| 1 | //===-- MemoryProfileInfo.cpp - memory profile info ------------------------==// |
| 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 contains utilities to analyze memory profile information. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/Analysis/MemoryProfileInfo.h" |
| 14 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 15 | #include "llvm/IR/Constants.h" |
| 16 | #include "llvm/Support/CommandLine.h" |
| 17 | #include "llvm/Support/Compiler.h" |
| 18 | #include "llvm/Support/Format.h" |
| 19 | |
| 20 | using namespace llvm; |
| 21 | using namespace llvm::memprof; |
| 22 | |
| 23 | #define DEBUG_TYPE "memory-profile-info" |
| 24 | |
| 25 | namespace llvm { |
| 26 | |
| 27 | cl::opt<bool> MemProfReportHintedSizes( |
| 28 | "memprof-report-hinted-sizes" , cl::init(Val: false), cl::Hidden, |
| 29 | cl::desc("Report total allocation sizes of hinted allocations" )); |
| 30 | |
| 31 | // This is useful if we have enabled reporting of hinted sizes, and want to get |
| 32 | // information from the indexing step for all contexts (especially for testing), |
| 33 | // or have specified a value less than 100% for -memprof-cloning-cold-threshold. |
| 34 | LLVM_ABI cl::opt<bool> MemProfKeepAllNotColdContexts( |
| 35 | "memprof-keep-all-not-cold-contexts" , cl::init(Val: false), cl::Hidden, |
| 36 | cl::desc("Keep all non-cold contexts (increases cloning overheads)" )); |
| 37 | |
| 38 | cl::opt<unsigned> MinClonedColdBytePercent( |
| 39 | "memprof-cloning-cold-threshold" , cl::init(Val: 100), cl::Hidden, |
| 40 | cl::desc("Min percent of cold bytes to hint alloc cold during cloning" )); |
| 41 | |
| 42 | // Discard non-cold contexts if they overlap with much larger cold contexts, |
| 43 | // specifically, if all contexts reaching a given callsite are at least this |
| 44 | // percent cold byte allocations. This reduces the amount of cloning required |
| 45 | // to expose the cold contexts when they greatly dominate non-cold contexts. |
| 46 | cl::opt<unsigned> MinCallsiteColdBytePercent( |
| 47 | "memprof-callsite-cold-threshold" , cl::init(Val: 100), cl::Hidden, |
| 48 | cl::desc("Min percent of cold bytes at a callsite to discard non-cold " |
| 49 | "contexts" )); |
| 50 | |
| 51 | // Enable saving context size information for largest cold contexts, which can |
| 52 | // be used to flag contexts for more aggressive cloning and reporting. |
| 53 | cl::opt<unsigned> MinPercentMaxColdSize( |
| 54 | "memprof-min-percent-max-cold-size" , cl::init(Val: 100), cl::Hidden, |
| 55 | cl::desc("Min percent of max cold bytes for critical cold context" )); |
| 56 | |
| 57 | LLVM_ABI cl::opt<bool> MemProfUseAmbiguousAttributes( |
| 58 | "memprof-ambiguous-attributes" , cl::init(Val: true), cl::Hidden, |
| 59 | cl::desc("Apply ambiguous memprof attribute to ambiguous allocations" )); |
| 60 | |
| 61 | } // end namespace llvm |
| 62 | |
| 63 | bool llvm::memprof::metadataIncludesAllContextSizeInfo() { |
| 64 | return MemProfReportHintedSizes || MinClonedColdBytePercent < 100; |
| 65 | } |
| 66 | |
| 67 | bool llvm::memprof::metadataMayIncludeContextSizeInfo() { |
| 68 | return metadataIncludesAllContextSizeInfo() || MinPercentMaxColdSize < 100; |
| 69 | } |
| 70 | |
| 71 | bool llvm::memprof::recordContextSizeInfoForAnalysis() { |
| 72 | return metadataMayIncludeContextSizeInfo() || |
| 73 | MinCallsiteColdBytePercent < 100; |
| 74 | } |
| 75 | |
| 76 | MDNode *llvm::memprof::buildCallstackMetadata(ArrayRef<uint64_t> CallStack, |
| 77 | LLVMContext &Ctx) { |
| 78 | SmallVector<Metadata *, 8> StackVals; |
| 79 | StackVals.reserve(N: CallStack.size()); |
| 80 | for (auto Id : CallStack) { |
| 81 | auto *StackValMD = |
| 82 | ValueAsMetadata::get(V: ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: Id)); |
| 83 | StackVals.push_back(Elt: StackValMD); |
| 84 | } |
| 85 | return MDNode::get(Context&: Ctx, MDs: StackVals); |
| 86 | } |
| 87 | |
| 88 | MDNode *llvm::memprof::getMIBStackNode(const MDNode *MIB) { |
| 89 | assert(MIB->getNumOperands() >= 2); |
| 90 | // The stack metadata is the first operand of each memprof MIB metadata. |
| 91 | return cast<MDNode>(Val: MIB->getOperand(I: 0)); |
| 92 | } |
| 93 | |
| 94 | AllocationType llvm::memprof::getMIBAllocType(const MDNode *MIB) { |
| 95 | assert(MIB->getNumOperands() >= 2); |
| 96 | // The allocation type is currently the second operand of each memprof |
| 97 | // MIB metadata. This will need to change as we add additional allocation |
| 98 | // types that can be applied based on the allocation profile data. |
| 99 | auto *MDS = dyn_cast<MDString>(Val: MIB->getOperand(I: 1)); |
| 100 | assert(MDS); |
| 101 | if (MDS->getString() == "cold" ) { |
| 102 | return AllocationType::Cold; |
| 103 | } else if (MDS->getString() == "hot" ) { |
| 104 | return AllocationType::Hot; |
| 105 | } |
| 106 | return AllocationType::NotCold; |
| 107 | } |
| 108 | |
| 109 | std::string llvm::memprof::getAllocTypeAttributeString(AllocationType Type) { |
| 110 | switch (Type) { |
| 111 | case AllocationType::NotCold: |
| 112 | return "notcold" ; |
| 113 | break; |
| 114 | case AllocationType::Cold: |
| 115 | return "cold" ; |
| 116 | break; |
| 117 | case AllocationType::Hot: |
| 118 | return "hot" ; |
| 119 | break; |
| 120 | default: |
| 121 | assert(false && "Unexpected alloc type" ); |
| 122 | } |
| 123 | llvm_unreachable("invalid alloc type" ); |
| 124 | } |
| 125 | |
| 126 | bool llvm::memprof::hasSingleAllocType(uint8_t AllocTypes) { |
| 127 | const unsigned NumAllocTypes = llvm::popcount(Value: AllocTypes); |
| 128 | assert(NumAllocTypes != 0); |
| 129 | return NumAllocTypes == 1; |
| 130 | } |
| 131 | |
| 132 | void llvm::memprof::removeAnyExistingAmbiguousAttribute(CallBase *CB) { |
| 133 | if (!CB->hasFnAttr(Kind: "memprof" )) |
| 134 | return; |
| 135 | assert(CB->getFnAttr("memprof" ).getValueAsString() == "ambiguous" ); |
| 136 | CB->removeFnAttr(Kind: "memprof" ); |
| 137 | } |
| 138 | |
| 139 | void llvm::memprof::addAmbiguousAttribute(CallBase *CB) { |
| 140 | if (!MemProfUseAmbiguousAttributes) |
| 141 | return; |
| 142 | // We may have an existing ambiguous attribute if we are reanalyzing |
| 143 | // after inlining. |
| 144 | if (CB->hasFnAttr(Kind: "memprof" )) { |
| 145 | assert(CB->getFnAttr("memprof" ).getValueAsString() == "ambiguous" ); |
| 146 | } else { |
| 147 | auto A = llvm::Attribute::get(Context&: CB->getContext(), Kind: "memprof" , Val: "ambiguous" ); |
| 148 | CB->addFnAttr(Attr: A); |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | void CallStackTrie::addCallStack( |
| 153 | AllocationType AllocType, ArrayRef<uint64_t> StackIds, |
| 154 | std::vector<ContextTotalSize> ContextSizeInfo) { |
| 155 | bool First = true; |
| 156 | CallStackTrieNode *Curr = nullptr; |
| 157 | for (auto StackId : StackIds) { |
| 158 | // If this is the first stack frame, add or update alloc node. |
| 159 | if (First) { |
| 160 | First = false; |
| 161 | if (Alloc) { |
| 162 | assert(AllocStackId == StackId); |
| 163 | Alloc->addAllocType(AllocType); |
| 164 | } else { |
| 165 | AllocStackId = StackId; |
| 166 | Alloc = new CallStackTrieNode(AllocType); |
| 167 | } |
| 168 | Curr = Alloc; |
| 169 | continue; |
| 170 | } |
| 171 | // Update existing caller node if it exists. |
| 172 | auto [Next, Inserted] = Curr->Callers.try_emplace(k: StackId); |
| 173 | if (!Inserted) { |
| 174 | Curr = Next->second; |
| 175 | Curr->addAllocType(AllocType); |
| 176 | continue; |
| 177 | } |
| 178 | // Otherwise add a new caller node. |
| 179 | auto *New = new CallStackTrieNode(AllocType); |
| 180 | Next->second = New; |
| 181 | Curr = New; |
| 182 | } |
| 183 | assert(Curr); |
| 184 | // Append all of the ContextSizeInfo, along with their original AllocType. |
| 185 | llvm::append_range(C&: Curr->ContextInfo, |
| 186 | R: llvm::map_range(C&: ContextSizeInfo, |
| 187 | F: [AllocType](const ContextTotalSize &CTS) { |
| 188 | return ContextSizeTypePair(CTS, |
| 189 | AllocType); |
| 190 | })); |
| 191 | } |
| 192 | |
| 193 | void CallStackTrie::addCallStack(MDNode *MIB) { |
| 194 | // Note that we are building this from existing MD_memprof metadata. |
| 195 | BuiltFromExistingMetadata = true; |
| 196 | MDNode *StackMD = getMIBStackNode(MIB); |
| 197 | assert(StackMD); |
| 198 | std::vector<uint64_t> CallStack; |
| 199 | CallStack.reserve(n: StackMD->getNumOperands()); |
| 200 | for (const auto &MIBStackIter : StackMD->operands()) { |
| 201 | auto *StackId = mdconst::dyn_extract<ConstantInt>(MD: MIBStackIter); |
| 202 | assert(StackId); |
| 203 | CallStack.push_back(x: StackId->getZExtValue()); |
| 204 | } |
| 205 | std::vector<ContextTotalSize> ContextSizeInfo; |
| 206 | // Collect the context size information if it exists. |
| 207 | if (MIB->getNumOperands() > 2) { |
| 208 | for (unsigned I = 2; I < MIB->getNumOperands(); I++) { |
| 209 | MDNode *ContextSizePair = dyn_cast<MDNode>(Val: MIB->getOperand(I)); |
| 210 | assert(ContextSizePair->getNumOperands() == 2); |
| 211 | uint64_t FullStackId = |
| 212 | mdconst::dyn_extract<ConstantInt>(MD: ContextSizePair->getOperand(I: 0)) |
| 213 | ->getZExtValue(); |
| 214 | uint64_t TotalSize = |
| 215 | mdconst::dyn_extract<ConstantInt>(MD: ContextSizePair->getOperand(I: 1)) |
| 216 | ->getZExtValue(); |
| 217 | ContextSizeInfo.push_back(x: {.FullStackId: FullStackId, .TotalSize: TotalSize}); |
| 218 | } |
| 219 | } |
| 220 | addCallStack(AllocType: getMIBAllocType(MIB), StackIds: CallStack, ContextSizeInfo: std::move(ContextSizeInfo)); |
| 221 | } |
| 222 | |
| 223 | static MDNode *createMIBNode(LLVMContext &Ctx, ArrayRef<uint64_t> MIBCallStack, |
| 224 | AllocationType AllocType, |
| 225 | ArrayRef<ContextSizeTypePair> ContextInfo, |
| 226 | const uint64_t MaxColdSize, |
| 227 | bool BuiltFromExistingMetadata, |
| 228 | uint64_t &TotalBytes, uint64_t &ColdBytes) { |
| 229 | SmallVector<Metadata *> MIBPayload( |
| 230 | {buildCallstackMetadata(CallStack: MIBCallStack, Ctx)}); |
| 231 | MIBPayload.push_back( |
| 232 | Elt: MDString::get(Context&: Ctx, Str: getAllocTypeAttributeString(Type: AllocType))); |
| 233 | |
| 234 | if (ContextInfo.empty()) { |
| 235 | // The profile matcher should have provided context size info if there was a |
| 236 | // MinCallsiteColdBytePercent < 100. Here we check >=100 to gracefully |
| 237 | // handle a user-provided percent larger than 100. However, we may not have |
| 238 | // this information if we built the Trie from existing MD_memprof metadata. |
| 239 | assert(BuiltFromExistingMetadata || MinCallsiteColdBytePercent >= 100); |
| 240 | return MDNode::get(Context&: Ctx, MDs: MIBPayload); |
| 241 | } |
| 242 | |
| 243 | for (const auto &[CSI, AT] : ContextInfo) { |
| 244 | const auto &[FullStackId, TotalSize] = CSI; |
| 245 | TotalBytes += TotalSize; |
| 246 | bool LargeColdContext = false; |
| 247 | if (AllocType == AllocationType::Cold) { |
| 248 | ColdBytes += TotalSize; |
| 249 | // If we have the max cold context size from summary information and have |
| 250 | // requested identification of contexts above a percentage of the max, see |
| 251 | // if this context qualifies. We should assume this is large if we rebuilt |
| 252 | // the trie from existing metadata (i.e. to update after inlining), in |
| 253 | // which case we don't have a MaxSize from the profile - we assume any |
| 254 | // context size info in existence on the metadata should be propagated. |
| 255 | if (BuiltFromExistingMetadata || |
| 256 | (MaxColdSize > 0 && MinPercentMaxColdSize < 100 && |
| 257 | TotalSize * 100 >= MaxColdSize * MinPercentMaxColdSize)) |
| 258 | LargeColdContext = true; |
| 259 | } |
| 260 | // Only add the context size info as metadata if we need it in the thin |
| 261 | // link (currently if reporting of hinted sizes is enabled, we have |
| 262 | // specified a threshold for marking allocations cold after cloning, or we |
| 263 | // have identified this as a large cold context of interest above). |
| 264 | if (metadataIncludesAllContextSizeInfo() || LargeColdContext) { |
| 265 | auto *FullStackIdMD = ValueAsMetadata::get( |
| 266 | V: ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: FullStackId)); |
| 267 | auto *TotalSizeMD = ValueAsMetadata::get( |
| 268 | V: ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: TotalSize)); |
| 269 | auto *ContextSizeMD = MDNode::get(Context&: Ctx, MDs: {FullStackIdMD, TotalSizeMD}); |
| 270 | MIBPayload.push_back(Elt: ContextSizeMD); |
| 271 | } |
| 272 | } |
| 273 | assert(TotalBytes > 0); |
| 274 | return MDNode::get(Context&: Ctx, MDs: MIBPayload); |
| 275 | } |
| 276 | |
| 277 | void CallStackTrie::collectContextInfo( |
| 278 | CallStackTrieNode *Node, std::vector<ContextSizeTypePair> &ContextInfo) { |
| 279 | llvm::append_range(C&: ContextInfo, R&: Node->ContextInfo); |
| 280 | for (auto &Caller : Node->Callers) |
| 281 | collectContextInfo(Node: Caller.second, ContextInfo); |
| 282 | } |
| 283 | |
| 284 | void CallStackTrie::convertHotToNotCold(CallStackTrieNode *Node) { |
| 285 | if (Node->hasAllocType(AllocType: AllocationType::Hot)) { |
| 286 | Node->removeAllocType(AllocType: AllocationType::Hot); |
| 287 | Node->addAllocType(AllocType: AllocationType::NotCold); |
| 288 | } |
| 289 | for (auto &Caller : Node->Callers) |
| 290 | convertHotToNotCold(Node: Caller.second); |
| 291 | } |
| 292 | |
| 293 | // Helper to emit messages for non-cold contexts that are ignored for various |
| 294 | // reasons when reporting of hinted bytes is enabled. |
| 295 | static void emitIgnoredNonColdContextMessage(StringRef Tag, |
| 296 | uint64_t FullStackId, |
| 297 | StringRef , |
| 298 | uint64_t TotalSize) { |
| 299 | errs() << "MemProf hinting: Total size for " << Tag |
| 300 | << " non-cold full allocation context hash " << FullStackId << Extra |
| 301 | << ": " << TotalSize << "\n" ; |
| 302 | } |
| 303 | |
| 304 | // Copy over some or all of NewMIBNodes to the SavedMIBNodes vector, depending |
| 305 | // on options that enable filtering out some NotCold contexts. |
| 306 | static void saveFilteredNewMIBNodes(std::vector<Metadata *> &NewMIBNodes, |
| 307 | std::vector<Metadata *> &SavedMIBNodes, |
| 308 | unsigned CallerContextLength, |
| 309 | uint64_t TotalBytes, uint64_t ColdBytes, |
| 310 | bool BuiltFromExistingMetadata) { |
| 311 | const bool MostlyCold = |
| 312 | // If we have built the Trie from existing MD_memprof metadata, we may or |
| 313 | // may not have context size information (in which case ColdBytes and |
| 314 | // TotalBytes are 0, which is not also guarded against below). Even if we |
| 315 | // do have some context size information from the the metadata, we have |
| 316 | // already gone through a round of discarding of small non-cold contexts |
| 317 | // during matching, and it would be overly aggressive to do it again, and |
| 318 | // we also want to maintain the same behavior with and without reporting |
| 319 | // of hinted bytes enabled. |
| 320 | !BuiltFromExistingMetadata && MinCallsiteColdBytePercent < 100 && |
| 321 | ColdBytes > 0 && |
| 322 | ColdBytes * 100 >= MinCallsiteColdBytePercent * TotalBytes; |
| 323 | |
| 324 | // In the simplest case, with pruning disabled, keep all the new MIB nodes. |
| 325 | if (MemProfKeepAllNotColdContexts && !MostlyCold) { |
| 326 | append_range(C&: SavedMIBNodes, R&: NewMIBNodes); |
| 327 | return; |
| 328 | } |
| 329 | |
| 330 | auto EmitMessageForRemovedContexts = [](const MDNode *MIBMD, StringRef Tag, |
| 331 | StringRef ) { |
| 332 | assert(MIBMD->getNumOperands() > 2); |
| 333 | for (unsigned I = 2; I < MIBMD->getNumOperands(); I++) { |
| 334 | MDNode *ContextSizePair = dyn_cast<MDNode>(Val: MIBMD->getOperand(I)); |
| 335 | assert(ContextSizePair->getNumOperands() == 2); |
| 336 | uint64_t FullStackId = |
| 337 | mdconst::dyn_extract<ConstantInt>(MD: ContextSizePair->getOperand(I: 0)) |
| 338 | ->getZExtValue(); |
| 339 | uint64_t TS = |
| 340 | mdconst::dyn_extract<ConstantInt>(MD: ContextSizePair->getOperand(I: 1)) |
| 341 | ->getZExtValue(); |
| 342 | emitIgnoredNonColdContextMessage(Tag, FullStackId, Extra, TotalSize: TS); |
| 343 | } |
| 344 | }; |
| 345 | |
| 346 | // If the cold bytes at the current callsite exceed the given threshold, we |
| 347 | // discard all non-cold contexts so do not need any of the later pruning |
| 348 | // handling. We can simply copy over all the cold contexts and return early. |
| 349 | if (MostlyCold) { |
| 350 | auto NewColdMIBNodes = |
| 351 | make_filter_range(Range&: NewMIBNodes, Pred: [&](const Metadata *M) { |
| 352 | auto MIBMD = cast<MDNode>(Val: M); |
| 353 | // Only append cold contexts. |
| 354 | if (getMIBAllocType(MIB: MIBMD) == AllocationType::Cold) |
| 355 | return true; |
| 356 | if (MemProfReportHintedSizes) { |
| 357 | const float PercentCold = ColdBytes * 100.0 / TotalBytes; |
| 358 | std::string PercentStr; |
| 359 | llvm::raw_string_ostream OS(PercentStr); |
| 360 | OS << format(Fmt: " for %5.2f%% cold bytes" , Vals: PercentCold); |
| 361 | EmitMessageForRemovedContexts(MIBMD, "discarded" , OS.str()); |
| 362 | } |
| 363 | return false; |
| 364 | }); |
| 365 | for (auto *M : NewColdMIBNodes) |
| 366 | SavedMIBNodes.push_back(x: M); |
| 367 | return; |
| 368 | } |
| 369 | |
| 370 | // Prune unneeded NotCold contexts, taking advantage of the fact |
| 371 | // that we later will only clone Cold contexts, as NotCold is the allocation |
| 372 | // default. We only need to keep as metadata the NotCold contexts that |
| 373 | // overlap the longest with Cold allocations, so that we know how deeply we |
| 374 | // need to clone. For example, assume we add the following contexts to the |
| 375 | // trie: |
| 376 | // 1 3 (notcold) |
| 377 | // 1 2 4 (cold) |
| 378 | // 1 2 5 (notcold) |
| 379 | // 1 2 6 (notcold) |
| 380 | // the trie looks like: |
| 381 | // 1 |
| 382 | // / \ |
| 383 | // 2 3 |
| 384 | // /|\ |
| 385 | // 4 5 6 |
| 386 | // |
| 387 | // It is sufficient to prune all but one not-cold contexts (either 1,2,5 or |
| 388 | // 1,2,6, we arbitrarily keep the first one we encounter which will be |
| 389 | // 1,2,5). |
| 390 | // |
| 391 | // To do this pruning, we first check if there were any not-cold |
| 392 | // contexts kept for a deeper caller, which will have a context length larger |
| 393 | // than the CallerContextLength being handled here (i.e. kept by a deeper |
| 394 | // recursion step). If so, none of the not-cold MIB nodes added for the |
| 395 | // immediate callers need to be kept. If not, we keep the first (created |
| 396 | // for the immediate caller) not-cold MIB node. |
| 397 | bool LongerNotColdContextKept = false; |
| 398 | for (auto *MIB : NewMIBNodes) { |
| 399 | auto MIBMD = cast<MDNode>(Val: MIB); |
| 400 | if (getMIBAllocType(MIB: MIBMD) == AllocationType::Cold) |
| 401 | continue; |
| 402 | MDNode *StackMD = getMIBStackNode(MIB: MIBMD); |
| 403 | assert(StackMD); |
| 404 | if (StackMD->getNumOperands() > CallerContextLength) { |
| 405 | LongerNotColdContextKept = true; |
| 406 | break; |
| 407 | } |
| 408 | } |
| 409 | // Don't need to emit any for the immediate caller if we already have |
| 410 | // longer overlapping contexts; |
| 411 | bool KeepFirstNewNotCold = !LongerNotColdContextKept; |
| 412 | auto NewColdMIBNodes = make_filter_range(Range&: NewMIBNodes, Pred: [&](const Metadata *M) { |
| 413 | auto MIBMD = cast<MDNode>(Val: M); |
| 414 | // Only keep cold contexts and first (longest non-cold context). |
| 415 | if (getMIBAllocType(MIB: MIBMD) != AllocationType::Cold) { |
| 416 | MDNode *StackMD = getMIBStackNode(MIB: MIBMD); |
| 417 | assert(StackMD); |
| 418 | // Keep any already kept for longer contexts. |
| 419 | if (StackMD->getNumOperands() > CallerContextLength) |
| 420 | return true; |
| 421 | // Otherwise keep the first one added by the immediate caller if there |
| 422 | // were no longer contexts. |
| 423 | if (KeepFirstNewNotCold) { |
| 424 | KeepFirstNewNotCold = false; |
| 425 | return true; |
| 426 | } |
| 427 | if (MemProfReportHintedSizes) |
| 428 | EmitMessageForRemovedContexts(MIBMD, "pruned" , "" ); |
| 429 | return false; |
| 430 | } |
| 431 | return true; |
| 432 | }); |
| 433 | for (auto *M : NewColdMIBNodes) |
| 434 | SavedMIBNodes.push_back(x: M); |
| 435 | } |
| 436 | |
| 437 | // Recursive helper to trim contexts and create metadata nodes. |
| 438 | // Caller should have pushed Node's loc to MIBCallStack. Doing this in the |
| 439 | // caller makes it simpler to handle the many early returns in this method. |
| 440 | // Updates the total and cold profiled bytes in the subtrie rooted at this node. |
| 441 | bool CallStackTrie::buildMIBNodes(CallStackTrieNode *Node, LLVMContext &Ctx, |
| 442 | std::vector<uint64_t> &MIBCallStack, |
| 443 | std::vector<Metadata *> &MIBNodes, |
| 444 | bool CalleeHasAmbiguousCallerContext, |
| 445 | uint64_t &TotalBytes, uint64_t &ColdBytes) { |
| 446 | // Trim context below the first node in a prefix with a single alloc type. |
| 447 | // Add an MIB record for the current call stack prefix. |
| 448 | if (hasSingleAllocType(AllocTypes: Node->AllocTypes)) { |
| 449 | std::vector<ContextSizeTypePair> ContextInfo; |
| 450 | collectContextInfo(Node, ContextInfo); |
| 451 | MIBNodes.push_back(x: createMIBNode( |
| 452 | Ctx, MIBCallStack, AllocType: (AllocationType)Node->AllocTypes, ContextInfo, |
| 453 | MaxColdSize, BuiltFromExistingMetadata, TotalBytes, ColdBytes)); |
| 454 | return true; |
| 455 | } |
| 456 | |
| 457 | // We don't have a single allocation for all the contexts sharing this prefix, |
| 458 | // so recursively descend into callers in trie. |
| 459 | if (!Node->Callers.empty()) { |
| 460 | bool NodeHasAmbiguousCallerContext = Node->Callers.size() > 1; |
| 461 | bool AddedMIBNodesForAllCallerContexts = true; |
| 462 | // Accumulate all new MIB nodes by the recursive calls below into a vector |
| 463 | // that will later be filtered before adding to the caller's MIBNodes |
| 464 | // vector. |
| 465 | std::vector<Metadata *> NewMIBNodes; |
| 466 | // Determine the total and cold byte counts for all callers, then add to the |
| 467 | // caller's counts further below. |
| 468 | uint64_t CallerTotalBytes = 0; |
| 469 | uint64_t CallerColdBytes = 0; |
| 470 | for (auto &Caller : Node->Callers) { |
| 471 | MIBCallStack.push_back(x: Caller.first); |
| 472 | AddedMIBNodesForAllCallerContexts &= buildMIBNodes( |
| 473 | Node: Caller.second, Ctx, MIBCallStack, MIBNodes&: NewMIBNodes, |
| 474 | CalleeHasAmbiguousCallerContext: NodeHasAmbiguousCallerContext, TotalBytes&: CallerTotalBytes, ColdBytes&: CallerColdBytes); |
| 475 | // Remove Caller. |
| 476 | MIBCallStack.pop_back(); |
| 477 | } |
| 478 | // Pass in the stack length of the MIB nodes added for the immediate caller, |
| 479 | // which is the current stack length plus 1. |
| 480 | saveFilteredNewMIBNodes(NewMIBNodes, SavedMIBNodes&: MIBNodes, CallerContextLength: MIBCallStack.size() + 1, |
| 481 | TotalBytes: CallerTotalBytes, ColdBytes: CallerColdBytes, |
| 482 | BuiltFromExistingMetadata); |
| 483 | TotalBytes += CallerTotalBytes; |
| 484 | ColdBytes += CallerColdBytes; |
| 485 | |
| 486 | if (AddedMIBNodesForAllCallerContexts) |
| 487 | return true; |
| 488 | // We expect that the callers should be forced to add MIBs to disambiguate |
| 489 | // the context in this case (see below). |
| 490 | assert(!NodeHasAmbiguousCallerContext); |
| 491 | } |
| 492 | |
| 493 | // If we reached here, then this node does not have a single allocation type, |
| 494 | // and we didn't add metadata for a longer call stack prefix including any of |
| 495 | // Node's callers. That means we never hit a single allocation type along all |
| 496 | // call stacks with this prefix. This can happen due to recursion collapsing |
| 497 | // or the stack being deeper than tracked by the profiler runtime, leading to |
| 498 | // contexts with different allocation types being merged. In that case, we |
| 499 | // trim the context just below the deepest context split, which is this |
| 500 | // node if the callee has an ambiguous caller context (multiple callers), |
| 501 | // since the recursive calls above returned false. Conservatively give it |
| 502 | // non-cold allocation type. |
| 503 | if (!CalleeHasAmbiguousCallerContext) |
| 504 | return false; |
| 505 | std::vector<ContextSizeTypePair> ContextInfo; |
| 506 | collectContextInfo(Node, ContextInfo); |
| 507 | MIBNodes.push_back(x: createMIBNode( |
| 508 | Ctx, MIBCallStack, AllocType: AllocationType::NotCold, ContextInfo, MaxColdSize, |
| 509 | BuiltFromExistingMetadata, TotalBytes, ColdBytes)); |
| 510 | return true; |
| 511 | } |
| 512 | |
| 513 | void CallStackTrie::addSingleAllocTypeAttribute(CallBase *CI, AllocationType AT, |
| 514 | StringRef Descriptor) { |
| 515 | auto AllocTypeString = getAllocTypeAttributeString(Type: AT); |
| 516 | auto A = llvm::Attribute::get(Context&: CI->getContext(), Kind: "memprof" , Val: AllocTypeString); |
| 517 | // After inlining we may be able to convert an existing ambiguous allocation |
| 518 | // to an unambiguous one. |
| 519 | removeAnyExistingAmbiguousAttribute(CB: CI); |
| 520 | CI->addFnAttr(Attr: A); |
| 521 | if (MemProfReportHintedSizes) { |
| 522 | std::vector<ContextSizeTypePair> ContextInfo; |
| 523 | collectContextInfo(Node: Alloc, ContextInfo); |
| 524 | for (const auto &[CSI, OrigAT] : ContextInfo) { |
| 525 | const auto &[FullStackId, TotalSize] = CSI; |
| 526 | // If the original alloc type is not the one being applied as the hint, |
| 527 | // report that we ignored this context. |
| 528 | if (AT != OrigAT) { |
| 529 | emitIgnoredNonColdContextMessage(Tag: "ignored" , FullStackId, Extra: "" , TotalSize); |
| 530 | continue; |
| 531 | } |
| 532 | errs() << "MemProf hinting: Total size for full allocation context hash " |
| 533 | << FullStackId << " and " << Descriptor << " alloc type " |
| 534 | << getAllocTypeAttributeString(Type: AT) << ": " << TotalSize << "\n" ; |
| 535 | } |
| 536 | } |
| 537 | if (ORE) |
| 538 | ORE->emit(OptDiag: OptimizationRemark(DEBUG_TYPE, "MemprofAttribute" , CI) |
| 539 | << ore::NV("AllocationCall" , CI) << " in function " |
| 540 | << ore::NV("Caller" , CI->getFunction()) |
| 541 | << " marked with memprof allocation attribute " |
| 542 | << ore::NV("Attribute" , AllocTypeString)); |
| 543 | } |
| 544 | |
| 545 | // Build and attach the minimal necessary MIB metadata. If the alloc has a |
| 546 | // single allocation type, add a function attribute instead. Returns true if |
| 547 | // memprof metadata attached, false if not (attribute added). |
| 548 | bool CallStackTrie::buildAndAttachMIBMetadata(CallBase *CI) { |
| 549 | if (hasSingleAllocType(AllocTypes: Alloc->AllocTypes)) { |
| 550 | addSingleAllocTypeAttribute(CI, AT: (AllocationType)Alloc->AllocTypes, |
| 551 | Descriptor: "single" ); |
| 552 | return false; |
| 553 | } |
| 554 | // If there were any hot allocation contexts, the Alloc trie node would have |
| 555 | // the Hot type set. If so, because we don't currently support cloning for hot |
| 556 | // contexts, they should be converted to NotCold. This happens in the cloning |
| 557 | // support anyway, however, doing this now enables more aggressive context |
| 558 | // trimming when building the MIB metadata (and possibly may make the |
| 559 | // allocation have a single NotCold allocation type), greatly reducing |
| 560 | // overheads in bitcode, cloning memory and cloning time. |
| 561 | if (Alloc->hasAllocType(AllocType: AllocationType::Hot)) { |
| 562 | convertHotToNotCold(Node: Alloc); |
| 563 | // Check whether we now have a single alloc type. |
| 564 | if (hasSingleAllocType(AllocTypes: Alloc->AllocTypes)) { |
| 565 | addSingleAllocTypeAttribute(CI, AT: (AllocationType)Alloc->AllocTypes, |
| 566 | Descriptor: "single" ); |
| 567 | return false; |
| 568 | } |
| 569 | } |
| 570 | auto &Ctx = CI->getContext(); |
| 571 | std::vector<uint64_t> MIBCallStack; |
| 572 | MIBCallStack.push_back(x: AllocStackId); |
| 573 | std::vector<Metadata *> MIBNodes; |
| 574 | uint64_t TotalBytes = 0; |
| 575 | uint64_t ColdBytes = 0; |
| 576 | assert(!Alloc->Callers.empty() && "addCallStack has not been called yet" ); |
| 577 | // The CalleeHasAmbiguousCallerContext flag is meant to say whether the |
| 578 | // callee of the given node has more than one caller. Here the node being |
| 579 | // passed in is the alloc and it has no callees. So it's false. |
| 580 | if (buildMIBNodes(Node: Alloc, Ctx, MIBCallStack, MIBNodes, |
| 581 | /*CalleeHasAmbiguousCallerContext=*/false, TotalBytes, |
| 582 | ColdBytes)) { |
| 583 | assert(MIBCallStack.size() == 1 && |
| 584 | "Should only be left with Alloc's location in stack" ); |
| 585 | CI->setMetadata(KindID: LLVMContext::MD_memprof, Node: MDNode::get(Context&: Ctx, MDs: MIBNodes)); |
| 586 | addAmbiguousAttribute(CB: CI); |
| 587 | return true; |
| 588 | } |
| 589 | // If there exists corner case that CallStackTrie has one chain to leaf |
| 590 | // and all node in the chain have multi alloc type, conservatively give |
| 591 | // it non-cold allocation type. |
| 592 | // FIXME: Avoid this case before memory profile created. Alternatively, select |
| 593 | // hint based on fraction cold. |
| 594 | addSingleAllocTypeAttribute(CI, AT: AllocationType::NotCold, Descriptor: "indistinguishable" ); |
| 595 | return false; |
| 596 | } |
| 597 | |
| 598 | template <> |
| 599 | CallStack<MDNode, MDNode::op_iterator>::CallStackIterator::CallStackIterator( |
| 600 | const MDNode *N, bool End) |
| 601 | : N(N) { |
| 602 | if (!N) |
| 603 | return; |
| 604 | Iter = End ? N->op_end() : N->op_begin(); |
| 605 | } |
| 606 | |
| 607 | template <> |
| 608 | uint64_t |
| 609 | CallStack<MDNode, MDNode::op_iterator>::CallStackIterator::operator*() { |
| 610 | assert(Iter != N->op_end()); |
| 611 | ConstantInt *StackIdCInt = mdconst::dyn_extract<ConstantInt>(MD: *Iter); |
| 612 | assert(StackIdCInt); |
| 613 | return StackIdCInt->getZExtValue(); |
| 614 | } |
| 615 | |
| 616 | template <> uint64_t CallStack<MDNode, MDNode::op_iterator>::back() const { |
| 617 | assert(N); |
| 618 | return mdconst::dyn_extract<ConstantInt>(MD: N->operands().back()) |
| 619 | ->getZExtValue(); |
| 620 | } |
| 621 | |
| 622 | MDNode *MDNode::getMergedMemProfMetadata(MDNode *A, MDNode *B) { |
| 623 | // TODO: Support more sophisticated merging, such as selecting the one with |
| 624 | // more bytes allocated, or implement support for carrying multiple allocation |
| 625 | // leaf contexts. For now, keep the first one. |
| 626 | if (A) |
| 627 | return A; |
| 628 | return B; |
| 629 | } |
| 630 | |
| 631 | MDNode *MDNode::getMergedCallsiteMetadata(MDNode *A, MDNode *B) { |
| 632 | // TODO: Support more sophisticated merging, which will require support for |
| 633 | // carrying multiple contexts. For now, keep the first one. |
| 634 | if (A) |
| 635 | return A; |
| 636 | return B; |
| 637 | } |
| 638 | |