| 1 | //===- MemProfUse.cpp - memory allocation profile use pass --*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file implements the MemProfUsePass which reads memory profiling data |
| 10 | // and uses it to add metadata to instructions to guide optimization. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/Transforms/Instrumentation/MemProfUse.h" |
| 15 | #include "llvm/ADT/DenseSet.h" |
| 16 | #include "llvm/ADT/SmallVector.h" |
| 17 | #include "llvm/ADT/Statistic.h" |
| 18 | #include "llvm/ADT/StringRef.h" |
| 19 | #include "llvm/Analysis/MemoryProfileInfo.h" |
| 20 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 21 | #include "llvm/Analysis/StaticDataProfileInfo.h" |
| 22 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 23 | #include "llvm/IR/DiagnosticInfo.h" |
| 24 | #include "llvm/IR/Function.h" |
| 25 | #include "llvm/IR/IntrinsicInst.h" |
| 26 | #include "llvm/IR/Module.h" |
| 27 | #include "llvm/ProfileData/DataAccessProf.h" |
| 28 | #include "llvm/ProfileData/InstrProf.h" |
| 29 | #include "llvm/ProfileData/InstrProfReader.h" |
| 30 | #include "llvm/ProfileData/MemProfCommon.h" |
| 31 | #include "llvm/Support/BLAKE3.h" |
| 32 | #include "llvm/Support/CommandLine.h" |
| 33 | #include "llvm/Support/Debug.h" |
| 34 | #include "llvm/Support/Format.h" |
| 35 | #include "llvm/Support/HashBuilder.h" |
| 36 | #include "llvm/Support/MD5.h" |
| 37 | #include "llvm/Support/VirtualFileSystem.h" |
| 38 | #include "llvm/Transforms/Utils/LongestCommonSequence.h" |
| 39 | #include <map> |
| 40 | #include <set> |
| 41 | |
| 42 | using namespace llvm; |
| 43 | using namespace llvm::memprof; |
| 44 | |
| 45 | #define DEBUG_TYPE "memprof" |
| 46 | |
| 47 | namespace llvm { |
| 48 | extern cl::opt<bool> PGOWarnMissing; |
| 49 | extern cl::opt<bool> NoPGOWarnMismatch; |
| 50 | extern cl::opt<bool> NoPGOWarnMismatchComdatWeak; |
| 51 | extern cl::opt<bool> AnnotateStringLiteralSectionPrefix; |
| 52 | } // namespace llvm |
| 53 | |
| 54 | // By default disable matching of allocation profiles onto operator new that |
| 55 | // already explicitly pass a hot/cold hint, since we don't currently |
| 56 | // override these hints anyway. |
| 57 | static cl::opt<bool> ClMemProfMatchHotColdNew( |
| 58 | "memprof-match-hot-cold-new" , |
| 59 | cl::desc( |
| 60 | "Match allocation profiles onto existing hot/cold operator new calls" ), |
| 61 | cl::Hidden, cl::init(Val: false)); |
| 62 | |
| 63 | static cl::opt<bool> |
| 64 | ClPrintMemProfMatchInfo("memprof-print-match-info" , |
| 65 | cl::desc("Print matching stats for each allocation " |
| 66 | "context in this module's profiles" ), |
| 67 | cl::Hidden, cl::init(Val: false)); |
| 68 | |
| 69 | static cl::opt<bool> PrintMatchedAllocStack( |
| 70 | "memprof-print-matched-alloc-stack" , |
| 71 | cl::desc("Print full stack context for matched " |
| 72 | "allocations with -memprof-print-match-info." ), |
| 73 | cl::Hidden, cl::init(Val: false)); |
| 74 | |
| 75 | static cl::opt<bool> |
| 76 | PrintFunctionGuids("memprof-print-function-guids" , |
| 77 | cl::desc("Print function GUIDs computed for matching" ), |
| 78 | cl::Hidden, cl::init(Val: false)); |
| 79 | |
| 80 | static cl::opt<bool> |
| 81 | SalvageStaleProfile("memprof-salvage-stale-profile" , |
| 82 | cl::desc("Salvage stale MemProf profile" ), |
| 83 | cl::init(Val: false), cl::Hidden); |
| 84 | |
| 85 | static cl::opt<bool> ClMemProfAttachCalleeGuids( |
| 86 | "memprof-attach-calleeguids" , |
| 87 | cl::desc( |
| 88 | "Attach calleeguids as value profile metadata for indirect calls." ), |
| 89 | cl::init(Val: true), cl::Hidden); |
| 90 | |
| 91 | static cl::opt<unsigned> MinMatchedColdBytePercent( |
| 92 | "memprof-matching-cold-threshold" , cl::init(Val: 100), cl::Hidden, |
| 93 | cl::desc("Min percent of cold bytes matched to hint allocation cold" )); |
| 94 | |
| 95 | static cl::opt<bool> AnnotateStaticDataSectionPrefix( |
| 96 | "memprof-annotate-static-data-prefix" , cl::init(Val: false), cl::Hidden, |
| 97 | cl::desc("If true, annotate the static data section prefix" )); |
| 98 | |
| 99 | // Matching statistics |
| 100 | STATISTIC(NumOfMemProfMissing, "Number of functions without memory profile." ); |
| 101 | STATISTIC(NumOfMemProfMismatch, |
| 102 | "Number of functions having mismatched memory profile hash." ); |
| 103 | STATISTIC(NumOfMemProfFunc, "Number of functions having valid memory profile." ); |
| 104 | STATISTIC(NumOfMemProfAllocContextProfiles, |
| 105 | "Number of alloc contexts in memory profile." ); |
| 106 | STATISTIC(NumOfMemProfCallSiteProfiles, |
| 107 | "Number of callsites in memory profile." ); |
| 108 | STATISTIC(NumOfMemProfMatchedAllocContexts, |
| 109 | "Number of matched memory profile alloc contexts." ); |
| 110 | STATISTIC(NumOfMemProfMatchedAllocs, |
| 111 | "Number of matched memory profile allocs." ); |
| 112 | STATISTIC(NumOfMemProfMatchedCallSites, |
| 113 | "Number of matched memory profile callsites." ); |
| 114 | STATISTIC(NumOfMemProfHotGlobalVars, |
| 115 | "Number of global vars annotated with 'hot' section prefix." ); |
| 116 | STATISTIC(NumOfMemProfColdGlobalVars, |
| 117 | "Number of global vars annotated with 'unlikely' section prefix." ); |
| 118 | STATISTIC(NumOfMemProfUnknownGlobalVars, |
| 119 | "Number of global vars with unknown hotness (no section prefix)." ); |
| 120 | STATISTIC(NumOfMemProfExplicitSectionGlobalVars, |
| 121 | "Number of global vars with user-specified section (not annotated)." ); |
| 122 | |
| 123 | static void addCallsiteMetadata(Instruction &I, |
| 124 | ArrayRef<uint64_t> InlinedCallStack, |
| 125 | LLVMContext &Ctx) { |
| 126 | I.setMetadata(KindID: LLVMContext::MD_callsite, |
| 127 | Node: buildCallstackMetadata(CallStack: InlinedCallStack, Ctx)); |
| 128 | } |
| 129 | |
| 130 | static uint64_t computeStackId(GlobalValue::GUID Function, uint32_t LineOffset, |
| 131 | uint32_t Column) { |
| 132 | llvm::HashBuilder<llvm::TruncatedBLAKE3<8>, llvm::endianness::little> |
| 133 | HashBuilder; |
| 134 | HashBuilder.add(Args: Function, Args: LineOffset, Args: Column); |
| 135 | llvm::BLAKE3Result<8> Hash = HashBuilder.final(); |
| 136 | uint64_t Id; |
| 137 | std::memcpy(dest: &Id, src: Hash.data(), n: sizeof(Hash)); |
| 138 | return Id; |
| 139 | } |
| 140 | |
| 141 | static uint64_t computeStackId(const memprof::Frame &Frame) { |
| 142 | return computeStackId(Function: Frame.Function, LineOffset: Frame.LineOffset, Column: Frame.Column); |
| 143 | } |
| 144 | |
| 145 | static AllocationType getAllocType(const AllocationInfo *AllocInfo) { |
| 146 | return getAllocType(TotalLifetimeAccessDensity: AllocInfo->Info.getTotalLifetimeAccessDensity(), |
| 147 | AllocCount: AllocInfo->Info.getAllocCount(), |
| 148 | TotalLifetime: AllocInfo->Info.getTotalLifetime()); |
| 149 | } |
| 150 | |
| 151 | static AllocationType addCallStack(CallStackTrie &AllocTrie, |
| 152 | const AllocationInfo *AllocInfo, |
| 153 | uint64_t FullStackId) { |
| 154 | SmallVector<uint64_t> StackIds; |
| 155 | for (const auto &StackFrame : AllocInfo->CallStack) |
| 156 | StackIds.push_back(Elt: computeStackId(Frame: StackFrame)); |
| 157 | auto AllocType = getAllocType(AllocInfo); |
| 158 | std::vector<ContextTotalSize> ContextSizeInfo; |
| 159 | if (recordContextSizeInfoForAnalysis()) { |
| 160 | auto TotalSize = AllocInfo->Info.getTotalSize(); |
| 161 | assert(TotalSize); |
| 162 | assert(FullStackId != 0); |
| 163 | ContextSizeInfo.push_back(x: {.FullStackId: FullStackId, .TotalSize: TotalSize}); |
| 164 | } |
| 165 | AllocTrie.addCallStack(AllocType, StackIds, ContextSizeInfo: std::move(ContextSizeInfo)); |
| 166 | return AllocType; |
| 167 | } |
| 168 | |
| 169 | // Return true if InlinedCallStack, computed from a call instruction's debug |
| 170 | // info, is a prefix of ProfileCallStack, a list of Frames from profile data |
| 171 | // (either the allocation data or a callsite). |
| 172 | static bool |
| 173 | stackFrameIncludesInlinedCallStack(ArrayRef<Frame> ProfileCallStack, |
| 174 | ArrayRef<uint64_t> InlinedCallStack) { |
| 175 | return ProfileCallStack.size() >= InlinedCallStack.size() && |
| 176 | llvm::equal(LRange: ProfileCallStack.take_front(N: InlinedCallStack.size()), |
| 177 | RRange&: InlinedCallStack, P: [](const Frame &F, uint64_t StackId) { |
| 178 | return computeStackId(Frame: F) == StackId; |
| 179 | }); |
| 180 | } |
| 181 | |
| 182 | static bool isAllocationWithHotColdVariant(const Function *Callee, |
| 183 | const TargetLibraryInfo &TLI) { |
| 184 | if (!Callee) |
| 185 | return false; |
| 186 | LibFunc Func; |
| 187 | if (!TLI.getLibFunc(FDecl: *Callee, F&: Func)) |
| 188 | return false; |
| 189 | switch (Func) { |
| 190 | case LibFunc_Znwm: |
| 191 | case LibFunc_ZnwmRKSt9nothrow_t: |
| 192 | case LibFunc_ZnwmSt11align_val_t: |
| 193 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t: |
| 194 | case LibFunc_Znam: |
| 195 | case LibFunc_ZnamRKSt9nothrow_t: |
| 196 | case LibFunc_ZnamSt11align_val_t: |
| 197 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t: |
| 198 | case LibFunc_size_returning_new: |
| 199 | case LibFunc_size_returning_new_aligned: |
| 200 | return true; |
| 201 | case LibFunc_Znwm12__hot_cold_t: |
| 202 | case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t: |
| 203 | case LibFunc_ZnwmSt11align_val_t12__hot_cold_t: |
| 204 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 205 | case LibFunc_Znam12__hot_cold_t: |
| 206 | case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t: |
| 207 | case LibFunc_ZnamSt11align_val_t12__hot_cold_t: |
| 208 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 209 | case LibFunc_size_returning_new_hot_cold: |
| 210 | case LibFunc_size_returning_new_aligned_hot_cold: |
| 211 | return ClMemProfMatchHotColdNew; |
| 212 | default: |
| 213 | return false; |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | static void HandleUnsupportedAnnotationKinds(GlobalVariable &GVar, |
| 218 | AnnotationKind Kind) { |
| 219 | assert(Kind != llvm::memprof::AnnotationKind::AnnotationOK && |
| 220 | "Should not handle AnnotationOK here" ); |
| 221 | SmallString<32> Reason; |
| 222 | switch (Kind) { |
| 223 | case llvm::memprof::AnnotationKind::ExplicitSection: |
| 224 | ++NumOfMemProfExplicitSectionGlobalVars; |
| 225 | Reason.append(RHS: "explicit section name" ); |
| 226 | break; |
| 227 | case llvm::memprof::AnnotationKind::DeclForLinker: |
| 228 | Reason.append(RHS: "linker declaration" ); |
| 229 | break; |
| 230 | case llvm::memprof::AnnotationKind::ReservedName: |
| 231 | Reason.append(RHS: "name starts with `llvm.`" ); |
| 232 | break; |
| 233 | default: |
| 234 | llvm_unreachable("Unexpected annotation kind" ); |
| 235 | } |
| 236 | LLVM_DEBUG(dbgs() << "Skip annotation for " << GVar.getName() << " due to " |
| 237 | << Reason << ".\n" ); |
| 238 | } |
| 239 | |
| 240 | // Computes the LLVM version of MD5 hash for the content of a string |
| 241 | // literal. |
| 242 | static std::optional<uint64_t> |
| 243 | getStringContentHash(const GlobalVariable &GVar) { |
| 244 | auto *Initializer = GVar.getInitializer(); |
| 245 | if (!Initializer) |
| 246 | return std::nullopt; |
| 247 | if (auto *C = dyn_cast<ConstantDataSequential>(Val: Initializer)) |
| 248 | if (C->isString()) { |
| 249 | // Note the hash computed for the literal would include the null byte. |
| 250 | return llvm::MD5Hash(Str: C->getAsString()); |
| 251 | } |
| 252 | return std::nullopt; |
| 253 | } |
| 254 | |
| 255 | // Structure for tracking info about matched allocation contexts for use with |
| 256 | // -memprof-print-match-info and -memprof-print-matched-alloc-stack. |
| 257 | struct AllocMatchInfo { |
| 258 | // Total size in bytes of matched context. |
| 259 | uint64_t TotalSize = 0; |
| 260 | // Matched allocation's type. |
| 261 | AllocationType AllocType = AllocationType::None; |
| 262 | // Number of frames matched to the allocation itself (values will be >1 in |
| 263 | // cases where allocation was already inlined). Use a set because there can |
| 264 | // be multiple inlined instances and each may have a different inline depth. |
| 265 | // Use std::set to iterate in sorted order when printing. |
| 266 | std::set<unsigned> MatchedFramesSet; |
| 267 | // The full call stack of the allocation, for cases where requested via |
| 268 | // -memprof-print-matched-alloc-stack. |
| 269 | std::vector<Frame> CallStack; |
| 270 | |
| 271 | // Caller responsible for inserting the matched frames and the call stack when |
| 272 | // appropriate. |
| 273 | AllocMatchInfo(uint64_t TotalSize, AllocationType AllocType) |
| 274 | : TotalSize(TotalSize), AllocType(AllocType) {} |
| 275 | }; |
| 276 | |
| 277 | DenseMap<uint64_t, SmallVector<CallEdgeTy, 0>> |
| 278 | memprof::(Module &M, const TargetLibraryInfo &TLI, |
| 279 | function_ref<bool(uint64_t)> IsPresentInProfile) { |
| 280 | DenseMap<uint64_t, SmallVector<CallEdgeTy, 0>> Calls; |
| 281 | |
| 282 | auto GetOffset = [](const DILocation *DIL) { |
| 283 | return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) & |
| 284 | 0xffff; |
| 285 | }; |
| 286 | |
| 287 | for (Function &F : M) { |
| 288 | if (F.isDeclaration()) |
| 289 | continue; |
| 290 | |
| 291 | for (auto &BB : F) { |
| 292 | for (auto &I : BB) { |
| 293 | if (!isa<CallBase>(Val: &I) || isa<IntrinsicInst>(Val: &I)) |
| 294 | continue; |
| 295 | |
| 296 | auto *CB = dyn_cast<CallBase>(Val: &I); |
| 297 | auto *CalledFunction = CB->getCalledFunction(); |
| 298 | // Disregard indirect calls and intrinsics. |
| 299 | if (!CalledFunction || CalledFunction->isIntrinsic()) |
| 300 | continue; |
| 301 | |
| 302 | StringRef CalleeName = CalledFunction->getName(); |
| 303 | // True if we are calling a heap allocation function that supports |
| 304 | // hot/cold variants. |
| 305 | bool IsAlloc = isAllocationWithHotColdVariant(Callee: CalledFunction, TLI); |
| 306 | // True for the first iteration below, indicating that we are looking at |
| 307 | // a leaf node. |
| 308 | bool IsLeaf = true; |
| 309 | for (const DILocation *DIL = I.getDebugLoc(); DIL; |
| 310 | DIL = DIL->getInlinedAt()) { |
| 311 | StringRef CallerName = DIL->getSubprogramLinkageName(); |
| 312 | assert(!CallerName.empty() && |
| 313 | "Be sure to enable -fdebug-info-for-profiling" ); |
| 314 | uint64_t CallerGUID = memprof::getGUID(FunctionName: CallerName); |
| 315 | uint64_t CalleeGUID = memprof::getGUID(FunctionName: CalleeName); |
| 316 | // Pretend that we are calling a function with GUID == 0 if we are |
| 317 | // in the inline stack leading to a heap allocation function. |
| 318 | if (IsAlloc) { |
| 319 | if (IsLeaf) { |
| 320 | // For leaf nodes, set CalleeGUID to 0 without consulting |
| 321 | // IsPresentInProfile. |
| 322 | CalleeGUID = 0; |
| 323 | } else if (!IsPresentInProfile(CalleeGUID)) { |
| 324 | // In addition to the leaf case above, continue to set CalleeGUID |
| 325 | // to 0 as long as we don't see CalleeGUID in the profile. |
| 326 | CalleeGUID = 0; |
| 327 | } else { |
| 328 | // Once we encounter a callee that exists in the profile, stop |
| 329 | // setting CalleeGUID to 0. |
| 330 | IsAlloc = false; |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | LineLocation Loc = {GetOffset(DIL), DIL->getColumn()}; |
| 335 | Calls[CallerGUID].emplace_back(Args&: Loc, Args&: CalleeGUID); |
| 336 | CalleeName = CallerName; |
| 337 | IsLeaf = false; |
| 338 | } |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | // Sort each call list by the source location. |
| 344 | for (auto &[CallerGUID, CallList] : Calls) { |
| 345 | llvm::sort(C&: CallList); |
| 346 | CallList.erase(CS: llvm::unique(R&: CallList), CE: CallList.end()); |
| 347 | } |
| 348 | |
| 349 | return Calls; |
| 350 | } |
| 351 | |
| 352 | DenseMap<uint64_t, LocToLocMap> |
| 353 | memprof::computeUndriftMap(Module &M, IndexedInstrProfReader *MemProfReader, |
| 354 | const TargetLibraryInfo &TLI) { |
| 355 | DenseMap<uint64_t, LocToLocMap> UndriftMaps; |
| 356 | |
| 357 | DenseMap<uint64_t, SmallVector<memprof::CallEdgeTy, 0>> CallsFromProfile = |
| 358 | MemProfReader->getMemProfCallerCalleePairs(); |
| 359 | DenseMap<uint64_t, SmallVector<memprof::CallEdgeTy, 0>> CallsFromIR = |
| 360 | extractCallsFromIR(M, TLI, IsPresentInProfile: [&](uint64_t GUID) { |
| 361 | return CallsFromProfile.contains(Val: GUID); |
| 362 | }); |
| 363 | |
| 364 | // Compute an undrift map for each CallerGUID. |
| 365 | for (const auto &[CallerGUID, IRAnchors] : CallsFromIR) { |
| 366 | auto It = CallsFromProfile.find(Val: CallerGUID); |
| 367 | if (It == CallsFromProfile.end()) |
| 368 | continue; |
| 369 | const auto &ProfileAnchors = It->second; |
| 370 | |
| 371 | LocToLocMap Matchings; |
| 372 | longestCommonSequence<LineLocation, GlobalValue::GUID>( |
| 373 | AnchorList1: ProfileAnchors, AnchorList2: IRAnchors, FunctionMatchesProfile: std::equal_to<GlobalValue::GUID>(), |
| 374 | InsertMatching: [&](LineLocation A, LineLocation B) { Matchings.try_emplace(Key: A, Args&: B); }); |
| 375 | [[maybe_unused]] bool Inserted = |
| 376 | UndriftMaps.try_emplace(Key: CallerGUID, Args: std::move(Matchings)).second; |
| 377 | |
| 378 | // The insertion must succeed because we visit each GUID exactly once. |
| 379 | assert(Inserted); |
| 380 | } |
| 381 | |
| 382 | return UndriftMaps; |
| 383 | } |
| 384 | |
| 385 | // Given a MemProfRecord, undrift all the source locations present in the |
| 386 | // record in place. |
| 387 | static void |
| 388 | undriftMemProfRecord(const DenseMap<uint64_t, LocToLocMap> &UndriftMaps, |
| 389 | memprof::MemProfRecord &MemProfRec) { |
| 390 | // Undrift a call stack in place. |
| 391 | auto UndriftCallStack = [&](std::vector<Frame> &CallStack) { |
| 392 | for (auto &F : CallStack) { |
| 393 | auto I = UndriftMaps.find(Val: F.Function); |
| 394 | if (I == UndriftMaps.end()) |
| 395 | continue; |
| 396 | auto J = I->second.find(Val: LineLocation(F.LineOffset, F.Column)); |
| 397 | if (J == I->second.end()) |
| 398 | continue; |
| 399 | auto &NewLoc = J->second; |
| 400 | F.LineOffset = NewLoc.LineOffset; |
| 401 | F.Column = NewLoc.Column; |
| 402 | } |
| 403 | }; |
| 404 | |
| 405 | for (auto &AS : MemProfRec.AllocSites) |
| 406 | UndriftCallStack(AS.CallStack); |
| 407 | |
| 408 | for (auto &CS : MemProfRec.CallSites) |
| 409 | UndriftCallStack(CS.Frames); |
| 410 | } |
| 411 | |
| 412 | // Helper function to process CalleeGuids and create value profile metadata |
| 413 | static void addVPMetadata(Module &M, Instruction &I, |
| 414 | ArrayRef<GlobalValue::GUID> CalleeGuids) { |
| 415 | if (!ClMemProfAttachCalleeGuids || CalleeGuids.empty()) |
| 416 | return; |
| 417 | |
| 418 | // Prepare the vector of value data, initializing from any existing |
| 419 | // value-profile metadata present on the instruction so that we merge the |
| 420 | // new CalleeGuids into the existing entries. |
| 421 | SmallVector<InstrProfValueData> VDs; |
| 422 | uint64_t TotalCount = 0; |
| 423 | |
| 424 | if (I.getMetadata(KindID: LLVMContext::MD_prof)) { |
| 425 | // Read all existing entries so we can merge them. Use a large |
| 426 | // MaxNumValueData to retrieve all existing entries. |
| 427 | VDs = getValueProfDataFromInst(Inst: I, ValueKind: IPVK_IndirectCallTarget, |
| 428 | /*MaxNumValueData=*/UINT32_MAX, TotalC&: TotalCount); |
| 429 | } |
| 430 | |
| 431 | // Save the original size for use later in detecting whether any were added. |
| 432 | const size_t OriginalSize = VDs.size(); |
| 433 | |
| 434 | // Initialize the set of existing guids with the original list. |
| 435 | DenseSet<uint64_t> ExistingValues( |
| 436 | llvm::from_range, |
| 437 | llvm::map_range( |
| 438 | C&: VDs, F: [](const InstrProfValueData &Entry) { return Entry.Value; })); |
| 439 | |
| 440 | // Merge CalleeGuids into list of existing VDs, by appending any that are not |
| 441 | // already included. |
| 442 | VDs.reserve(N: OriginalSize + CalleeGuids.size()); |
| 443 | for (auto G : CalleeGuids) { |
| 444 | if (!ExistingValues.insert(V: G).second) |
| 445 | continue; |
| 446 | InstrProfValueData NewEntry; |
| 447 | NewEntry.Value = G; |
| 448 | // For MemProf, we don't have actual call counts, so we assign |
| 449 | // a weight of 1 to each potential target. |
| 450 | // TODO: Consider making this weight configurable or increasing it to |
| 451 | // improve effectiveness for ICP. |
| 452 | NewEntry.Count = 1; |
| 453 | TotalCount += NewEntry.Count; |
| 454 | VDs.push_back(Elt: NewEntry); |
| 455 | } |
| 456 | |
| 457 | // Update the VP metadata if we added any new callee GUIDs to the list. |
| 458 | assert(VDs.size() >= OriginalSize); |
| 459 | if (VDs.size() == OriginalSize) |
| 460 | return; |
| 461 | |
| 462 | // First clear the existing !prof. |
| 463 | I.setMetadata(KindID: LLVMContext::MD_prof, Node: nullptr); |
| 464 | |
| 465 | // No need to sort the updated VDs as all appended entries have the same count |
| 466 | // of 1, which is no larger than any existing entries. The incoming list of |
| 467 | // CalleeGuids should already be deterministic for a given profile. |
| 468 | annotateValueSite(M, Inst&: I, VDs, Sum: TotalCount, ValueKind: IPVK_IndirectCallTarget, MaxMDCount: VDs.size()); |
| 469 | } |
| 470 | |
| 471 | static void handleAllocSite( |
| 472 | Instruction &I, CallBase *CI, ArrayRef<uint64_t> InlinedCallStack, |
| 473 | LLVMContext &Ctx, OptimizationRemarkEmitter &ORE, uint64_t MaxColdSize, |
| 474 | const std::set<const AllocationInfo *> &AllocInfoSet, |
| 475 | std::map<uint64_t, AllocMatchInfo> &FullStackIdToAllocMatchInfo) { |
| 476 | // TODO: Remove this once the profile creation logic deduplicates contexts |
| 477 | // that are the same other than the IsInlineFrame bool. Until then, keep the |
| 478 | // largest. |
| 479 | DenseMap<uint64_t, const AllocationInfo *> UniqueFullContextIdAllocInfo; |
| 480 | for (auto *AllocInfo : AllocInfoSet) { |
| 481 | auto FullStackId = computeFullStackId(CallStack: AllocInfo->CallStack); |
| 482 | auto [It, Inserted] = |
| 483 | UniqueFullContextIdAllocInfo.insert(KV: {FullStackId, AllocInfo}); |
| 484 | // If inserted entry, done. |
| 485 | if (Inserted) |
| 486 | continue; |
| 487 | // Keep the larger one, or the noncold one if they are the same size. |
| 488 | auto CurSize = It->second->Info.getTotalSize(); |
| 489 | auto NewSize = AllocInfo->Info.getTotalSize(); |
| 490 | if ((CurSize > NewSize) || |
| 491 | (CurSize == NewSize && |
| 492 | getAllocType(AllocInfo) != AllocationType::NotCold)) |
| 493 | continue; |
| 494 | It->second = AllocInfo; |
| 495 | } |
| 496 | // We may match this instruction's location list to multiple MIB |
| 497 | // contexts. Add them to a Trie specialized for trimming the contexts to |
| 498 | // the minimal needed to disambiguate contexts with unique behavior. |
| 499 | CallStackTrie AllocTrie(&ORE, MaxColdSize); |
| 500 | uint64_t TotalSize = 0; |
| 501 | uint64_t TotalColdSize = 0; |
| 502 | for (auto &[FullStackId, AllocInfo] : UniqueFullContextIdAllocInfo) { |
| 503 | // Check the full inlined call stack against this one. |
| 504 | // If we found and thus matched all frames on the call, include |
| 505 | // this MIB. |
| 506 | if (stackFrameIncludesInlinedCallStack(ProfileCallStack: AllocInfo->CallStack, |
| 507 | InlinedCallStack)) { |
| 508 | NumOfMemProfMatchedAllocContexts++; |
| 509 | auto AllocType = addCallStack(AllocTrie, AllocInfo, FullStackId); |
| 510 | TotalSize += AllocInfo->Info.getTotalSize(); |
| 511 | if (AllocType == AllocationType::Cold) |
| 512 | TotalColdSize += AllocInfo->Info.getTotalSize(); |
| 513 | // Record information about the allocation if match info printing |
| 514 | // was requested. |
| 515 | if (ClPrintMemProfMatchInfo) { |
| 516 | assert(FullStackId != 0); |
| 517 | auto [Iter, Inserted] = FullStackIdToAllocMatchInfo.try_emplace( |
| 518 | k: FullStackId, |
| 519 | args: AllocMatchInfo(AllocInfo->Info.getTotalSize(), AllocType)); |
| 520 | // Always insert the new matched frame count, since it may differ. |
| 521 | Iter->second.MatchedFramesSet.insert(x: InlinedCallStack.size()); |
| 522 | if (Inserted && PrintMatchedAllocStack) |
| 523 | Iter->second.CallStack.insert(position: Iter->second.CallStack.begin(), |
| 524 | first: AllocInfo->CallStack.begin(), |
| 525 | last: AllocInfo->CallStack.end()); |
| 526 | } |
| 527 | ORE.emit( |
| 528 | OptDiag: OptimizationRemark(DEBUG_TYPE, "MemProfUse" , CI) |
| 529 | << ore::NV("AllocationCall" , CI) << " in function " |
| 530 | << ore::NV("Caller" , CI->getFunction()) |
| 531 | << " matched alloc context with alloc type " |
| 532 | << ore::NV("Attribute" , getAllocTypeAttributeString(Type: AllocType)) |
| 533 | << " total size " << ore::NV("Size" , AllocInfo->Info.getTotalSize()) |
| 534 | << " full context id " << ore::NV("Context" , FullStackId) |
| 535 | << " frame count " << ore::NV("Frames" , InlinedCallStack.size())); |
| 536 | } |
| 537 | } |
| 538 | // If the threshold for the percent of cold bytes is less than 100%, |
| 539 | // and not all bytes are cold, see if we should still hint this |
| 540 | // allocation as cold without context sensitivity. |
| 541 | if (TotalColdSize < TotalSize && MinMatchedColdBytePercent < 100 && |
| 542 | TotalColdSize * 100 >= MinMatchedColdBytePercent * TotalSize) { |
| 543 | AllocTrie.addSingleAllocTypeAttribute(CI, AT: AllocationType::Cold, Descriptor: "dominant" ); |
| 544 | return; |
| 545 | } |
| 546 | |
| 547 | // We might not have matched any to the full inlined call stack. |
| 548 | // But if we did, create and attach metadata, or a function attribute if |
| 549 | // all contexts have identical profiled behavior. |
| 550 | if (!AllocTrie.empty()) { |
| 551 | NumOfMemProfMatchedAllocs++; |
| 552 | // MemprofMDAttached will be false if a function attribute was |
| 553 | // attached. |
| 554 | bool MemprofMDAttached = AllocTrie.buildAndAttachMIBMetadata(CI); |
| 555 | assert(MemprofMDAttached == I.hasMetadata(LLVMContext::MD_memprof)); |
| 556 | if (MemprofMDAttached) { |
| 557 | // Add callsite metadata for the instruction's location list so that |
| 558 | // it simpler later on to identify which part of the MIB contexts |
| 559 | // are from this particular instruction (including during inlining, |
| 560 | // when the callsite metadata will be updated appropriately). |
| 561 | // FIXME: can this be changed to strip out the matching stack |
| 562 | // context ids from the MIB contexts and not add any callsite |
| 563 | // metadata here to save space? |
| 564 | addCallsiteMetadata(I, InlinedCallStack, Ctx); |
| 565 | } |
| 566 | } |
| 567 | } |
| 568 | |
| 569 | // Helper struct for maintaining refs to callsite data. As an alternative we |
| 570 | // could store a pointer to the CallSiteInfo struct but we also need the frame |
| 571 | // index. Using ArrayRefs instead makes it a little easier to read. |
| 572 | struct CallSiteEntry { |
| 573 | // Subset of frames for the corresponding CallSiteInfo. |
| 574 | ArrayRef<Frame> Frames; |
| 575 | // Potential targets for indirect calls. |
| 576 | ArrayRef<GlobalValue::GUID> CalleeGuids; |
| 577 | }; |
| 578 | |
| 579 | static void handleCallSite(Instruction &I, const Function *CalledFunction, |
| 580 | ArrayRef<uint64_t> InlinedCallStack, |
| 581 | const std::vector<CallSiteEntry> &CallSiteEntries, |
| 582 | Module &M, |
| 583 | std::set<std::vector<uint64_t>> &MatchedCallSites, |
| 584 | OptimizationRemarkEmitter &ORE) { |
| 585 | auto &Ctx = M.getContext(); |
| 586 | // Set of Callee GUIDs to attach to indirect calls. We accumulate all of them |
| 587 | // to support cases where the instuction's inlined frames match multiple call |
| 588 | // site entries, which can happen if the profile was collected from a binary |
| 589 | // where this instruction was eventually inlined into multiple callers. |
| 590 | SetVector<GlobalValue::GUID> CalleeGuids; |
| 591 | bool CallsiteMDAdded = false; |
| 592 | for (const auto &CallSiteEntry : CallSiteEntries) { |
| 593 | // If we found and thus matched all frames on the call, create and |
| 594 | // attach call stack metadata. |
| 595 | if (stackFrameIncludesInlinedCallStack(ProfileCallStack: CallSiteEntry.Frames, |
| 596 | InlinedCallStack)) { |
| 597 | NumOfMemProfMatchedCallSites++; |
| 598 | // Only need to find one with a matching call stack and add a single |
| 599 | // callsite metadata. |
| 600 | if (!CallsiteMDAdded) { |
| 601 | addCallsiteMetadata(I, InlinedCallStack, Ctx); |
| 602 | |
| 603 | // Accumulate call site matching information upon request. |
| 604 | if (ClPrintMemProfMatchInfo) { |
| 605 | std::vector<uint64_t> CallStack; |
| 606 | append_range(C&: CallStack, R&: InlinedCallStack); |
| 607 | MatchedCallSites.insert(x: std::move(CallStack)); |
| 608 | } |
| 609 | OptimizationRemark (DEBUG_TYPE, "MemProfUse" , &I); |
| 610 | Remark << ore::NV("CallSite" , &I) << " in function " |
| 611 | << ore::NV("Caller" , I.getFunction()) |
| 612 | << " matched callsite with frame count " |
| 613 | << ore::NV("Frames" , InlinedCallStack.size()) |
| 614 | << " and stack ids" ; |
| 615 | for (uint64_t StackId : InlinedCallStack) |
| 616 | Remark << " " << ore::NV("StackId" , StackId); |
| 617 | ORE.emit(OptDiag&: Remark); |
| 618 | |
| 619 | // If this is a direct call, we're done. |
| 620 | if (CalledFunction) |
| 621 | break; |
| 622 | CallsiteMDAdded = true; |
| 623 | } |
| 624 | |
| 625 | assert(!CalledFunction && "Didn't expect direct call" ); |
| 626 | |
| 627 | // Collect Callee GUIDs from all matching CallSiteEntries. |
| 628 | CalleeGuids.insert(Start: CallSiteEntry.CalleeGuids.begin(), |
| 629 | End: CallSiteEntry.CalleeGuids.end()); |
| 630 | } |
| 631 | } |
| 632 | // Try to attach indirect call metadata if possible. |
| 633 | addVPMetadata(M, I, CalleeGuids: CalleeGuids.getArrayRef()); |
| 634 | } |
| 635 | |
| 636 | // Dump inline call stack for debugging purposes. |
| 637 | static void (Instruction &I, CallBase *CI, |
| 638 | OptimizationRemarkEmitter &ORE, |
| 639 | DenseSet<uint64_t> &SeenFrames, |
| 640 | DenseSet<uint64_t> &SeenStacks, |
| 641 | bool ProfileHasColumns) { |
| 642 | auto GetOffset = [](const DILocation *DIL) { |
| 643 | return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) & |
| 644 | 0xffff; |
| 645 | }; |
| 646 | |
| 647 | // Dump frame info. Frames are deduplicated using FrameID. |
| 648 | std::string CallStack; |
| 649 | raw_string_ostream CallStackOS(CallStack); |
| 650 | bool First = true; |
| 651 | for (const DILocation *DIL = I.getDebugLoc(); DIL; |
| 652 | DIL = DIL->getInlinedAt()) { |
| 653 | StringRef Name = DIL->getScope()->getSubprogram()->getLinkageName(); |
| 654 | if (Name.empty()) |
| 655 | Name = DIL->getScope()->getSubprogram()->getName(); |
| 656 | auto CalleeGUID = Function::getGUIDAssumingExternalLinkage(GlobalName: Name); |
| 657 | uint64_t FrameID = computeStackId(Function: CalleeGUID, LineOffset: GetOffset(DIL), |
| 658 | Column: ProfileHasColumns ? DIL->getColumn() : 0); |
| 659 | if (SeenFrames.insert(V: FrameID).second) { |
| 660 | std::string DictMsg; |
| 661 | raw_string_ostream DictOS(DictMsg); |
| 662 | DictOS << "frame: " << FrameID << " " << Name << ":" << GetOffset(DIL) |
| 663 | << ":" << (ProfileHasColumns ? DIL->getColumn() : 0); |
| 664 | ORE.emit(OptDiag: OptimizationRemarkAnalysis(DEBUG_TYPE, "MemProfUse" , CI) |
| 665 | << DictOS.str()); |
| 666 | } |
| 667 | |
| 668 | if (First) |
| 669 | First = false; |
| 670 | else |
| 671 | CallStackOS << "," ; |
| 672 | CallStackOS << FrameID; |
| 673 | } |
| 674 | |
| 675 | // Dump inline call stack info. Stacks are deduplicated using StackHash. |
| 676 | uint64_t StackHash = llvm::MD5Hash(Str: CallStack); |
| 677 | if (SeenStacks.insert(V: StackHash).second) { |
| 678 | std::string Msg; |
| 679 | raw_string_ostream OS(Msg); |
| 680 | OS << "inline call stack: " << CallStack; |
| 681 | ORE.emit(OptDiag: OptimizationRemarkAnalysis(DEBUG_TYPE, "MemProfUse" , CI) |
| 682 | << OS.str()); |
| 683 | } |
| 684 | } |
| 685 | |
| 686 | static void |
| 687 | readMemprof(Module &M, Function &F, IndexedInstrProfReader *MemProfReader, |
| 688 | const TargetLibraryInfo &TLI, |
| 689 | std::map<uint64_t, AllocMatchInfo> &FullStackIdToAllocMatchInfo, |
| 690 | std::set<std::vector<uint64_t>> &MatchedCallSites, |
| 691 | DenseMap<uint64_t, LocToLocMap> &UndriftMaps, |
| 692 | OptimizationRemarkEmitter &ORE, uint64_t MaxColdSize, |
| 693 | DenseSet<uint64_t> &SeenStacks, DenseSet<uint64_t> &SeenFrames) { |
| 694 | auto &Ctx = M.getContext(); |
| 695 | // Previously we used getIRPGOFuncName() here. If F is local linkage, |
| 696 | // getIRPGOFuncName() returns FuncName with prefix 'FileName;'. But |
| 697 | // llvm-profdata uses FuncName in dwarf to create GUID which doesn't |
| 698 | // contain FileName's prefix. It caused local linkage function can't |
| 699 | // find MemProfRecord. So we use getName() now. |
| 700 | // 'unique-internal-linkage-names' can make MemProf work better for local |
| 701 | // linkage function. |
| 702 | auto FuncName = F.getName(); |
| 703 | auto FuncGUID = Function::getGUIDAssumingExternalLinkage(GlobalName: FuncName); |
| 704 | if (PrintFunctionGuids) |
| 705 | errs() << "MemProf: Function GUID " << FuncGUID << " is " << FuncName |
| 706 | << "\n" ; |
| 707 | std::optional<memprof::MemProfRecord> MemProfRec; |
| 708 | auto Err = MemProfReader->getMemProfRecord(FuncNameHash: FuncGUID).moveInto(Value&: MemProfRec); |
| 709 | if (Err) { |
| 710 | handleAllErrors(E: std::move(Err), Handlers: [&](const InstrProfError &IPE) { |
| 711 | auto Err = IPE.get(); |
| 712 | bool SkipWarning = false; |
| 713 | LLVM_DEBUG(dbgs() << "Error in reading profile for Func " << FuncName |
| 714 | << ": " ); |
| 715 | if (Err == instrprof_error::unknown_function) { |
| 716 | NumOfMemProfMissing++; |
| 717 | SkipWarning = !PGOWarnMissing; |
| 718 | LLVM_DEBUG(dbgs() << "unknown function" ); |
| 719 | } else if (Err == instrprof_error::hash_mismatch) { |
| 720 | NumOfMemProfMismatch++; |
| 721 | SkipWarning = |
| 722 | NoPGOWarnMismatch || |
| 723 | (NoPGOWarnMismatchComdatWeak && |
| 724 | (F.hasComdat() || |
| 725 | F.getLinkage() == GlobalValue::AvailableExternallyLinkage)); |
| 726 | LLVM_DEBUG(dbgs() << "hash mismatch (skip=" << SkipWarning << ")" ); |
| 727 | } |
| 728 | |
| 729 | if (SkipWarning) |
| 730 | return; |
| 731 | |
| 732 | std::string Msg = (IPE.message() + Twine(" " ) + F.getName().str() + |
| 733 | Twine(" Hash = " ) + std::to_string(val: FuncGUID)) |
| 734 | .str(); |
| 735 | |
| 736 | Ctx.diagnose( |
| 737 | DI: DiagnosticInfoPGOProfile(M.getName().data(), Msg, DS_Warning)); |
| 738 | }); |
| 739 | return; |
| 740 | } |
| 741 | |
| 742 | NumOfMemProfFunc++; |
| 743 | |
| 744 | // If requested, undrfit MemProfRecord so that the source locations in it |
| 745 | // match those in the IR. |
| 746 | if (SalvageStaleProfile) |
| 747 | undriftMemProfRecord(UndriftMaps, MemProfRec&: *MemProfRec); |
| 748 | |
| 749 | // Detect if there are non-zero column numbers in the profile. If not, |
| 750 | // treat all column numbers as 0 when matching (i.e. ignore any non-zero |
| 751 | // columns in the IR). The profiled binary might have been built with |
| 752 | // column numbers disabled, for example. |
| 753 | bool ProfileHasColumns = false; |
| 754 | |
| 755 | // Build maps of the location hash to all profile data with that leaf location |
| 756 | // (allocation info and the callsites). |
| 757 | std::map<uint64_t, std::set<const AllocationInfo *>> LocHashToAllocInfo; |
| 758 | |
| 759 | // For the callsites we need to record slices of the frame array (see comments |
| 760 | // below where the map entries are added) along with their CalleeGuids. |
| 761 | std::map<uint64_t, std::vector<CallSiteEntry>> LocHashToCallSites; |
| 762 | for (auto &AI : MemProfRec->AllocSites) { |
| 763 | NumOfMemProfAllocContextProfiles++; |
| 764 | // Associate the allocation info with the leaf frame. The later matching |
| 765 | // code will match any inlined call sequences in the IR with a longer prefix |
| 766 | // of call stack frames. |
| 767 | uint64_t StackId = computeStackId(Frame: AI.CallStack[0]); |
| 768 | LocHashToAllocInfo[StackId].insert(x: &AI); |
| 769 | ProfileHasColumns |= AI.CallStack[0].Column; |
| 770 | } |
| 771 | for (auto &CS : MemProfRec->CallSites) { |
| 772 | NumOfMemProfCallSiteProfiles++; |
| 773 | // Need to record all frames from leaf up to and including this function, |
| 774 | // as any of these may or may not have been inlined at this point. |
| 775 | unsigned Idx = 0; |
| 776 | for (auto &StackFrame : CS.Frames) { |
| 777 | uint64_t StackId = computeStackId(Frame: StackFrame); |
| 778 | ArrayRef<Frame> FrameSlice = ArrayRef<Frame>(CS.Frames).drop_front(N: Idx++); |
| 779 | // The callee guids for the slice containing all frames (due to the |
| 780 | // increment above Idx is now 1) comes from the CalleeGuids recorded in |
| 781 | // the CallSite. For the slices not containing the leaf-most frame, the |
| 782 | // callee guid is simply the function GUID of the prior frame. |
| 783 | LocHashToCallSites[StackId].push_back( |
| 784 | x: {.Frames: FrameSlice, .CalleeGuids: (Idx == 1 ? CS.CalleeGuids |
| 785 | : ArrayRef<GlobalValue::GUID>( |
| 786 | CS.Frames[Idx - 2].Function))}); |
| 787 | |
| 788 | ProfileHasColumns |= StackFrame.Column; |
| 789 | // Once we find this function, we can stop recording. |
| 790 | if (StackFrame.Function == FuncGUID) |
| 791 | break; |
| 792 | } |
| 793 | assert(Idx <= CS.Frames.size() && CS.Frames[Idx - 1].Function == FuncGUID); |
| 794 | } |
| 795 | |
| 796 | auto GetOffset = [](const DILocation *DIL) { |
| 797 | return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) & |
| 798 | 0xffff; |
| 799 | }; |
| 800 | |
| 801 | // Now walk the instructions, looking up the associated profile data using |
| 802 | // debug locations. |
| 803 | for (auto &BB : F) { |
| 804 | for (auto &I : BB) { |
| 805 | if (I.isDebugOrPseudoInst()) |
| 806 | continue; |
| 807 | // We are only interested in calls (allocation or interior call stack |
| 808 | // context calls). |
| 809 | auto *CI = dyn_cast<CallBase>(Val: &I); |
| 810 | if (!CI) |
| 811 | continue; |
| 812 | auto *CalledFunction = CI->getCalledFunction(); |
| 813 | if (CalledFunction && CalledFunction->isIntrinsic()) |
| 814 | continue; |
| 815 | |
| 816 | if (ORE.allowExtraAnalysis(DEBUG_TYPE)) |
| 817 | dumpInlineCallStack(I, CI, ORE, SeenFrames, SeenStacks, |
| 818 | ProfileHasColumns); |
| 819 | |
| 820 | // List of call stack ids computed from the location hashes on debug |
| 821 | // locations (leaf to inlined at root). |
| 822 | SmallVector<uint64_t, 8> InlinedCallStack; |
| 823 | // Was the leaf location found in one of the profile maps? |
| 824 | bool LeafFound = false; |
| 825 | // If leaf was found in a map, iterators pointing to its location in both |
| 826 | // of the maps. It might exist in neither, one, or both (the latter case |
| 827 | // can happen because we don't currently have discriminators to |
| 828 | // distinguish the case when a single line/col maps to both an allocation |
| 829 | // and another callsite). |
| 830 | auto AllocInfoIter = LocHashToAllocInfo.end(); |
| 831 | auto CallSitesIter = LocHashToCallSites.end(); |
| 832 | for (const DILocation *DIL = I.getDebugLoc(); DIL != nullptr; |
| 833 | DIL = DIL->getInlinedAt()) { |
| 834 | // Use C++ linkage name if possible. Need to compile with |
| 835 | // -fdebug-info-for-profiling to get linkage name. |
| 836 | StringRef Name = DIL->getScope()->getSubprogram()->getLinkageName(); |
| 837 | if (Name.empty()) |
| 838 | Name = DIL->getScope()->getSubprogram()->getName(); |
| 839 | auto CalleeGUID = Function::getGUIDAssumingExternalLinkage(GlobalName: Name); |
| 840 | auto StackId = computeStackId(Function: CalleeGUID, LineOffset: GetOffset(DIL), |
| 841 | Column: ProfileHasColumns ? DIL->getColumn() : 0); |
| 842 | // Check if we have found the profile's leaf frame. If yes, collect |
| 843 | // the rest of the call's inlined context starting here. If not, see if |
| 844 | // we find a match further up the inlined context (in case the profile |
| 845 | // was missing debug frames at the leaf). |
| 846 | if (!LeafFound) { |
| 847 | AllocInfoIter = LocHashToAllocInfo.find(x: StackId); |
| 848 | CallSitesIter = LocHashToCallSites.find(x: StackId); |
| 849 | if (AllocInfoIter != LocHashToAllocInfo.end() || |
| 850 | CallSitesIter != LocHashToCallSites.end()) |
| 851 | LeafFound = true; |
| 852 | } |
| 853 | if (LeafFound) |
| 854 | InlinedCallStack.push_back(Elt: StackId); |
| 855 | } |
| 856 | // If leaf not in either of the maps, skip inst. |
| 857 | if (!LeafFound) |
| 858 | continue; |
| 859 | |
| 860 | // First add !memprof metadata from allocation info, if we found the |
| 861 | // instruction's leaf location in that map, and if the rest of the |
| 862 | // instruction's locations match the prefix Frame locations on an |
| 863 | // allocation context with the same leaf. |
| 864 | if (AllocInfoIter != LocHashToAllocInfo.end() && |
| 865 | // Only consider allocations which support hinting. |
| 866 | isAllocationWithHotColdVariant(Callee: CI->getCalledFunction(), TLI)) |
| 867 | handleAllocSite(I, CI, InlinedCallStack, Ctx, ORE, MaxColdSize, |
| 868 | AllocInfoSet: AllocInfoIter->second, FullStackIdToAllocMatchInfo); |
| 869 | else if (CallSitesIter != LocHashToCallSites.end()) |
| 870 | // Otherwise, add callsite metadata. If we reach here then we found the |
| 871 | // instruction's leaf location in the callsites map and not the |
| 872 | // allocation map. |
| 873 | handleCallSite(I, CalledFunction, InlinedCallStack, |
| 874 | CallSiteEntries: CallSitesIter->second, M, MatchedCallSites, ORE); |
| 875 | } |
| 876 | } |
| 877 | } |
| 878 | |
| 879 | MemProfUsePass::MemProfUsePass(std::string MemoryProfileFile, |
| 880 | IntrusiveRefCntPtr<vfs::FileSystem> FS) |
| 881 | : MemoryProfileFileName(MemoryProfileFile), FS(FS) { |
| 882 | if (!FS) |
| 883 | this->FS = vfs::getRealFileSystem(); |
| 884 | } |
| 885 | |
| 886 | PreservedAnalyses MemProfUsePass::run(Module &M, ModuleAnalysisManager &AM) { |
| 887 | // Return immediately if the module doesn't contain any function or global |
| 888 | // variables. |
| 889 | if (M.empty() && M.globals().empty()) |
| 890 | return PreservedAnalyses::all(); |
| 891 | |
| 892 | LLVM_DEBUG(dbgs() << "Read in memory profile:\n" ); |
| 893 | auto &Ctx = M.getContext(); |
| 894 | auto ReaderOrErr = IndexedInstrProfReader::create(Path: MemoryProfileFileName, FS&: *FS); |
| 895 | if (Error E = ReaderOrErr.takeError()) { |
| 896 | handleAllErrors(E: std::move(E), Handlers: [&](const ErrorInfoBase &EI) { |
| 897 | Ctx.diagnose( |
| 898 | DI: DiagnosticInfoPGOProfile(MemoryProfileFileName.data(), EI.message())); |
| 899 | }); |
| 900 | return PreservedAnalyses::all(); |
| 901 | } |
| 902 | |
| 903 | std::unique_ptr<IndexedInstrProfReader> MemProfReader = |
| 904 | std::move(ReaderOrErr.get()); |
| 905 | if (!MemProfReader) { |
| 906 | Ctx.diagnose(DI: DiagnosticInfoPGOProfile( |
| 907 | MemoryProfileFileName.data(), StringRef("Cannot get MemProfReader" ))); |
| 908 | return PreservedAnalyses::all(); |
| 909 | } |
| 910 | |
| 911 | if (!MemProfReader->hasMemoryProfile()) { |
| 912 | Ctx.diagnose(DI: DiagnosticInfoPGOProfile(MemoryProfileFileName.data(), |
| 913 | "Not a memory profile" )); |
| 914 | return PreservedAnalyses::all(); |
| 915 | } |
| 916 | |
| 917 | const bool Changed = |
| 918 | annotateGlobalVariables(M, DataAccessProf: MemProfReader->getDataAccessProfileData()); |
| 919 | |
| 920 | // If the module doesn't contain any function, return after we process all |
| 921 | // global variables. |
| 922 | if (M.empty()) |
| 923 | return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); |
| 924 | |
| 925 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 926 | |
| 927 | TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: *M.begin()); |
| 928 | DenseMap<uint64_t, LocToLocMap> UndriftMaps; |
| 929 | if (SalvageStaleProfile) |
| 930 | UndriftMaps = computeUndriftMap(M, MemProfReader: MemProfReader.get(), TLI); |
| 931 | |
| 932 | // Map from the stack hash of each matched allocation context in the function |
| 933 | // profiles to match info such as the total profiled size (bytes), allocation |
| 934 | // type, number of frames matched to the allocation itself, and the full array |
| 935 | // of call stack ids. |
| 936 | std::map<uint64_t, AllocMatchInfo> FullStackIdToAllocMatchInfo; |
| 937 | |
| 938 | // Set of the matched call sites, each expressed as a sequence of an inline |
| 939 | // call stack. |
| 940 | std::set<std::vector<uint64_t>> MatchedCallSites; |
| 941 | |
| 942 | DenseSet<uint64_t> SeenStacks; |
| 943 | DenseSet<uint64_t> SeenFrames; |
| 944 | |
| 945 | uint64_t MaxColdSize = 0; |
| 946 | if (auto *MemProfSum = MemProfReader->getMemProfSummary()) |
| 947 | MaxColdSize = MemProfSum->getMaxColdTotalSize(); |
| 948 | |
| 949 | for (auto &F : M) { |
| 950 | if (F.isDeclaration()) |
| 951 | continue; |
| 952 | |
| 953 | const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 954 | auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
| 955 | readMemprof(M, F, MemProfReader: MemProfReader.get(), TLI, FullStackIdToAllocMatchInfo, |
| 956 | MatchedCallSites, UndriftMaps, ORE, MaxColdSize, SeenStacks, |
| 957 | SeenFrames); |
| 958 | } |
| 959 | |
| 960 | if (ClPrintMemProfMatchInfo) { |
| 961 | for (const auto &[Id, Info] : FullStackIdToAllocMatchInfo) { |
| 962 | for (auto Frames : Info.MatchedFramesSet) { |
| 963 | // TODO: To reduce verbosity, should we change the existing message |
| 964 | // so that we emit a list of matched frame counts in a single message |
| 965 | // about the context (instead of one message per frame count? |
| 966 | errs() << "MemProf " << getAllocTypeAttributeString(Type: Info.AllocType) |
| 967 | << " context with id " << Id << " has total profiled size " |
| 968 | << Info.TotalSize << " is matched with " << Frames << " frames" ; |
| 969 | if (PrintMatchedAllocStack) { |
| 970 | errs() << " and call stack" ; |
| 971 | for (auto &F : Info.CallStack) |
| 972 | errs() << " " << computeStackId(Frame: F); |
| 973 | } |
| 974 | errs() << "\n" ; |
| 975 | } |
| 976 | } |
| 977 | |
| 978 | for (const auto &CallStack : MatchedCallSites) { |
| 979 | errs() << "MemProf callsite match for inline call stack" ; |
| 980 | for (uint64_t StackId : CallStack) |
| 981 | errs() << " " << StackId; |
| 982 | errs() << "\n" ; |
| 983 | } |
| 984 | } |
| 985 | |
| 986 | return PreservedAnalyses::none(); |
| 987 | } |
| 988 | |
| 989 | bool MemProfUsePass::annotateGlobalVariables( |
| 990 | Module &M, const memprof::DataAccessProfData *DataAccessProf) { |
| 991 | if (!AnnotateStaticDataSectionPrefix || M.globals().empty()) |
| 992 | return false; |
| 993 | |
| 994 | if (!DataAccessProf) { |
| 995 | M.addModuleFlag(Behavior: Module::Warning, Key: "EnableDataAccessProf" , Val: 0U); |
| 996 | // FIXME: Add a diagnostic message without failing the compilation when |
| 997 | // data access profile payload is not available. |
| 998 | return false; |
| 999 | } |
| 1000 | M.addModuleFlag(Behavior: Module::Warning, Key: "EnableDataAccessProf" , Val: 1U); |
| 1001 | |
| 1002 | bool Changed = false; |
| 1003 | // Iterate all global variables in the module and annotate them based on |
| 1004 | // data access profiles. Note it's up to the linker to decide how to map input |
| 1005 | // sections to output sections, and one conservative practice is to map |
| 1006 | // unlikely-prefixed ones to unlikely output section, and map the rest |
| 1007 | // (hot-prefixed or prefix-less) to the canonical output section. |
| 1008 | for (GlobalVariable &GVar : M.globals()) { |
| 1009 | assert(!GVar.getSectionPrefix().has_value() && |
| 1010 | "GVar shouldn't have section prefix yet" ); |
| 1011 | auto Kind = llvm::memprof::getAnnotationKind(GV: GVar); |
| 1012 | if (Kind != llvm::memprof::AnnotationKind::AnnotationOK) { |
| 1013 | HandleUnsupportedAnnotationKinds(GVar, Kind); |
| 1014 | continue; |
| 1015 | } |
| 1016 | |
| 1017 | StringRef Name = GVar.getName(); |
| 1018 | SymbolHandleRef Handle = SymbolHandleRef(Name); |
| 1019 | // Skip string literals as their mangled names don't stay stable across |
| 1020 | // binary releases. |
| 1021 | if (!AnnotateStringLiteralSectionPrefix) |
| 1022 | if (Name.starts_with(Prefix: ".str" )) |
| 1023 | continue; |
| 1024 | |
| 1025 | if (Name.starts_with(Prefix: ".str" )) { |
| 1026 | std::optional<uint64_t> Hash = getStringContentHash(GVar); |
| 1027 | if (!Hash) { |
| 1028 | LLVM_DEBUG(dbgs() << "Cannot compute content hash for string literal " |
| 1029 | << Name << "\n" ); |
| 1030 | continue; |
| 1031 | } |
| 1032 | Handle = SymbolHandleRef(Hash.value()); |
| 1033 | } |
| 1034 | |
| 1035 | // DataAccessProfRecord's get* methods will canonicalize the name under the |
| 1036 | // hood before looking it up, so optimizer doesn't need to do it. |
| 1037 | std::optional<DataAccessProfRecord> Record = |
| 1038 | DataAccessProf->getProfileRecord(SymID: Handle); |
| 1039 | // Annotate a global variable as hot if it has non-zero sampled count, and |
| 1040 | // annotate it as cold if it's seen in the profiled binary |
| 1041 | // file but doesn't have any access sample. |
| 1042 | // For logging, optimization remark emitter requires a llvm::Function, but |
| 1043 | // it's not well defined how to associate a global variable with a function. |
| 1044 | // So we just print out the static data section prefix in LLVM_DEBUG. |
| 1045 | if (Record && Record->AccessCount > 0) { |
| 1046 | ++NumOfMemProfHotGlobalVars; |
| 1047 | Changed |= GVar.setSectionPrefix("hot" ); |
| 1048 | LLVM_DEBUG(dbgs() << "Global variable " << Name |
| 1049 | << " is annotated as hot\n" ); |
| 1050 | } else if (DataAccessProf->isKnownColdSymbol(SymID: Handle)) { |
| 1051 | ++NumOfMemProfColdGlobalVars; |
| 1052 | Changed |= GVar.setSectionPrefix("unlikely" ); |
| 1053 | Changed = true; |
| 1054 | LLVM_DEBUG(dbgs() << "Global variable " << Name |
| 1055 | << " is annotated as unlikely\n" ); |
| 1056 | } else { |
| 1057 | ++NumOfMemProfUnknownGlobalVars; |
| 1058 | LLVM_DEBUG(dbgs() << "Global variable " << Name << " is not annotated\n" ); |
| 1059 | } |
| 1060 | } |
| 1061 | |
| 1062 | return Changed; |
| 1063 | } |
| 1064 | |