| 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/SmallVector.h" |
| 16 | #include "llvm/ADT/Statistic.h" |
| 17 | #include "llvm/ADT/StringRef.h" |
| 18 | #include "llvm/Analysis/MemoryProfileInfo.h" |
| 19 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 20 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 21 | #include "llvm/IR/DiagnosticInfo.h" |
| 22 | #include "llvm/IR/Function.h" |
| 23 | #include "llvm/IR/IntrinsicInst.h" |
| 24 | #include "llvm/IR/Module.h" |
| 25 | #include "llvm/ProfileData/InstrProf.h" |
| 26 | #include "llvm/ProfileData/InstrProfReader.h" |
| 27 | #include "llvm/ProfileData/MemProfCommon.h" |
| 28 | #include "llvm/Support/BLAKE3.h" |
| 29 | #include "llvm/Support/CommandLine.h" |
| 30 | #include "llvm/Support/Debug.h" |
| 31 | #include "llvm/Support/HashBuilder.h" |
| 32 | #include "llvm/Support/VirtualFileSystem.h" |
| 33 | #include "llvm/Transforms/Utils/LongestCommonSequence.h" |
| 34 | #include <map> |
| 35 | #include <set> |
| 36 | |
| 37 | using namespace llvm; |
| 38 | using namespace llvm::memprof; |
| 39 | |
| 40 | #define DEBUG_TYPE "memprof" |
| 41 | |
| 42 | namespace llvm { |
| 43 | extern cl::opt<bool> PGOWarnMissing; |
| 44 | extern cl::opt<bool> NoPGOWarnMismatch; |
| 45 | extern cl::opt<bool> NoPGOWarnMismatchComdatWeak; |
| 46 | } // namespace llvm |
| 47 | |
| 48 | // By default disable matching of allocation profiles onto operator new that |
| 49 | // already explicitly pass a hot/cold hint, since we don't currently |
| 50 | // override these hints anyway. |
| 51 | static cl::opt<bool> ClMemProfMatchHotColdNew( |
| 52 | "memprof-match-hot-cold-new" , |
| 53 | cl::desc( |
| 54 | "Match allocation profiles onto existing hot/cold operator new calls" ), |
| 55 | cl::Hidden, cl::init(Val: false)); |
| 56 | |
| 57 | static cl::opt<bool> |
| 58 | ClPrintMemProfMatchInfo("memprof-print-match-info" , |
| 59 | cl::desc("Print matching stats for each allocation " |
| 60 | "context in this module's profiles" ), |
| 61 | cl::Hidden, cl::init(Val: false)); |
| 62 | |
| 63 | static cl::opt<bool> |
| 64 | SalvageStaleProfile("memprof-salvage-stale-profile" , |
| 65 | cl::desc("Salvage stale MemProf profile" ), |
| 66 | cl::init(Val: false), cl::Hidden); |
| 67 | |
| 68 | static cl::opt<bool> ClMemProfAttachCalleeGuids( |
| 69 | "memprof-attach-calleeguids" , |
| 70 | cl::desc( |
| 71 | "Attach calleeguids as value profile metadata for indirect calls." ), |
| 72 | cl::init(Val: true), cl::Hidden); |
| 73 | |
| 74 | static cl::opt<unsigned> MinMatchedColdBytePercent( |
| 75 | "memprof-matching-cold-threshold" , cl::init(Val: 100), cl::Hidden, |
| 76 | cl::desc("Min percent of cold bytes matched to hint allocation cold" )); |
| 77 | |
| 78 | // Matching statistics |
| 79 | STATISTIC(NumOfMemProfMissing, "Number of functions without memory profile." ); |
| 80 | STATISTIC(NumOfMemProfMismatch, |
| 81 | "Number of functions having mismatched memory profile hash." ); |
| 82 | STATISTIC(NumOfMemProfFunc, "Number of functions having valid memory profile." ); |
| 83 | STATISTIC(NumOfMemProfAllocContextProfiles, |
| 84 | "Number of alloc contexts in memory profile." ); |
| 85 | STATISTIC(NumOfMemProfCallSiteProfiles, |
| 86 | "Number of callsites in memory profile." ); |
| 87 | STATISTIC(NumOfMemProfMatchedAllocContexts, |
| 88 | "Number of matched memory profile alloc contexts." ); |
| 89 | STATISTIC(NumOfMemProfMatchedAllocs, |
| 90 | "Number of matched memory profile allocs." ); |
| 91 | STATISTIC(NumOfMemProfMatchedCallSites, |
| 92 | "Number of matched memory profile callsites." ); |
| 93 | |
| 94 | static void addCallsiteMetadata(Instruction &I, |
| 95 | ArrayRef<uint64_t> InlinedCallStack, |
| 96 | LLVMContext &Ctx) { |
| 97 | I.setMetadata(KindID: LLVMContext::MD_callsite, |
| 98 | Node: buildCallstackMetadata(CallStack: InlinedCallStack, Ctx)); |
| 99 | } |
| 100 | |
| 101 | static uint64_t computeStackId(GlobalValue::GUID Function, uint32_t LineOffset, |
| 102 | uint32_t Column) { |
| 103 | llvm::HashBuilder<llvm::TruncatedBLAKE3<8>, llvm::endianness::little> |
| 104 | HashBuilder; |
| 105 | HashBuilder.add(Args: Function, Args: LineOffset, Args: Column); |
| 106 | llvm::BLAKE3Result<8> Hash = HashBuilder.final(); |
| 107 | uint64_t Id; |
| 108 | std::memcpy(dest: &Id, src: Hash.data(), n: sizeof(Hash)); |
| 109 | return Id; |
| 110 | } |
| 111 | |
| 112 | static uint64_t computeStackId(const memprof::Frame &Frame) { |
| 113 | return computeStackId(Function: Frame.Function, LineOffset: Frame.LineOffset, Column: Frame.Column); |
| 114 | } |
| 115 | |
| 116 | static AllocationType addCallStack(CallStackTrie &AllocTrie, |
| 117 | const AllocationInfo *AllocInfo, |
| 118 | uint64_t FullStackId) { |
| 119 | SmallVector<uint64_t> StackIds; |
| 120 | for (const auto &StackFrame : AllocInfo->CallStack) |
| 121 | StackIds.push_back(Elt: computeStackId(Frame: StackFrame)); |
| 122 | auto AllocType = getAllocType(TotalLifetimeAccessDensity: AllocInfo->Info.getTotalLifetimeAccessDensity(), |
| 123 | AllocCount: AllocInfo->Info.getAllocCount(), |
| 124 | TotalLifetime: AllocInfo->Info.getTotalLifetime()); |
| 125 | std::vector<ContextTotalSize> ContextSizeInfo; |
| 126 | if (recordContextSizeInfoForAnalysis()) { |
| 127 | auto TotalSize = AllocInfo->Info.getTotalSize(); |
| 128 | assert(TotalSize); |
| 129 | assert(FullStackId != 0); |
| 130 | ContextSizeInfo.push_back(x: {.FullStackId: FullStackId, .TotalSize: TotalSize}); |
| 131 | } |
| 132 | AllocTrie.addCallStack(AllocType, StackIds, ContextSizeInfo: std::move(ContextSizeInfo)); |
| 133 | return AllocType; |
| 134 | } |
| 135 | |
| 136 | // Return true if InlinedCallStack, computed from a call instruction's debug |
| 137 | // info, is a prefix of ProfileCallStack, a list of Frames from profile data |
| 138 | // (either the allocation data or a callsite). |
| 139 | static bool |
| 140 | stackFrameIncludesInlinedCallStack(ArrayRef<Frame> ProfileCallStack, |
| 141 | ArrayRef<uint64_t> InlinedCallStack) { |
| 142 | return ProfileCallStack.size() >= InlinedCallStack.size() && |
| 143 | llvm::equal(LRange: ProfileCallStack.take_front(N: InlinedCallStack.size()), |
| 144 | RRange&: InlinedCallStack, P: [](const Frame &F, uint64_t StackId) { |
| 145 | return computeStackId(Frame: F) == StackId; |
| 146 | }); |
| 147 | } |
| 148 | |
| 149 | static bool isAllocationWithHotColdVariant(const Function *Callee, |
| 150 | const TargetLibraryInfo &TLI) { |
| 151 | if (!Callee) |
| 152 | return false; |
| 153 | LibFunc Func; |
| 154 | if (!TLI.getLibFunc(FDecl: *Callee, F&: Func)) |
| 155 | return false; |
| 156 | switch (Func) { |
| 157 | case LibFunc_Znwm: |
| 158 | case LibFunc_ZnwmRKSt9nothrow_t: |
| 159 | case LibFunc_ZnwmSt11align_val_t: |
| 160 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t: |
| 161 | case LibFunc_Znam: |
| 162 | case LibFunc_ZnamRKSt9nothrow_t: |
| 163 | case LibFunc_ZnamSt11align_val_t: |
| 164 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t: |
| 165 | case LibFunc_size_returning_new: |
| 166 | case LibFunc_size_returning_new_aligned: |
| 167 | return true; |
| 168 | case LibFunc_Znwm12__hot_cold_t: |
| 169 | case LibFunc_ZnwmRKSt9nothrow_t12__hot_cold_t: |
| 170 | case LibFunc_ZnwmSt11align_val_t12__hot_cold_t: |
| 171 | case LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 172 | case LibFunc_Znam12__hot_cold_t: |
| 173 | case LibFunc_ZnamRKSt9nothrow_t12__hot_cold_t: |
| 174 | case LibFunc_ZnamSt11align_val_t12__hot_cold_t: |
| 175 | case LibFunc_ZnamSt11align_val_tRKSt9nothrow_t12__hot_cold_t: |
| 176 | case LibFunc_size_returning_new_hot_cold: |
| 177 | case LibFunc_size_returning_new_aligned_hot_cold: |
| 178 | return ClMemProfMatchHotColdNew; |
| 179 | default: |
| 180 | return false; |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | struct AllocMatchInfo { |
| 185 | uint64_t TotalSize = 0; |
| 186 | AllocationType AllocType = AllocationType::None; |
| 187 | }; |
| 188 | |
| 189 | DenseMap<uint64_t, SmallVector<CallEdgeTy, 0>> |
| 190 | memprof::(Module &M, const TargetLibraryInfo &TLI, |
| 191 | function_ref<bool(uint64_t)> IsPresentInProfile) { |
| 192 | DenseMap<uint64_t, SmallVector<CallEdgeTy, 0>> Calls; |
| 193 | |
| 194 | auto GetOffset = [](const DILocation *DIL) { |
| 195 | return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) & |
| 196 | 0xffff; |
| 197 | }; |
| 198 | |
| 199 | for (Function &F : M) { |
| 200 | if (F.isDeclaration()) |
| 201 | continue; |
| 202 | |
| 203 | for (auto &BB : F) { |
| 204 | for (auto &I : BB) { |
| 205 | if (!isa<CallBase>(Val: &I) || isa<IntrinsicInst>(Val: &I)) |
| 206 | continue; |
| 207 | |
| 208 | auto *CB = dyn_cast<CallBase>(Val: &I); |
| 209 | auto *CalledFunction = CB->getCalledFunction(); |
| 210 | // Disregard indirect calls and intrinsics. |
| 211 | if (!CalledFunction || CalledFunction->isIntrinsic()) |
| 212 | continue; |
| 213 | |
| 214 | StringRef CalleeName = CalledFunction->getName(); |
| 215 | // True if we are calling a heap allocation function that supports |
| 216 | // hot/cold variants. |
| 217 | bool IsAlloc = isAllocationWithHotColdVariant(Callee: CalledFunction, TLI); |
| 218 | // True for the first iteration below, indicating that we are looking at |
| 219 | // a leaf node. |
| 220 | bool IsLeaf = true; |
| 221 | for (const DILocation *DIL = I.getDebugLoc(); DIL; |
| 222 | DIL = DIL->getInlinedAt()) { |
| 223 | StringRef CallerName = DIL->getSubprogramLinkageName(); |
| 224 | assert(!CallerName.empty() && |
| 225 | "Be sure to enable -fdebug-info-for-profiling" ); |
| 226 | uint64_t CallerGUID = memprof::getGUID(FunctionName: CallerName); |
| 227 | uint64_t CalleeGUID = memprof::getGUID(FunctionName: CalleeName); |
| 228 | // Pretend that we are calling a function with GUID == 0 if we are |
| 229 | // in the inline stack leading to a heap allocation function. |
| 230 | if (IsAlloc) { |
| 231 | if (IsLeaf) { |
| 232 | // For leaf nodes, set CalleeGUID to 0 without consulting |
| 233 | // IsPresentInProfile. |
| 234 | CalleeGUID = 0; |
| 235 | } else if (!IsPresentInProfile(CalleeGUID)) { |
| 236 | // In addition to the leaf case above, continue to set CalleeGUID |
| 237 | // to 0 as long as we don't see CalleeGUID in the profile. |
| 238 | CalleeGUID = 0; |
| 239 | } else { |
| 240 | // Once we encounter a callee that exists in the profile, stop |
| 241 | // setting CalleeGUID to 0. |
| 242 | IsAlloc = false; |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | LineLocation Loc = {GetOffset(DIL), DIL->getColumn()}; |
| 247 | Calls[CallerGUID].emplace_back(Args&: Loc, Args&: CalleeGUID); |
| 248 | CalleeName = CallerName; |
| 249 | IsLeaf = false; |
| 250 | } |
| 251 | } |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | // Sort each call list by the source location. |
| 256 | for (auto &[CallerGUID, CallList] : Calls) { |
| 257 | llvm::sort(C&: CallList); |
| 258 | CallList.erase(CS: llvm::unique(R&: CallList), CE: CallList.end()); |
| 259 | } |
| 260 | |
| 261 | return Calls; |
| 262 | } |
| 263 | |
| 264 | DenseMap<uint64_t, LocToLocMap> |
| 265 | memprof::computeUndriftMap(Module &M, IndexedInstrProfReader *MemProfReader, |
| 266 | const TargetLibraryInfo &TLI) { |
| 267 | DenseMap<uint64_t, LocToLocMap> UndriftMaps; |
| 268 | |
| 269 | DenseMap<uint64_t, SmallVector<memprof::CallEdgeTy, 0>> CallsFromProfile = |
| 270 | MemProfReader->getMemProfCallerCalleePairs(); |
| 271 | DenseMap<uint64_t, SmallVector<memprof::CallEdgeTy, 0>> CallsFromIR = |
| 272 | extractCallsFromIR(M, TLI, IsPresentInProfile: [&](uint64_t GUID) { |
| 273 | return CallsFromProfile.contains(Val: GUID); |
| 274 | }); |
| 275 | |
| 276 | // Compute an undrift map for each CallerGUID. |
| 277 | for (const auto &[CallerGUID, IRAnchors] : CallsFromIR) { |
| 278 | auto It = CallsFromProfile.find(Val: CallerGUID); |
| 279 | if (It == CallsFromProfile.end()) |
| 280 | continue; |
| 281 | const auto &ProfileAnchors = It->second; |
| 282 | |
| 283 | LocToLocMap Matchings; |
| 284 | longestCommonSequence<LineLocation, GlobalValue::GUID>( |
| 285 | AnchorList1: ProfileAnchors, AnchorList2: IRAnchors, FunctionMatchesProfile: std::equal_to<GlobalValue::GUID>(), |
| 286 | InsertMatching: [&](LineLocation A, LineLocation B) { Matchings.try_emplace(k: A, args&: B); }); |
| 287 | [[maybe_unused]] bool Inserted = |
| 288 | UndriftMaps.try_emplace(Key: CallerGUID, Args: std::move(Matchings)).second; |
| 289 | |
| 290 | // The insertion must succeed because we visit each GUID exactly once. |
| 291 | assert(Inserted); |
| 292 | } |
| 293 | |
| 294 | return UndriftMaps; |
| 295 | } |
| 296 | |
| 297 | // Given a MemProfRecord, undrift all the source locations present in the |
| 298 | // record in place. |
| 299 | static void |
| 300 | undriftMemProfRecord(const DenseMap<uint64_t, LocToLocMap> &UndriftMaps, |
| 301 | memprof::MemProfRecord &MemProfRec) { |
| 302 | // Undrift a call stack in place. |
| 303 | auto UndriftCallStack = [&](std::vector<Frame> &CallStack) { |
| 304 | for (auto &F : CallStack) { |
| 305 | auto I = UndriftMaps.find(Val: F.Function); |
| 306 | if (I == UndriftMaps.end()) |
| 307 | continue; |
| 308 | auto J = I->second.find(x: LineLocation(F.LineOffset, F.Column)); |
| 309 | if (J == I->second.end()) |
| 310 | continue; |
| 311 | auto &NewLoc = J->second; |
| 312 | F.LineOffset = NewLoc.LineOffset; |
| 313 | F.Column = NewLoc.Column; |
| 314 | } |
| 315 | }; |
| 316 | |
| 317 | for (auto &AS : MemProfRec.AllocSites) |
| 318 | UndriftCallStack(AS.CallStack); |
| 319 | |
| 320 | for (auto &CS : MemProfRec.CallSites) |
| 321 | UndriftCallStack(CS.Frames); |
| 322 | } |
| 323 | |
| 324 | // Helper function to process CalleeGuids and create value profile metadata |
| 325 | static void addVPMetadata(Module &M, Instruction &I, |
| 326 | ArrayRef<GlobalValue::GUID> CalleeGuids) { |
| 327 | if (!ClMemProfAttachCalleeGuids || CalleeGuids.empty()) |
| 328 | return; |
| 329 | |
| 330 | if (I.getMetadata(KindID: LLVMContext::MD_prof)) { |
| 331 | uint64_t Unused; |
| 332 | // TODO: When merging is implemented, increase this to a typical ICP value |
| 333 | // (e.g., 3-6) For now, we only need to check if existing data exists, so 1 |
| 334 | // is sufficient |
| 335 | auto ExistingVD = getValueProfDataFromInst(Inst: I, ValueKind: IPVK_IndirectCallTarget, |
| 336 | /*MaxNumValueData=*/1, TotalC&: Unused); |
| 337 | // We don't know how to merge value profile data yet. |
| 338 | if (!ExistingVD.empty()) { |
| 339 | return; |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | SmallVector<InstrProfValueData, 4> VDs; |
| 344 | uint64_t TotalCount = 0; |
| 345 | |
| 346 | for (const GlobalValue::GUID CalleeGUID : CalleeGuids) { |
| 347 | InstrProfValueData VD; |
| 348 | VD.Value = CalleeGUID; |
| 349 | // For MemProf, we don't have actual call counts, so we assign |
| 350 | // a weight of 1 to each potential target. |
| 351 | // TODO: Consider making this weight configurable or increasing it to |
| 352 | // improve effectiveness for ICP. |
| 353 | VD.Count = 1; |
| 354 | VDs.push_back(Elt: VD); |
| 355 | TotalCount += VD.Count; |
| 356 | } |
| 357 | |
| 358 | if (!VDs.empty()) { |
| 359 | annotateValueSite(M, Inst&: I, VDs, Sum: TotalCount, ValueKind: IPVK_IndirectCallTarget, |
| 360 | MaxMDCount: VDs.size()); |
| 361 | } |
| 362 | } |
| 363 | |
| 364 | static void readMemprof(Module &M, Function &F, |
| 365 | IndexedInstrProfReader *MemProfReader, |
| 366 | const TargetLibraryInfo &TLI, |
| 367 | std::map<std::pair<uint64_t, unsigned>, AllocMatchInfo> |
| 368 | &FullStackIdToAllocMatchInfo, |
| 369 | std::set<std::vector<uint64_t>> &MatchedCallSites, |
| 370 | DenseMap<uint64_t, LocToLocMap> &UndriftMaps, |
| 371 | OptimizationRemarkEmitter &ORE, uint64_t MaxColdSize) { |
| 372 | auto &Ctx = M.getContext(); |
| 373 | // Previously we used getIRPGOFuncName() here. If F is local linkage, |
| 374 | // getIRPGOFuncName() returns FuncName with prefix 'FileName;'. But |
| 375 | // llvm-profdata uses FuncName in dwarf to create GUID which doesn't |
| 376 | // contain FileName's prefix. It caused local linkage function can't |
| 377 | // find MemProfRecord. So we use getName() now. |
| 378 | // 'unique-internal-linkage-names' can make MemProf work better for local |
| 379 | // linkage function. |
| 380 | auto FuncName = F.getName(); |
| 381 | auto FuncGUID = Function::getGUIDAssumingExternalLinkage(GlobalName: FuncName); |
| 382 | std::optional<memprof::MemProfRecord> MemProfRec; |
| 383 | auto Err = MemProfReader->getMemProfRecord(FuncNameHash: FuncGUID).moveInto(Value&: MemProfRec); |
| 384 | if (Err) { |
| 385 | handleAllErrors(E: std::move(Err), Handlers: [&](const InstrProfError &IPE) { |
| 386 | auto Err = IPE.get(); |
| 387 | bool SkipWarning = false; |
| 388 | LLVM_DEBUG(dbgs() << "Error in reading profile for Func " << FuncName |
| 389 | << ": " ); |
| 390 | if (Err == instrprof_error::unknown_function) { |
| 391 | NumOfMemProfMissing++; |
| 392 | SkipWarning = !PGOWarnMissing; |
| 393 | LLVM_DEBUG(dbgs() << "unknown function" ); |
| 394 | } else if (Err == instrprof_error::hash_mismatch) { |
| 395 | NumOfMemProfMismatch++; |
| 396 | SkipWarning = |
| 397 | NoPGOWarnMismatch || |
| 398 | (NoPGOWarnMismatchComdatWeak && |
| 399 | (F.hasComdat() || |
| 400 | F.getLinkage() == GlobalValue::AvailableExternallyLinkage)); |
| 401 | LLVM_DEBUG(dbgs() << "hash mismatch (skip=" << SkipWarning << ")" ); |
| 402 | } |
| 403 | |
| 404 | if (SkipWarning) |
| 405 | return; |
| 406 | |
| 407 | std::string Msg = (IPE.message() + Twine(" " ) + F.getName().str() + |
| 408 | Twine(" Hash = " ) + std::to_string(val: FuncGUID)) |
| 409 | .str(); |
| 410 | |
| 411 | Ctx.diagnose( |
| 412 | DI: DiagnosticInfoPGOProfile(M.getName().data(), Msg, DS_Warning)); |
| 413 | }); |
| 414 | return; |
| 415 | } |
| 416 | |
| 417 | NumOfMemProfFunc++; |
| 418 | |
| 419 | // If requested, undrfit MemProfRecord so that the source locations in it |
| 420 | // match those in the IR. |
| 421 | if (SalvageStaleProfile) |
| 422 | undriftMemProfRecord(UndriftMaps, MemProfRec&: *MemProfRec); |
| 423 | |
| 424 | // Detect if there are non-zero column numbers in the profile. If not, |
| 425 | // treat all column numbers as 0 when matching (i.e. ignore any non-zero |
| 426 | // columns in the IR). The profiled binary might have been built with |
| 427 | // column numbers disabled, for example. |
| 428 | bool ProfileHasColumns = false; |
| 429 | |
| 430 | // Build maps of the location hash to all profile data with that leaf location |
| 431 | // (allocation info and the callsites). |
| 432 | std::map<uint64_t, std::set<const AllocationInfo *>> LocHashToAllocInfo; |
| 433 | |
| 434 | // Helper struct for maintaining refs to callsite data. As an alternative we |
| 435 | // could store a pointer to the CallSiteInfo struct but we also need the frame |
| 436 | // index. Using ArrayRefs instead makes it a little easier to read. |
| 437 | struct CallSiteEntry { |
| 438 | // Subset of frames for the corresponding CallSiteInfo. |
| 439 | ArrayRef<Frame> Frames; |
| 440 | // Potential targets for indirect calls. |
| 441 | ArrayRef<GlobalValue::GUID> CalleeGuids; |
| 442 | |
| 443 | // Only compare Frame contents. |
| 444 | // Use pointer-based equality instead of ArrayRef's operator== which does |
| 445 | // element-wise comparison. We want to check if it's the same slice of the |
| 446 | // underlying array, not just equivalent content. |
| 447 | bool operator==(const CallSiteEntry &Other) const { |
| 448 | return Frames.data() == Other.Frames.data() && |
| 449 | Frames.size() == Other.Frames.size(); |
| 450 | } |
| 451 | }; |
| 452 | |
| 453 | struct CallSiteEntryHash { |
| 454 | size_t operator()(const CallSiteEntry &Entry) const { |
| 455 | return computeFullStackId(CallStack: Entry.Frames); |
| 456 | } |
| 457 | }; |
| 458 | |
| 459 | // For the callsites we need to record slices of the frame array (see comments |
| 460 | // below where the map entries are added) along with their CalleeGuids. |
| 461 | std::map<uint64_t, std::unordered_set<CallSiteEntry, CallSiteEntryHash>> |
| 462 | LocHashToCallSites; |
| 463 | for (auto &AI : MemProfRec->AllocSites) { |
| 464 | NumOfMemProfAllocContextProfiles++; |
| 465 | // Associate the allocation info with the leaf frame. The later matching |
| 466 | // code will match any inlined call sequences in the IR with a longer prefix |
| 467 | // of call stack frames. |
| 468 | uint64_t StackId = computeStackId(Frame: AI.CallStack[0]); |
| 469 | LocHashToAllocInfo[StackId].insert(x: &AI); |
| 470 | ProfileHasColumns |= AI.CallStack[0].Column; |
| 471 | } |
| 472 | for (auto &CS : MemProfRec->CallSites) { |
| 473 | NumOfMemProfCallSiteProfiles++; |
| 474 | // Need to record all frames from leaf up to and including this function, |
| 475 | // as any of these may or may not have been inlined at this point. |
| 476 | unsigned Idx = 0; |
| 477 | for (auto &StackFrame : CS.Frames) { |
| 478 | uint64_t StackId = computeStackId(Frame: StackFrame); |
| 479 | ArrayRef<Frame> FrameSlice = ArrayRef<Frame>(CS.Frames).drop_front(N: Idx++); |
| 480 | ArrayRef<GlobalValue::GUID> CalleeGuids(CS.CalleeGuids); |
| 481 | LocHashToCallSites[StackId].insert(x: {.Frames: FrameSlice, .CalleeGuids: CalleeGuids}); |
| 482 | |
| 483 | ProfileHasColumns |= StackFrame.Column; |
| 484 | // Once we find this function, we can stop recording. |
| 485 | if (StackFrame.Function == FuncGUID) |
| 486 | break; |
| 487 | } |
| 488 | assert(Idx <= CS.Frames.size() && CS.Frames[Idx - 1].Function == FuncGUID); |
| 489 | } |
| 490 | |
| 491 | auto GetOffset = [](const DILocation *DIL) { |
| 492 | return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) & |
| 493 | 0xffff; |
| 494 | }; |
| 495 | |
| 496 | // Now walk the instructions, looking up the associated profile data using |
| 497 | // debug locations. |
| 498 | for (auto &BB : F) { |
| 499 | for (auto &I : BB) { |
| 500 | if (I.isDebugOrPseudoInst()) |
| 501 | continue; |
| 502 | // We are only interested in calls (allocation or interior call stack |
| 503 | // context calls). |
| 504 | auto *CI = dyn_cast<CallBase>(Val: &I); |
| 505 | if (!CI) |
| 506 | continue; |
| 507 | auto *CalledFunction = CI->getCalledFunction(); |
| 508 | if (CalledFunction && CalledFunction->isIntrinsic()) |
| 509 | continue; |
| 510 | // List of call stack ids computed from the location hashes on debug |
| 511 | // locations (leaf to inlined at root). |
| 512 | SmallVector<uint64_t, 8> InlinedCallStack; |
| 513 | // Was the leaf location found in one of the profile maps? |
| 514 | bool LeafFound = false; |
| 515 | // If leaf was found in a map, iterators pointing to its location in both |
| 516 | // of the maps. It might exist in neither, one, or both (the latter case |
| 517 | // can happen because we don't currently have discriminators to |
| 518 | // distinguish the case when a single line/col maps to both an allocation |
| 519 | // and another callsite). |
| 520 | auto AllocInfoIter = LocHashToAllocInfo.end(); |
| 521 | auto CallSitesIter = LocHashToCallSites.end(); |
| 522 | for (const DILocation *DIL = I.getDebugLoc(); DIL != nullptr; |
| 523 | DIL = DIL->getInlinedAt()) { |
| 524 | // Use C++ linkage name if possible. Need to compile with |
| 525 | // -fdebug-info-for-profiling to get linkage name. |
| 526 | StringRef Name = DIL->getScope()->getSubprogram()->getLinkageName(); |
| 527 | if (Name.empty()) |
| 528 | Name = DIL->getScope()->getSubprogram()->getName(); |
| 529 | auto CalleeGUID = Function::getGUIDAssumingExternalLinkage(GlobalName: Name); |
| 530 | auto StackId = computeStackId(Function: CalleeGUID, LineOffset: GetOffset(DIL), |
| 531 | Column: ProfileHasColumns ? DIL->getColumn() : 0); |
| 532 | // Check if we have found the profile's leaf frame. If yes, collect |
| 533 | // the rest of the call's inlined context starting here. If not, see if |
| 534 | // we find a match further up the inlined context (in case the profile |
| 535 | // was missing debug frames at the leaf). |
| 536 | if (!LeafFound) { |
| 537 | AllocInfoIter = LocHashToAllocInfo.find(x: StackId); |
| 538 | CallSitesIter = LocHashToCallSites.find(x: StackId); |
| 539 | if (AllocInfoIter != LocHashToAllocInfo.end() || |
| 540 | CallSitesIter != LocHashToCallSites.end()) |
| 541 | LeafFound = true; |
| 542 | } |
| 543 | if (LeafFound) |
| 544 | InlinedCallStack.push_back(Elt: StackId); |
| 545 | } |
| 546 | // If leaf not in either of the maps, skip inst. |
| 547 | if (!LeafFound) |
| 548 | continue; |
| 549 | |
| 550 | // First add !memprof metadata from allocation info, if we found the |
| 551 | // instruction's leaf location in that map, and if the rest of the |
| 552 | // instruction's locations match the prefix Frame locations on an |
| 553 | // allocation context with the same leaf. |
| 554 | if (AllocInfoIter != LocHashToAllocInfo.end() && |
| 555 | // Only consider allocations which support hinting. |
| 556 | isAllocationWithHotColdVariant(Callee: CI->getCalledFunction(), TLI)) { |
| 557 | // We may match this instruction's location list to multiple MIB |
| 558 | // contexts. Add them to a Trie specialized for trimming the contexts to |
| 559 | // the minimal needed to disambiguate contexts with unique behavior. |
| 560 | CallStackTrie AllocTrie(&ORE, MaxColdSize); |
| 561 | uint64_t TotalSize = 0; |
| 562 | uint64_t TotalColdSize = 0; |
| 563 | for (auto *AllocInfo : AllocInfoIter->second) { |
| 564 | // Check the full inlined call stack against this one. |
| 565 | // If we found and thus matched all frames on the call, include |
| 566 | // this MIB. |
| 567 | if (stackFrameIncludesInlinedCallStack(ProfileCallStack: AllocInfo->CallStack, |
| 568 | InlinedCallStack)) { |
| 569 | NumOfMemProfMatchedAllocContexts++; |
| 570 | uint64_t FullStackId = 0; |
| 571 | if (ClPrintMemProfMatchInfo || recordContextSizeInfoForAnalysis()) |
| 572 | FullStackId = computeFullStackId(CallStack: AllocInfo->CallStack); |
| 573 | auto AllocType = addCallStack(AllocTrie, AllocInfo, FullStackId); |
| 574 | TotalSize += AllocInfo->Info.getTotalSize(); |
| 575 | if (AllocType == AllocationType::Cold) |
| 576 | TotalColdSize += AllocInfo->Info.getTotalSize(); |
| 577 | // Record information about the allocation if match info printing |
| 578 | // was requested. |
| 579 | if (ClPrintMemProfMatchInfo) { |
| 580 | assert(FullStackId != 0); |
| 581 | FullStackIdToAllocMatchInfo[std::make_pair( |
| 582 | x&: FullStackId, y: InlinedCallStack.size())] = { |
| 583 | .TotalSize: AllocInfo->Info.getTotalSize(), .AllocType: AllocType}; |
| 584 | } |
| 585 | } |
| 586 | } |
| 587 | // If the threshold for the percent of cold bytes is less than 100%, |
| 588 | // and not all bytes are cold, see if we should still hint this |
| 589 | // allocation as cold without context sensitivity. |
| 590 | if (TotalColdSize < TotalSize && MinMatchedColdBytePercent < 100 && |
| 591 | TotalColdSize * 100 >= MinMatchedColdBytePercent * TotalSize) { |
| 592 | AllocTrie.addSingleAllocTypeAttribute(CI, AT: AllocationType::Cold, |
| 593 | Descriptor: "dominant" ); |
| 594 | continue; |
| 595 | } |
| 596 | |
| 597 | // We might not have matched any to the full inlined call stack. |
| 598 | // But if we did, create and attach metadata, or a function attribute if |
| 599 | // all contexts have identical profiled behavior. |
| 600 | if (!AllocTrie.empty()) { |
| 601 | NumOfMemProfMatchedAllocs++; |
| 602 | // MemprofMDAttached will be false if a function attribute was |
| 603 | // attached. |
| 604 | bool MemprofMDAttached = AllocTrie.buildAndAttachMIBMetadata(CI); |
| 605 | assert(MemprofMDAttached == I.hasMetadata(LLVMContext::MD_memprof)); |
| 606 | if (MemprofMDAttached) { |
| 607 | // Add callsite metadata for the instruction's location list so that |
| 608 | // it simpler later on to identify which part of the MIB contexts |
| 609 | // are from this particular instruction (including during inlining, |
| 610 | // when the callsite metadata will be updated appropriately). |
| 611 | // FIXME: can this be changed to strip out the matching stack |
| 612 | // context ids from the MIB contexts and not add any callsite |
| 613 | // metadata here to save space? |
| 614 | addCallsiteMetadata(I, InlinedCallStack, Ctx); |
| 615 | } |
| 616 | } |
| 617 | continue; |
| 618 | } |
| 619 | |
| 620 | if (CallSitesIter == LocHashToCallSites.end()) |
| 621 | continue; |
| 622 | |
| 623 | // Otherwise, add callsite metadata. If we reach here then we found the |
| 624 | // instruction's leaf location in the callsites map and not the allocation |
| 625 | // map. |
| 626 | for (const auto &CallSiteEntry : CallSitesIter->second) { |
| 627 | // If we found and thus matched all frames on the call, create and |
| 628 | // attach call stack metadata. |
| 629 | if (stackFrameIncludesInlinedCallStack(ProfileCallStack: CallSiteEntry.Frames, |
| 630 | InlinedCallStack)) { |
| 631 | NumOfMemProfMatchedCallSites++; |
| 632 | addCallsiteMetadata(I, InlinedCallStack, Ctx); |
| 633 | |
| 634 | // Try to attach indirect call metadata if possible. |
| 635 | if (!CalledFunction) |
| 636 | addVPMetadata(M, I, CalleeGuids: CallSiteEntry.CalleeGuids); |
| 637 | |
| 638 | // Only need to find one with a matching call stack and add a single |
| 639 | // callsite metadata. |
| 640 | |
| 641 | // Accumulate call site matching information upon request. |
| 642 | if (ClPrintMemProfMatchInfo) { |
| 643 | std::vector<uint64_t> CallStack; |
| 644 | append_range(C&: CallStack, R&: InlinedCallStack); |
| 645 | MatchedCallSites.insert(x: std::move(CallStack)); |
| 646 | } |
| 647 | break; |
| 648 | } |
| 649 | } |
| 650 | } |
| 651 | } |
| 652 | } |
| 653 | |
| 654 | MemProfUsePass::MemProfUsePass(std::string MemoryProfileFile, |
| 655 | IntrusiveRefCntPtr<vfs::FileSystem> FS) |
| 656 | : MemoryProfileFileName(MemoryProfileFile), FS(FS) { |
| 657 | if (!FS) |
| 658 | this->FS = vfs::getRealFileSystem(); |
| 659 | } |
| 660 | |
| 661 | PreservedAnalyses MemProfUsePass::run(Module &M, ModuleAnalysisManager &AM) { |
| 662 | // Return immediately if the module doesn't contain any function. |
| 663 | if (M.empty()) |
| 664 | return PreservedAnalyses::all(); |
| 665 | |
| 666 | LLVM_DEBUG(dbgs() << "Read in memory profile:" ); |
| 667 | auto &Ctx = M.getContext(); |
| 668 | auto ReaderOrErr = IndexedInstrProfReader::create(Path: MemoryProfileFileName, FS&: *FS); |
| 669 | if (Error E = ReaderOrErr.takeError()) { |
| 670 | handleAllErrors(E: std::move(E), Handlers: [&](const ErrorInfoBase &EI) { |
| 671 | Ctx.diagnose( |
| 672 | DI: DiagnosticInfoPGOProfile(MemoryProfileFileName.data(), EI.message())); |
| 673 | }); |
| 674 | return PreservedAnalyses::all(); |
| 675 | } |
| 676 | |
| 677 | std::unique_ptr<IndexedInstrProfReader> MemProfReader = |
| 678 | std::move(ReaderOrErr.get()); |
| 679 | if (!MemProfReader) { |
| 680 | Ctx.diagnose(DI: DiagnosticInfoPGOProfile( |
| 681 | MemoryProfileFileName.data(), StringRef("Cannot get MemProfReader" ))); |
| 682 | return PreservedAnalyses::all(); |
| 683 | } |
| 684 | |
| 685 | if (!MemProfReader->hasMemoryProfile()) { |
| 686 | Ctx.diagnose(DI: DiagnosticInfoPGOProfile(MemoryProfileFileName.data(), |
| 687 | "Not a memory profile" )); |
| 688 | return PreservedAnalyses::all(); |
| 689 | } |
| 690 | |
| 691 | auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 692 | |
| 693 | TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: *M.begin()); |
| 694 | DenseMap<uint64_t, LocToLocMap> UndriftMaps; |
| 695 | if (SalvageStaleProfile) |
| 696 | UndriftMaps = computeUndriftMap(M, MemProfReader: MemProfReader.get(), TLI); |
| 697 | |
| 698 | // Map from the stack hash and matched frame count of each allocation context |
| 699 | // in the function profiles to the total profiled size (bytes) and allocation |
| 700 | // type. |
| 701 | std::map<std::pair<uint64_t, unsigned>, AllocMatchInfo> |
| 702 | FullStackIdToAllocMatchInfo; |
| 703 | |
| 704 | // Set of the matched call sites, each expressed as a sequence of an inline |
| 705 | // call stack. |
| 706 | std::set<std::vector<uint64_t>> MatchedCallSites; |
| 707 | |
| 708 | uint64_t MaxColdSize = 0; |
| 709 | if (auto *MemProfSum = MemProfReader->getMemProfSummary()) |
| 710 | MaxColdSize = MemProfSum->getMaxColdTotalSize(); |
| 711 | |
| 712 | for (auto &F : M) { |
| 713 | if (F.isDeclaration()) |
| 714 | continue; |
| 715 | |
| 716 | const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 717 | auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
| 718 | readMemprof(M, F, MemProfReader: MemProfReader.get(), TLI, FullStackIdToAllocMatchInfo, |
| 719 | MatchedCallSites, UndriftMaps, ORE, MaxColdSize); |
| 720 | } |
| 721 | |
| 722 | if (ClPrintMemProfMatchInfo) { |
| 723 | for (const auto &[IdLengthPair, Info] : FullStackIdToAllocMatchInfo) { |
| 724 | auto [Id, Length] = IdLengthPair; |
| 725 | errs() << "MemProf " << getAllocTypeAttributeString(Type: Info.AllocType) |
| 726 | << " context with id " << Id << " has total profiled size " |
| 727 | << Info.TotalSize << " is matched with " << Length << " frames\n" ; |
| 728 | } |
| 729 | |
| 730 | for (const auto &CallStack : MatchedCallSites) { |
| 731 | errs() << "MemProf callsite match for inline call stack" ; |
| 732 | for (uint64_t StackId : CallStack) |
| 733 | errs() << " " << StackId; |
| 734 | errs() << "\n" ; |
| 735 | } |
| 736 | } |
| 737 | |
| 738 | return PreservedAnalyses::none(); |
| 739 | } |
| 740 | |