(Operand);
116 }
117 }
118 return HasBlockAddress;
119 }
120
getHotness(uint64_t ProfileCount,ProfileSummaryInfo * PSI)121 static CalleeInfo::HotnessType getHotness(uint64_t ProfileCount,
122 ProfileSummaryInfo *PSI) {
123 if (!PSI)
124 return CalleeInfo::HotnessType::Unknown;
125 if (PSI->isHotCount(ProfileCount))
126 return CalleeInfo::HotnessType::Hot;
127 if (PSI->isColdCount(ProfileCount))
128 return CalleeInfo::HotnessType::Cold;
129 return CalleeInfo::HotnessType::None;
130 }
131
isNonRenamableLocal(const GlobalValue & GV)132 static bool isNonRenamableLocal(const GlobalValue &GV) {
133 return GV.hasSection() && GV.hasLocalLinkage();
134 }
135
136 /// Determine whether this call has all constant integer arguments (excluding
137 /// "this") and summarize it to VCalls or ConstVCalls as appropriate.
addVCallToSet(DevirtCallSite Call,GlobalValue::GUID Guid,SetVector<FunctionSummary::VFuncId> & VCalls,SetVector<FunctionSummary::ConstVCall> & ConstVCalls)138 static void addVCallToSet(DevirtCallSite Call, GlobalValue::GUID Guid,
139 SetVector<FunctionSummary::VFuncId> &VCalls,
140 SetVector<FunctionSummary::ConstVCall> &ConstVCalls) {
141 std::vector<uint64_t> Args;
142 // Start from the second argument to skip the "this" pointer.
143 for (auto &Arg : make_range(Call.CS.arg_begin() + 1, Call.CS.arg_end())) {
144 auto *CI = dyn_cast<ConstantInt>(Arg);
145 if (!CI || CI->getBitWidth() > 64) {
146 VCalls.insert({Guid, Call.Offset});
147 return;
148 }
149 Args.push_back(CI->getZExtValue());
150 }
151 ConstVCalls.insert({{Guid, Call.Offset}, std::move(Args)});
152 }
153
154 /// If this intrinsic call requires that we add information to the function
155 /// summary, do so via the non-constant reference arguments.
addIntrinsicToSummary(const CallInst * CI,SetVector<GlobalValue::GUID> & TypeTests,SetVector<FunctionSummary::VFuncId> & TypeTestAssumeVCalls,SetVector<FunctionSummary::VFuncId> & TypeCheckedLoadVCalls,SetVector<FunctionSummary::ConstVCall> & TypeTestAssumeConstVCalls,SetVector<FunctionSummary::ConstVCall> & TypeCheckedLoadConstVCalls,DominatorTree & DT)156 static void addIntrinsicToSummary(
157 const CallInst *CI, SetVector<GlobalValue::GUID> &TypeTests,
158 SetVector<FunctionSummary::VFuncId> &TypeTestAssumeVCalls,
159 SetVector<FunctionSummary::VFuncId> &TypeCheckedLoadVCalls,
160 SetVector<FunctionSummary::ConstVCall> &TypeTestAssumeConstVCalls,
161 SetVector<FunctionSummary::ConstVCall> &TypeCheckedLoadConstVCalls,
162 DominatorTree &DT) {
163 switch (CI->getCalledFunction()->getIntrinsicID()) {
164 case Intrinsic::type_test: {
165 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(1));
166 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata());
167 if (!TypeId)
168 break;
169 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString());
170
171 // Produce a summary from type.test intrinsics. We only summarize type.test
172 // intrinsics that are used other than by an llvm.assume intrinsic.
173 // Intrinsics that are assumed are relevant only to the devirtualization
174 // pass, not the type test lowering pass.
175 bool HasNonAssumeUses = llvm::any_of(CI->uses(), [](const Use &CIU) {
176 auto *AssumeCI = dyn_cast<CallInst>(CIU.getUser());
177 if (!AssumeCI)
178 return true;
179 Function *F = AssumeCI->getCalledFunction();
180 return !F || F->getIntrinsicID() != Intrinsic::assume;
181 });
182 if (HasNonAssumeUses)
183 TypeTests.insert(Guid);
184
185 SmallVector<DevirtCallSite, 4> DevirtCalls;
186 SmallVector<CallInst *, 4> Assumes;
187 findDevirtualizableCallsForTypeTest(DevirtCalls, Assumes, CI, DT);
188 for (auto &Call : DevirtCalls)
189 addVCallToSet(Call, Guid, TypeTestAssumeVCalls,
190 TypeTestAssumeConstVCalls);
191
192 break;
193 }
194
195 case Intrinsic::type_checked_load: {
196 auto *TypeMDVal = cast<MetadataAsValue>(CI->getArgOperand(2));
197 auto *TypeId = dyn_cast<MDString>(TypeMDVal->getMetadata());
198 if (!TypeId)
199 break;
200 GlobalValue::GUID Guid = GlobalValue::getGUID(TypeId->getString());
201
202 SmallVector<DevirtCallSite, 4> DevirtCalls;
203 SmallVector<Instruction *, 4> LoadedPtrs;
204 SmallVector<Instruction *, 4> Preds;
205 bool HasNonCallUses = false;
206 findDevirtualizableCallsForTypeCheckedLoad(DevirtCalls, LoadedPtrs, Preds,
207 HasNonCallUses, CI, DT);
208 // Any non-call uses of the result of llvm.type.checked.load will
209 // prevent us from optimizing away the llvm.type.test.
210 if (HasNonCallUses)
211 TypeTests.insert(Guid);
212 for (auto &Call : DevirtCalls)
213 addVCallToSet(Call, Guid, TypeCheckedLoadVCalls,
214 TypeCheckedLoadConstVCalls);
215
216 break;
217 }
218 default:
219 break;
220 }
221 }
222
isNonVolatileLoad(const Instruction * I)223 static bool isNonVolatileLoad(const Instruction *I) {
224 if (const auto *LI = dyn_cast<LoadInst>(I))
225 return !LI->isVolatile();
226
227 return false;
228 }
229
computeFunctionSummary(ModuleSummaryIndex & Index,const Module & M,const Function & F,BlockFrequencyInfo * BFI,ProfileSummaryInfo * PSI,DominatorTree & DT,bool HasLocalsInUsedOrAsm,DenseSet<GlobalValue::GUID> & CantBePromoted,bool IsThinLTO)230 static void computeFunctionSummary(ModuleSummaryIndex &Index, const Module &M,
231 const Function &F, BlockFrequencyInfo *BFI,
232 ProfileSummaryInfo *PSI, DominatorTree &DT,
233 bool HasLocalsInUsedOrAsm,
234 DenseSet<GlobalValue::GUID> &CantBePromoted,
235 bool IsThinLTO) {
236 // Summary not currently supported for anonymous functions, they should
237 // have been named.
238 assert(F.hasName());
239
240 unsigned NumInsts = 0;
241 // Map from callee ValueId to profile count. Used to accumulate profile
242 // counts for all static calls to a given callee.
243 MapVector<ValueInfo, CalleeInfo> CallGraphEdges;
244 SetVector<ValueInfo> RefEdges;
245 SetVector<GlobalValue::GUID> TypeTests;
246 SetVector<FunctionSummary::VFuncId> TypeTestAssumeVCalls,
247 TypeCheckedLoadVCalls;
248 SetVector<FunctionSummary::ConstVCall> TypeTestAssumeConstVCalls,
249 TypeCheckedLoadConstVCalls;
250 ICallPromotionAnalysis ICallAnalysis;
251 SmallPtrSet<const User *, 8> Visited;
252
253 // Add personality function, prefix data and prologue data to function's ref
254 // list.
255 findRefEdges(Index, &F, RefEdges, Visited);
256 std::vector<const Instruction *> NonVolatileLoads;
257
258 bool HasInlineAsmMaybeReferencingInternal = false;
259 for (const BasicBlock &BB : F)
260 for (const Instruction &I : BB) {
261 if (isa<DbgInfoIntrinsic>(I))
262 continue;
263 ++NumInsts;
264 if (isNonVolatileLoad(&I)) {
265 // Postpone processing of non-volatile load instructions
266 // See comments below
267 Visited.insert(&I);
268 NonVolatileLoads.push_back(&I);
269 continue;
270 }
271 findRefEdges(Index, &I, RefEdges, Visited);
272 auto CS = ImmutableCallSite(&I);
273 if (!CS)
274 continue;
275
276 const auto *CI = dyn_cast<CallInst>(&I);
277 // Since we don't know exactly which local values are referenced in inline
278 // assembly, conservatively mark the function as possibly referencing
279 // a local value from inline assembly to ensure we don't export a
280 // reference (which would require renaming and promotion of the
281 // referenced value).
282 if (HasLocalsInUsedOrAsm && CI && CI->isInlineAsm())
283 HasInlineAsmMaybeReferencingInternal = true;
284
285 auto *CalledValue = CS.getCalledValue();
286 auto *CalledFunction = CS.getCalledFunction();
287 if (CalledValue && !CalledFunction) {
288 CalledValue = CalledValue->stripPointerCastsNoFollowAliases();
289 // Stripping pointer casts can reveal a called function.
290 CalledFunction = dyn_cast<Function>(CalledValue);
291 }
292 // Check if this is an alias to a function. If so, get the
293 // called aliasee for the checks below.
294 if (auto *GA = dyn_cast<GlobalAlias>(CalledValue)) {
295 assert(!CalledFunction && "Expected null called function in callsite for alias");
296 CalledFunction = dyn_cast<Function>(GA->getBaseObject());
297 }
298 // Check if this is a direct call to a known function or a known
299 // intrinsic, or an indirect call with profile data.
300 if (CalledFunction) {
301 if (CI && CalledFunction->isIntrinsic()) {
302 addIntrinsicToSummary(
303 CI, TypeTests, TypeTestAssumeVCalls, TypeCheckedLoadVCalls,
304 TypeTestAssumeConstVCalls, TypeCheckedLoadConstVCalls, DT);
305 continue;
306 }
307 // We should have named any anonymous globals
308 assert(CalledFunction->hasName());
309 auto ScaledCount = PSI->getProfileCount(&I, BFI);
310 auto Hotness = ScaledCount ? getHotness(ScaledCount.getValue(), PSI)
311 : CalleeInfo::HotnessType::Unknown;
312 if (ForceSummaryEdgesCold != FunctionSummary::FSHT_None)
313 Hotness = CalleeInfo::HotnessType::Cold;
314
315 // Use the original CalledValue, in case it was an alias. We want
316 // to record the call edge to the alias in that case. Eventually
317 // an alias summary will be created to associate the alias and
318 // aliasee.
319 auto &ValueInfo = CallGraphEdges[Index.getOrInsertValueInfo(
320 cast<GlobalValue>(CalledValue))];
321 ValueInfo.updateHotness(Hotness);
322 // Add the relative block frequency to CalleeInfo if there is no profile
323 // information.
324 if (BFI != nullptr && Hotness == CalleeInfo::HotnessType::Unknown) {
325 uint64_t BBFreq = BFI->getBlockFreq(&BB).getFrequency();
326 uint64_t EntryFreq = BFI->getEntryFreq();
327 ValueInfo.updateRelBlockFreq(BBFreq, EntryFreq);
328 }
329 } else {
330 // Skip inline assembly calls.
331 if (CI && CI->isInlineAsm())
332 continue;
333 // Skip direct calls.
334 if (!CalledValue || isa<Constant>(CalledValue))
335 continue;
336
337 // Check if the instruction has a callees metadata. If so, add callees
338 // to CallGraphEdges to reflect the references from the metadata, and
339 // to enable importing for subsequent indirect call promotion and
340 // inlining.
341 if (auto *MD = I.getMetadata(LLVMContext::MD_callees)) {
342 for (auto &Op : MD->operands()) {
343 Function *Callee = mdconst::extract_or_null<Function>(Op);
344 if (Callee)
345 CallGraphEdges[Index.getOrInsertValueInfo(Callee)];
346 }
347 }
348
349 uint32_t NumVals, NumCandidates;
350 uint64_t TotalCount;
351 auto CandidateProfileData =
352 ICallAnalysis.getPromotionCandidatesForInstruction(
353 &I, NumVals, TotalCount, NumCandidates);
354 for (auto &Candidate : CandidateProfileData)
355 CallGraphEdges[Index.getOrInsertValueInfo(Candidate.Value)]
356 .updateHotness(getHotness(Candidate.Count, PSI));
357 }
358 }
359
360 // By now we processed all instructions in a function, except
361 // non-volatile loads. All new refs we add in a loop below
362 // are obviously constant. All constant refs are grouped in the
363 // end of RefEdges vector, so we can use a single integer value
364 // to identify them.
365 unsigned RefCnt = RefEdges.size();
366 for (const Instruction *I : NonVolatileLoads) {
367 Visited.erase(I);
368 findRefEdges(Index, I, RefEdges, Visited);
369 }
370 std::vector<ValueInfo> Refs = RefEdges.takeVector();
371 // Regular LTO module doesn't participate in ThinLTO import,
372 // so no reference from it can be readonly, since this would
373 // require importing variable as local copy
374 if (IsThinLTO)
375 for (; RefCnt < Refs.size(); ++RefCnt)
376 Refs[RefCnt].setReadOnly();
377
378 // Explicit add hot edges to enforce importing for designated GUIDs for
379 // sample PGO, to enable the same inlines as the profiled optimized binary.
380 for (auto &I : F.getImportGUIDs())
381 CallGraphEdges[Index.getOrInsertValueInfo(I)].updateHotness(
382 ForceSummaryEdgesCold == FunctionSummary::FSHT_All
383 ? CalleeInfo::HotnessType::Cold
384 : CalleeInfo::HotnessType::Critical);
385
386 bool NonRenamableLocal = isNonRenamableLocal(F);
387 bool NotEligibleForImport =
388 NonRenamableLocal || HasInlineAsmMaybeReferencingInternal;
389 GlobalValueSummary::GVFlags Flags(F.getLinkage(), NotEligibleForImport,
390 /* Live = */ false, F.isDSOLocal());
391 FunctionSummary::FFlags FunFlags{
392 F.hasFnAttribute(Attribute::ReadNone),
393 F.hasFnAttribute(Attribute::ReadOnly),
394 F.hasFnAttribute(Attribute::NoRecurse), F.returnDoesNotAlias(),
395 // FIXME: refactor this to use the same code that inliner is using.
396 // Don't try to import functions with noinline attribute.
397 F.getAttributes().hasFnAttribute(Attribute::NoInline)};
398 auto FuncSummary = llvm::make_unique<FunctionSummary>(
399 Flags, NumInsts, FunFlags, /*EntryCount=*/0, std::move(Refs),
400 CallGraphEdges.takeVector(), TypeTests.takeVector(),
401 TypeTestAssumeVCalls.takeVector(), TypeCheckedLoadVCalls.takeVector(),
402 TypeTestAssumeConstVCalls.takeVector(),
403 TypeCheckedLoadConstVCalls.takeVector());
404 if (NonRenamableLocal)
405 CantBePromoted.insert(F.getGUID());
406 Index.addGlobalValueSummary(F, std::move(FuncSummary));
407 }
408
409 static void
computeVariableSummary(ModuleSummaryIndex & Index,const GlobalVariable & V,DenseSet<GlobalValue::GUID> & CantBePromoted)410 computeVariableSummary(ModuleSummaryIndex &Index, const GlobalVariable &V,
411 DenseSet<GlobalValue::GUID> &CantBePromoted) {
412 SetVector<ValueInfo> RefEdges;
413 SmallPtrSet<const User *, 8> Visited;
414 bool HasBlockAddress = findRefEdges(Index, &V, RefEdges, Visited);
415 bool NonRenamableLocal = isNonRenamableLocal(V);
416 GlobalValueSummary::GVFlags Flags(V.getLinkage(), NonRenamableLocal,
417 /* Live = */ false, V.isDSOLocal());
418
419 // Don't mark variables we won't be able to internalize as read-only.
420 GlobalVarSummary::GVarFlags VarFlags(
421 !V.hasComdat() && !V.hasAppendingLinkage() && !V.isInterposable() &&
422 !V.hasAvailableExternallyLinkage() && !V.hasDLLExportStorageClass());
423 auto GVarSummary = llvm::make_unique<GlobalVarSummary>(Flags, VarFlags,
424 RefEdges.takeVector());
425 if (NonRenamableLocal)
426 CantBePromoted.insert(V.getGUID());
427 if (HasBlockAddress)
428 GVarSummary->setNotEligibleToImport();
429 Index.addGlobalValueSummary(V, std::move(GVarSummary));
430 }
431
432 static void
computeAliasSummary(ModuleSummaryIndex & Index,const GlobalAlias & A,DenseSet<GlobalValue::GUID> & CantBePromoted)433 computeAliasSummary(ModuleSummaryIndex &Index, const GlobalAlias &A,
434 DenseSet<GlobalValue::GUID> &CantBePromoted) {
435 bool NonRenamableLocal = isNonRenamableLocal(A);
436 GlobalValueSummary::GVFlags Flags(A.getLinkage(), NonRenamableLocal,
437 /* Live = */ false, A.isDSOLocal());
438 auto AS = llvm::make_unique<AliasSummary>(Flags);
439 auto *Aliasee = A.getBaseObject();
440 auto *AliaseeSummary = Index.getGlobalValueSummary(*Aliasee);
441 assert(AliaseeSummary && "Alias expects aliasee summary to be parsed");
442 AS->setAliasee(AliaseeSummary);
443 if (NonRenamableLocal)
444 CantBePromoted.insert(A.getGUID());
445 Index.addGlobalValueSummary(A, std::move(AS));
446 }
447
448 // Set LiveRoot flag on entries matching the given value name.
setLiveRoot(ModuleSummaryIndex & Index,StringRef Name)449 static void setLiveRoot(ModuleSummaryIndex &Index, StringRef Name) {
450 if (ValueInfo VI = Index.getValueInfo(GlobalValue::getGUID(Name)))
451 for (auto &Summary : VI.getSummaryList())
452 Summary->setLive(true);
453 }
454
buildModuleSummaryIndex(const Module & M,std::function<BlockFrequencyInfo * (const Function & F)> GetBFICallback,ProfileSummaryInfo * PSI)455 ModuleSummaryIndex llvm::buildModuleSummaryIndex(
456 const Module &M,
457 std::function<BlockFrequencyInfo *(const Function &F)> GetBFICallback,
458 ProfileSummaryInfo *PSI) {
459 assert(PSI);
460 bool EnableSplitLTOUnit = false;
461 if (auto *MD = mdconst::extract_or_null<ConstantInt>(
462 M.getModuleFlag("EnableSplitLTOUnit")))
463 EnableSplitLTOUnit = MD->getZExtValue();
464 ModuleSummaryIndex Index(/*HaveGVs=*/true, EnableSplitLTOUnit);
465
466 // Identify the local values in the llvm.used and llvm.compiler.used sets,
467 // which should not be exported as they would then require renaming and
468 // promotion, but we may have opaque uses e.g. in inline asm. We collect them
469 // here because we use this information to mark functions containing inline
470 // assembly calls as not importable.
471 SmallPtrSet<GlobalValue *, 8> LocalsUsed;
472 SmallPtrSet<GlobalValue *, 8> Used;
473 // First collect those in the llvm.used set.
474 collectUsedGlobalVariables(M, Used, /*CompilerUsed*/ false);
475 // Next collect those in the llvm.compiler.used set.
476 collectUsedGlobalVariables(M, Used, /*CompilerUsed*/ true);
477 DenseSet<GlobalValue::GUID> CantBePromoted;
478 for (auto *V : Used) {
479 if (V->hasLocalLinkage()) {
480 LocalsUsed.insert(V);
481 CantBePromoted.insert(V->getGUID());
482 }
483 }
484
485 bool HasLocalInlineAsmSymbol = false;
486 if (!M.getModuleInlineAsm().empty()) {
487 // Collect the local values defined by module level asm, and set up
488 // summaries for these symbols so that they can be marked as NoRename,
489 // to prevent export of any use of them in regular IR that would require
490 // renaming within the module level asm. Note we don't need to create a
491 // summary for weak or global defs, as they don't need to be flagged as
492 // NoRename, and defs in module level asm can't be imported anyway.
493 // Also, any values used but not defined within module level asm should
494 // be listed on the llvm.used or llvm.compiler.used global and marked as
495 // referenced from there.
496 ModuleSymbolTable::CollectAsmSymbols(
497 M, [&](StringRef Name, object::BasicSymbolRef::Flags Flags) {
498 // Symbols not marked as Weak or Global are local definitions.
499 if (Flags & (object::BasicSymbolRef::SF_Weak |
500 object::BasicSymbolRef::SF_Global))
501 return;
502 HasLocalInlineAsmSymbol = true;
503 GlobalValue *GV = M.getNamedValue(Name);
504 if (!GV)
505 return;
506 assert(GV->isDeclaration() && "Def in module asm already has definition");
507 GlobalValueSummary::GVFlags GVFlags(GlobalValue::InternalLinkage,
508 /* NotEligibleToImport = */ true,
509 /* Live = */ true,
510 /* Local */ GV->isDSOLocal());
511 CantBePromoted.insert(GV->getGUID());
512 // Create the appropriate summary type.
513 if (Function *F = dyn_cast<Function>(GV)) {
514 std::unique_ptr<FunctionSummary> Summary =
515 llvm::make_unique<FunctionSummary>(
516 GVFlags, /*InstCount=*/0,
517 FunctionSummary::FFlags{
518 F->hasFnAttribute(Attribute::ReadNone),
519 F->hasFnAttribute(Attribute::ReadOnly),
520 F->hasFnAttribute(Attribute::NoRecurse),
521 F->returnDoesNotAlias(),
522 /* NoInline = */ false},
523 /*EntryCount=*/0, ArrayRef<ValueInfo>{},
524 ArrayRef<FunctionSummary::EdgeTy>{},
525 ArrayRef<GlobalValue::GUID>{},
526 ArrayRef<FunctionSummary::VFuncId>{},
527 ArrayRef<FunctionSummary::VFuncId>{},
528 ArrayRef<FunctionSummary::ConstVCall>{},
529 ArrayRef<FunctionSummary::ConstVCall>{});
530 Index.addGlobalValueSummary(*GV, std::move(Summary));
531 } else {
532 std::unique_ptr<GlobalVarSummary> Summary =
533 llvm::make_unique<GlobalVarSummary>(
534 GVFlags, GlobalVarSummary::GVarFlags(),
535 ArrayRef<ValueInfo>{});
536 Index.addGlobalValueSummary(*GV, std::move(Summary));
537 }
538 });
539 }
540
541 bool IsThinLTO = true;
542 if (auto *MD =
543 mdconst::extract_or_null<ConstantInt>(M.getModuleFlag("ThinLTO")))
544 IsThinLTO = MD->getZExtValue();
545
546 // Compute summaries for all functions defined in module, and save in the
547 // index.
548 for (auto &F : M) {
549 if (F.isDeclaration())
550 continue;
551
552 DominatorTree DT(const_cast<Function &>(F));
553 BlockFrequencyInfo *BFI = nullptr;
554 std::unique_ptr<BlockFrequencyInfo> BFIPtr;
555 if (GetBFICallback)
556 BFI = GetBFICallback(F);
557 else if (F.hasProfileData()) {
558 LoopInfo LI{DT};
559 BranchProbabilityInfo BPI{F, LI};
560 BFIPtr = llvm::make_unique<BlockFrequencyInfo>(F, BPI, LI);
561 BFI = BFIPtr.get();
562 }
563
564 computeFunctionSummary(Index, M, F, BFI, PSI, DT,
565 !LocalsUsed.empty() || HasLocalInlineAsmSymbol,
566 CantBePromoted, IsThinLTO);
567 }
568
569 // Compute summaries for all variables defined in module, and save in the
570 // index.
571 for (const GlobalVariable &G : M.globals()) {
572 if (G.isDeclaration())
573 continue;
574 computeVariableSummary(Index, G, CantBePromoted);
575 }
576
577 // Compute summaries for all aliases defined in module, and save in the
578 // index.
579 for (const GlobalAlias &A : M.aliases())
580 computeAliasSummary(Index, A, CantBePromoted);
581
582 for (auto *V : LocalsUsed) {
583 auto *Summary = Index.getGlobalValueSummary(*V);
584 assert(Summary && "Missing summary for global value");
585 Summary->setNotEligibleToImport();
586 }
587
588 // The linker doesn't know about these LLVM produced values, so we need
589 // to flag them as live in the index to ensure index-based dead value
590 // analysis treats them as live roots of the analysis.
591 setLiveRoot(Index, "llvm.used");
592 setLiveRoot(Index, "llvm.compiler.used");
593 setLiveRoot(Index, "llvm.global_ctors");
594 setLiveRoot(Index, "llvm.global_dtors");
595 setLiveRoot(Index, "llvm.global.annotations");
596
597 for (auto &GlobalList : Index) {
598 // Ignore entries for references that are undefined in the current module.
599 if (GlobalList.second.SummaryList.empty())
600 continue;
601
602 assert(GlobalList.second.SummaryList.size() == 1 &&
603 "Expected module's index to have one summary per GUID");
604 auto &Summary = GlobalList.second.SummaryList[0];
605 if (!IsThinLTO) {
606 Summary->setNotEligibleToImport();
607 continue;
608 }
609
610 bool AllRefsCanBeExternallyReferenced =
611 llvm::all_of(Summary->refs(), [&](const ValueInfo &VI) {
612 return !CantBePromoted.count(VI.getGUID());
613 });
614 if (!AllRefsCanBeExternallyReferenced) {
615 Summary->setNotEligibleToImport();
616 continue;
617 }
618
619 if (auto *FuncSummary = dyn_cast<FunctionSummary>(Summary.get())) {
620 bool AllCallsCanBeExternallyReferenced = llvm::all_of(
621 FuncSummary->calls(), [&](const FunctionSummary::EdgeTy &Edge) {
622 return !CantBePromoted.count(Edge.first.getGUID());
623 });
624 if (!AllCallsCanBeExternallyReferenced)
625 Summary->setNotEligibleToImport();
626 }
627 }
628
629 return Index;
630 }
631
632 AnalysisKey ModuleSummaryIndexAnalysis::Key;
633
634 ModuleSummaryIndex
run(Module & M,ModuleAnalysisManager & AM)635 ModuleSummaryIndexAnalysis::run(Module &M, ModuleAnalysisManager &AM) {
636 ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M);
637 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
638 return buildModuleSummaryIndex(
639 M,
640 [&FAM](const Function &F) {
641 return &FAM.getResult<BlockFrequencyAnalysis>(
642 *const_cast<Function *>(&F));
643 },
644 &PSI);
645 }
646
647 char ModuleSummaryIndexWrapperPass::ID = 0;
648
649 INITIALIZE_PASS_BEGIN(ModuleSummaryIndexWrapperPass, "module-summary-analysis",
650 "Module Summary Analysis", false, true)
INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)651 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
652 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
653 INITIALIZE_PASS_END(ModuleSummaryIndexWrapperPass, "module-summary-analysis",
654 "Module Summary Analysis", false, true)
655
656 ModulePass *llvm::createModuleSummaryIndexWrapperPass() {
657 return new ModuleSummaryIndexWrapperPass();
658 }
659
ModuleSummaryIndexWrapperPass()660 ModuleSummaryIndexWrapperPass::ModuleSummaryIndexWrapperPass()
661 : ModulePass(ID) {
662 initializeModuleSummaryIndexWrapperPassPass(*PassRegistry::getPassRegistry());
663 }
664
runOnModule(Module & M)665 bool ModuleSummaryIndexWrapperPass::runOnModule(Module &M) {
666 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
667 Index.emplace(buildModuleSummaryIndex(
668 M,
669 [this](const Function &F) {
670 return &(this->getAnalysis<BlockFrequencyInfoWrapperPass>(
671 *const_cast<Function *>(&F))
672 .getBFI());
673 },
674 PSI));
675 return false;
676 }
677
doFinalization(Module & M)678 bool ModuleSummaryIndexWrapperPass::doFinalization(Module &M) {
679 Index.reset();
680 return false;
681 }
682
getAnalysisUsage(AnalysisUsage & AU) const683 void ModuleSummaryIndexWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
684 AU.setPreservesAll();
685 AU.addRequired<BlockFrequencyInfoWrapperPass>();
686 AU.addRequired<ProfileSummaryInfoWrapperPass>();
687 }
688