| 1 | //===-- xray_function_call_trie.h ------------------------------*- 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 is a part of XRay, a dynamic runtime instrumentation system. |
| 10 | // |
| 11 | // This file defines the interface for a function call trie. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | #ifndef XRAY_FUNCTION_CALL_TRIE_H |
| 15 | #define XRAY_FUNCTION_CALL_TRIE_H |
| 16 | |
| 17 | #include "xray_buffer_queue.h" |
| 18 | #include "xray_defs.h" |
| 19 | #include "xray_profiling_flags.h" |
| 20 | #include "xray_segmented_array.h" |
| 21 | #include <limits> |
| 22 | #include <memory> // For placement new. |
| 23 | #include <utility> |
| 24 | |
| 25 | namespace __xray { |
| 26 | |
| 27 | /// A FunctionCallTrie represents the stack traces of XRay instrumented |
| 28 | /// functions that we've encountered, where a node corresponds to a function and |
| 29 | /// the path from the root to the node its stack trace. Each node in the trie |
| 30 | /// will contain some useful values, including: |
| 31 | /// |
| 32 | /// * The cumulative amount of time spent in this particular node/stack. |
| 33 | /// * The number of times this stack has appeared. |
| 34 | /// * A histogram of latencies for that particular node. |
| 35 | /// |
| 36 | /// Each node in the trie will also contain a list of callees, represented using |
| 37 | /// a Array<NodeIdPair> -- each NodeIdPair instance will contain the function |
| 38 | /// ID of the callee, and a pointer to the node. |
| 39 | /// |
| 40 | /// If we visualise this data structure, we'll find the following potential |
| 41 | /// representation: |
| 42 | /// |
| 43 | /// [function id node] -> [callees] [cumulative time] |
| 44 | /// [call counter] [latency histogram] |
| 45 | /// |
| 46 | /// As an example, when we have a function in this pseudocode: |
| 47 | /// |
| 48 | /// func f(N) { |
| 49 | /// g() |
| 50 | /// h() |
| 51 | /// for i := 1..N { j() } |
| 52 | /// } |
| 53 | /// |
| 54 | /// We may end up with a trie of the following form: |
| 55 | /// |
| 56 | /// f -> [ g, h, j ] [...] [1] [...] |
| 57 | /// g -> [ ... ] [...] [1] [...] |
| 58 | /// h -> [ ... ] [...] [1] [...] |
| 59 | /// j -> [ ... ] [...] [N] [...] |
| 60 | /// |
| 61 | /// If for instance the function g() called j() like so: |
| 62 | /// |
| 63 | /// func g() { |
| 64 | /// for i := 1..10 { j() } |
| 65 | /// } |
| 66 | /// |
| 67 | /// We'll find the following updated trie: |
| 68 | /// |
| 69 | /// f -> [ g, h, j ] [...] [1] [...] |
| 70 | /// g -> [ j' ] [...] [1] [...] |
| 71 | /// h -> [ ... ] [...] [1] [...] |
| 72 | /// j -> [ ... ] [...] [N] [...] |
| 73 | /// j' -> [ ... ] [...] [10] [...] |
| 74 | /// |
| 75 | /// Note that we'll have a new node representing the path `f -> g -> j'` with |
| 76 | /// isolated data. This isolation gives us a means of representing the stack |
| 77 | /// traces as a path, as opposed to a key in a table. The alternative |
| 78 | /// implementation here would be to use a separate table for the path, and use |
| 79 | /// hashes of the path as an identifier to accumulate the information. We've |
| 80 | /// moved away from this approach as it takes a lot of time to compute the hash |
| 81 | /// every time we need to update a function's call information as we're handling |
| 82 | /// the entry and exit events. |
| 83 | /// |
| 84 | /// This approach allows us to maintain a shadow stack, which represents the |
| 85 | /// currently executing path, and on function exits quickly compute the amount |
| 86 | /// of time elapsed from the entry, then update the counters for the node |
| 87 | /// already represented in the trie. This necessitates an efficient |
| 88 | /// representation of the various data structures (the list of callees must be |
| 89 | /// cache-aware and efficient to look up, and the histogram must be compact and |
| 90 | /// quick to update) to enable us to keep the overheads of this implementation |
| 91 | /// to the minimum. |
| 92 | class FunctionCallTrie { |
| 93 | public: |
| 94 | struct Node; |
| 95 | |
| 96 | // We use a NodeIdPair type instead of a std::pair<...> to not rely on the |
| 97 | // standard library types in this header. |
| 98 | struct NodeIdPair { |
| 99 | Node *NodePtr; |
| 100 | int32_t FId; |
| 101 | }; |
| 102 | |
| 103 | using NodeIdPairArray = Array<NodeIdPair>; |
| 104 | using NodeIdPairAllocatorType = NodeIdPairArray::AllocatorType; |
| 105 | |
| 106 | // A Node in the FunctionCallTrie gives us a list of callees, the cumulative |
| 107 | // number of times this node actually appeared, the cumulative amount of time |
| 108 | // for this particular node including its children call times, and just the |
| 109 | // local time spent on this node. Each Node will have the ID of the XRay |
| 110 | // instrumented function that it is associated to. |
| 111 | struct Node { |
| 112 | Node *Parent; |
| 113 | NodeIdPairArray Callees; |
| 114 | uint64_t CallCount; |
| 115 | uint64_t CumulativeLocalTime; // Typically in TSC deltas, not wall-time. |
| 116 | int32_t FId; |
| 117 | |
| 118 | // TODO: Include the compact histogram. |
| 119 | }; |
| 120 | |
| 121 | private: |
| 122 | struct ShadowStackEntry { |
| 123 | uint64_t EntryTSC; |
| 124 | Node *NodePtr; |
| 125 | uint16_t EntryCPU; |
| 126 | }; |
| 127 | |
| 128 | using NodeArray = Array<Node>; |
| 129 | using RootArray = Array<Node *>; |
| 130 | using ShadowStackArray = Array<ShadowStackEntry>; |
| 131 | |
| 132 | public: |
| 133 | // We collate the allocators we need into a single struct, as a convenience to |
| 134 | // allow us to initialize these as a group. |
| 135 | struct Allocators { |
| 136 | using NodeAllocatorType = NodeArray::AllocatorType; |
| 137 | using RootAllocatorType = RootArray::AllocatorType; |
| 138 | using ShadowStackAllocatorType = ShadowStackArray::AllocatorType; |
| 139 | |
| 140 | // Use hosted aligned storage members to allow for trivial move and init. |
| 141 | // This also allows us to sidestep the potential-failing allocation issue. |
| 142 | alignas(NodeAllocatorType) std::byte |
| 143 | NodeAllocatorStorage[sizeof(NodeAllocatorType)]; |
| 144 | alignas(RootAllocatorType) std::byte |
| 145 | RootAllocatorStorage[sizeof(RootAllocatorType)]; |
| 146 | alignas(ShadowStackAllocatorType) std::byte |
| 147 | ShadowStackAllocatorStorage[sizeof(ShadowStackAllocatorType)]; |
| 148 | alignas(NodeIdPairAllocatorType) std::byte |
| 149 | NodeIdPairAllocatorStorage[sizeof(NodeIdPairAllocatorType)]; |
| 150 | |
| 151 | NodeAllocatorType *NodeAllocator = nullptr; |
| 152 | RootAllocatorType *RootAllocator = nullptr; |
| 153 | ShadowStackAllocatorType *ShadowStackAllocator = nullptr; |
| 154 | NodeIdPairAllocatorType *NodeIdPairAllocator = nullptr; |
| 155 | |
| 156 | Allocators() = default; |
| 157 | Allocators(const Allocators &) = delete; |
| 158 | Allocators &operator=(const Allocators &) = delete; |
| 159 | |
| 160 | struct Buffers { |
| 161 | BufferQueue::Buffer NodeBuffer; |
| 162 | BufferQueue::Buffer RootsBuffer; |
| 163 | BufferQueue::Buffer ShadowStackBuffer; |
| 164 | BufferQueue::Buffer NodeIdPairBuffer; |
| 165 | }; |
| 166 | |
| 167 | explicit Allocators(Buffers &B) XRAY_NEVER_INSTRUMENT { |
| 168 | new (&NodeAllocatorStorage) |
| 169 | NodeAllocatorType(B.NodeBuffer.Data, B.NodeBuffer.Size); |
| 170 | NodeAllocator = |
| 171 | reinterpret_cast<NodeAllocatorType *>(&NodeAllocatorStorage); |
| 172 | |
| 173 | new (&RootAllocatorStorage) |
| 174 | RootAllocatorType(B.RootsBuffer.Data, B.RootsBuffer.Size); |
| 175 | RootAllocator = |
| 176 | reinterpret_cast<RootAllocatorType *>(&RootAllocatorStorage); |
| 177 | |
| 178 | new (&ShadowStackAllocatorStorage) ShadowStackAllocatorType( |
| 179 | B.ShadowStackBuffer.Data, B.ShadowStackBuffer.Size); |
| 180 | ShadowStackAllocator = reinterpret_cast<ShadowStackAllocatorType *>( |
| 181 | &ShadowStackAllocatorStorage); |
| 182 | |
| 183 | new (&NodeIdPairAllocatorStorage) NodeIdPairAllocatorType( |
| 184 | B.NodeIdPairBuffer.Data, B.NodeIdPairBuffer.Size); |
| 185 | NodeIdPairAllocator = reinterpret_cast<NodeIdPairAllocatorType *>( |
| 186 | &NodeIdPairAllocatorStorage); |
| 187 | } |
| 188 | |
| 189 | explicit Allocators(uptr Max) XRAY_NEVER_INSTRUMENT { |
| 190 | new (&NodeAllocatorStorage) NodeAllocatorType(Max); |
| 191 | NodeAllocator = |
| 192 | reinterpret_cast<NodeAllocatorType *>(&NodeAllocatorStorage); |
| 193 | |
| 194 | new (&RootAllocatorStorage) RootAllocatorType(Max); |
| 195 | RootAllocator = |
| 196 | reinterpret_cast<RootAllocatorType *>(&RootAllocatorStorage); |
| 197 | |
| 198 | new (&ShadowStackAllocatorStorage) ShadowStackAllocatorType(Max); |
| 199 | ShadowStackAllocator = reinterpret_cast<ShadowStackAllocatorType *>( |
| 200 | &ShadowStackAllocatorStorage); |
| 201 | |
| 202 | new (&NodeIdPairAllocatorStorage) NodeIdPairAllocatorType(Max); |
| 203 | NodeIdPairAllocator = reinterpret_cast<NodeIdPairAllocatorType *>( |
| 204 | &NodeIdPairAllocatorStorage); |
| 205 | } |
| 206 | |
| 207 | Allocators(Allocators &&O) XRAY_NEVER_INSTRUMENT { |
| 208 | // Here we rely on the safety of memcpy'ing contents of the storage |
| 209 | // members, and then pointing the source pointers to nullptr. |
| 210 | internal_memcpy(dest: &NodeAllocatorStorage, src: &O.NodeAllocatorStorage, |
| 211 | n: sizeof(NodeAllocatorType)); |
| 212 | internal_memcpy(dest: &RootAllocatorStorage, src: &O.RootAllocatorStorage, |
| 213 | n: sizeof(RootAllocatorType)); |
| 214 | internal_memcpy(dest: &ShadowStackAllocatorStorage, |
| 215 | src: &O.ShadowStackAllocatorStorage, |
| 216 | n: sizeof(ShadowStackAllocatorType)); |
| 217 | internal_memcpy(dest: &NodeIdPairAllocatorStorage, |
| 218 | src: &O.NodeIdPairAllocatorStorage, |
| 219 | n: sizeof(NodeIdPairAllocatorType)); |
| 220 | |
| 221 | NodeAllocator = |
| 222 | reinterpret_cast<NodeAllocatorType *>(&NodeAllocatorStorage); |
| 223 | RootAllocator = |
| 224 | reinterpret_cast<RootAllocatorType *>(&RootAllocatorStorage); |
| 225 | ShadowStackAllocator = reinterpret_cast<ShadowStackAllocatorType *>( |
| 226 | &ShadowStackAllocatorStorage); |
| 227 | NodeIdPairAllocator = reinterpret_cast<NodeIdPairAllocatorType *>( |
| 228 | &NodeIdPairAllocatorStorage); |
| 229 | |
| 230 | O.NodeAllocator = nullptr; |
| 231 | O.RootAllocator = nullptr; |
| 232 | O.ShadowStackAllocator = nullptr; |
| 233 | O.NodeIdPairAllocator = nullptr; |
| 234 | } |
| 235 | |
| 236 | Allocators &operator=(Allocators &&O) XRAY_NEVER_INSTRUMENT { |
| 237 | // When moving into an existing instance, we ensure that we clean up the |
| 238 | // current allocators. |
| 239 | if (NodeAllocator) |
| 240 | NodeAllocator->~NodeAllocatorType(); |
| 241 | if (O.NodeAllocator) { |
| 242 | new (&NodeAllocatorStorage) |
| 243 | NodeAllocatorType(std::move(t&: *O.NodeAllocator)); |
| 244 | NodeAllocator = |
| 245 | reinterpret_cast<NodeAllocatorType *>(&NodeAllocatorStorage); |
| 246 | O.NodeAllocator = nullptr; |
| 247 | } else { |
| 248 | NodeAllocator = nullptr; |
| 249 | } |
| 250 | |
| 251 | if (RootAllocator) |
| 252 | RootAllocator->~RootAllocatorType(); |
| 253 | if (O.RootAllocator) { |
| 254 | new (&RootAllocatorStorage) |
| 255 | RootAllocatorType(std::move(t&: *O.RootAllocator)); |
| 256 | RootAllocator = |
| 257 | reinterpret_cast<RootAllocatorType *>(&RootAllocatorStorage); |
| 258 | O.RootAllocator = nullptr; |
| 259 | } else { |
| 260 | RootAllocator = nullptr; |
| 261 | } |
| 262 | |
| 263 | if (ShadowStackAllocator) |
| 264 | ShadowStackAllocator->~ShadowStackAllocatorType(); |
| 265 | if (O.ShadowStackAllocator) { |
| 266 | new (&ShadowStackAllocatorStorage) |
| 267 | ShadowStackAllocatorType(std::move(t&: *O.ShadowStackAllocator)); |
| 268 | ShadowStackAllocator = reinterpret_cast<ShadowStackAllocatorType *>( |
| 269 | &ShadowStackAllocatorStorage); |
| 270 | O.ShadowStackAllocator = nullptr; |
| 271 | } else { |
| 272 | ShadowStackAllocator = nullptr; |
| 273 | } |
| 274 | |
| 275 | if (NodeIdPairAllocator) |
| 276 | NodeIdPairAllocator->~NodeIdPairAllocatorType(); |
| 277 | if (O.NodeIdPairAllocator) { |
| 278 | new (&NodeIdPairAllocatorStorage) |
| 279 | NodeIdPairAllocatorType(std::move(t&: *O.NodeIdPairAllocator)); |
| 280 | NodeIdPairAllocator = reinterpret_cast<NodeIdPairAllocatorType *>( |
| 281 | &NodeIdPairAllocatorStorage); |
| 282 | O.NodeIdPairAllocator = nullptr; |
| 283 | } else { |
| 284 | NodeIdPairAllocator = nullptr; |
| 285 | } |
| 286 | |
| 287 | return *this; |
| 288 | } |
| 289 | |
| 290 | ~Allocators() XRAY_NEVER_INSTRUMENT { |
| 291 | if (NodeAllocator != nullptr) |
| 292 | NodeAllocator->~NodeAllocatorType(); |
| 293 | if (RootAllocator != nullptr) |
| 294 | RootAllocator->~RootAllocatorType(); |
| 295 | if (ShadowStackAllocator != nullptr) |
| 296 | ShadowStackAllocator->~ShadowStackAllocatorType(); |
| 297 | if (NodeIdPairAllocator != nullptr) |
| 298 | NodeIdPairAllocator->~NodeIdPairAllocatorType(); |
| 299 | } |
| 300 | }; |
| 301 | |
| 302 | static Allocators InitAllocators() XRAY_NEVER_INSTRUMENT { |
| 303 | return InitAllocatorsCustom(Max: profilingFlags()->per_thread_allocator_max); |
| 304 | } |
| 305 | |
| 306 | static Allocators InitAllocatorsCustom(uptr Max) XRAY_NEVER_INSTRUMENT { |
| 307 | Allocators A(Max); |
| 308 | return A; |
| 309 | } |
| 310 | |
| 311 | static Allocators |
| 312 | InitAllocatorsFromBuffers(Allocators::Buffers &Bufs) XRAY_NEVER_INSTRUMENT { |
| 313 | Allocators A(Bufs); |
| 314 | return A; |
| 315 | } |
| 316 | |
| 317 | private: |
| 318 | NodeArray Nodes; |
| 319 | RootArray Roots; |
| 320 | ShadowStackArray ShadowStack; |
| 321 | NodeIdPairAllocatorType *NodeIdPairAllocator; |
| 322 | uint32_t OverflowedFunctions; |
| 323 | |
| 324 | public: |
| 325 | explicit FunctionCallTrie(const Allocators &A) XRAY_NEVER_INSTRUMENT |
| 326 | : Nodes(*A.NodeAllocator), |
| 327 | Roots(*A.RootAllocator), |
| 328 | ShadowStack(*A.ShadowStackAllocator), |
| 329 | NodeIdPairAllocator(A.NodeIdPairAllocator), |
| 330 | OverflowedFunctions(0) {} |
| 331 | |
| 332 | FunctionCallTrie() = delete; |
| 333 | FunctionCallTrie(const FunctionCallTrie &) = delete; |
| 334 | FunctionCallTrie &operator=(const FunctionCallTrie &) = delete; |
| 335 | |
| 336 | FunctionCallTrie(FunctionCallTrie &&O) XRAY_NEVER_INSTRUMENT |
| 337 | : Nodes(std::move(t&: O.Nodes)), |
| 338 | Roots(std::move(t&: O.Roots)), |
| 339 | ShadowStack(std::move(t&: O.ShadowStack)), |
| 340 | NodeIdPairAllocator(O.NodeIdPairAllocator), |
| 341 | OverflowedFunctions(O.OverflowedFunctions) {} |
| 342 | |
| 343 | FunctionCallTrie &operator=(FunctionCallTrie &&O) XRAY_NEVER_INSTRUMENT { |
| 344 | Nodes = std::move(t&: O.Nodes); |
| 345 | Roots = std::move(t&: O.Roots); |
| 346 | ShadowStack = std::move(t&: O.ShadowStack); |
| 347 | NodeIdPairAllocator = O.NodeIdPairAllocator; |
| 348 | OverflowedFunctions = O.OverflowedFunctions; |
| 349 | return *this; |
| 350 | } |
| 351 | |
| 352 | ~FunctionCallTrie() XRAY_NEVER_INSTRUMENT {} |
| 353 | |
| 354 | void enterFunction(const int32_t FId, uint64_t TSC, |
| 355 | uint16_t CPU) XRAY_NEVER_INSTRUMENT { |
| 356 | DCHECK_NE(FId, 0); |
| 357 | |
| 358 | // If we're already overflowed the function call stack, do not bother |
| 359 | // attempting to record any more function entries. |
| 360 | if (UNLIKELY(OverflowedFunctions)) { |
| 361 | ++OverflowedFunctions; |
| 362 | return; |
| 363 | } |
| 364 | |
| 365 | // If this is the first function we've encountered, we want to set up the |
| 366 | // node(s) and treat it as a root. |
| 367 | if (UNLIKELY(ShadowStack.empty())) { |
| 368 | auto *NewRoot = Nodes.AppendEmplace( |
| 369 | args: nullptr, args: NodeIdPairArray(*NodeIdPairAllocator), args: 0u, args: 0u, args: FId); |
| 370 | if (UNLIKELY(NewRoot == nullptr)) |
| 371 | return; |
| 372 | if (Roots.AppendEmplace(args&: NewRoot) == nullptr) { |
| 373 | Nodes.trim(Elements: 1); |
| 374 | return; |
| 375 | } |
| 376 | if (ShadowStack.AppendEmplace(args&: TSC, args&: NewRoot, args&: CPU) == nullptr) { |
| 377 | Nodes.trim(Elements: 1); |
| 378 | Roots.trim(Elements: 1); |
| 379 | ++OverflowedFunctions; |
| 380 | return; |
| 381 | } |
| 382 | return; |
| 383 | } |
| 384 | |
| 385 | // From this point on, we require that the stack is not empty. |
| 386 | DCHECK(!ShadowStack.empty()); |
| 387 | auto TopNode = ShadowStack.back().NodePtr; |
| 388 | DCHECK_NE(TopNode, nullptr); |
| 389 | |
| 390 | // If we've seen this callee before, then we access that node and place that |
| 391 | // on the top of the stack. |
| 392 | auto* Callee = TopNode->Callees.find_element( |
| 393 | P: [FId](const NodeIdPair &NR) { return NR.FId == FId; }); |
| 394 | if (Callee != nullptr) { |
| 395 | CHECK_NE(Callee->NodePtr, nullptr); |
| 396 | if (ShadowStack.AppendEmplace(args&: TSC, args&: Callee->NodePtr, args&: CPU) == nullptr) |
| 397 | ++OverflowedFunctions; |
| 398 | return; |
| 399 | } |
| 400 | |
| 401 | // This means we've never seen this stack before, create a new node here. |
| 402 | auto* NewNode = Nodes.AppendEmplace( |
| 403 | args&: TopNode, args: NodeIdPairArray(*NodeIdPairAllocator), args: 0u, args: 0u, args: FId); |
| 404 | if (UNLIKELY(NewNode == nullptr)) |
| 405 | return; |
| 406 | DCHECK_NE(NewNode, nullptr); |
| 407 | TopNode->Callees.AppendEmplace(args&: NewNode, args: FId); |
| 408 | if (ShadowStack.AppendEmplace(args&: TSC, args&: NewNode, args&: CPU) == nullptr) |
| 409 | ++OverflowedFunctions; |
| 410 | return; |
| 411 | } |
| 412 | |
| 413 | void exitFunction(int32_t FId, uint64_t TSC, |
| 414 | uint16_t CPU) XRAY_NEVER_INSTRUMENT { |
| 415 | // If we're exiting functions that have "overflowed" or don't fit into the |
| 416 | // stack due to allocator constraints, we then decrement that count first. |
| 417 | if (OverflowedFunctions) { |
| 418 | --OverflowedFunctions; |
| 419 | return; |
| 420 | } |
| 421 | |
| 422 | // When we exit a function, we look up the ShadowStack to see whether we've |
| 423 | // entered this function before. We do as little processing here as we can, |
| 424 | // since most of the hard work would have already been done at function |
| 425 | // entry. |
| 426 | uint64_t CumulativeTreeTime = 0; |
| 427 | |
| 428 | while (!ShadowStack.empty()) { |
| 429 | const auto &Top = ShadowStack.back(); |
| 430 | auto TopNode = Top.NodePtr; |
| 431 | DCHECK_NE(TopNode, nullptr); |
| 432 | |
| 433 | // We may encounter overflow on the TSC we're provided, which may end up |
| 434 | // being less than the TSC when we first entered the function. |
| 435 | // |
| 436 | // To get the accurate measurement of cycles, we need to check whether |
| 437 | // we've overflowed (TSC < Top.EntryTSC) and then account the difference |
| 438 | // between the entry TSC and the max for the TSC counter (max of uint64_t) |
| 439 | // then add the value of TSC. We can prove that the maximum delta we will |
| 440 | // get is at most the 64-bit unsigned value, since the difference between |
| 441 | // a TSC of 0 and a Top.EntryTSC of 1 is (numeric_limits<uint64_t>::max() |
| 442 | // - 1) + 1. |
| 443 | // |
| 444 | // NOTE: This assumes that TSCs are synchronised across CPUs. |
| 445 | // TODO: Count the number of times we've seen CPU migrations. |
| 446 | uint64_t LocalTime = |
| 447 | Top.EntryTSC > TSC |
| 448 | ? (std::numeric_limits<uint64_t>::max() - Top.EntryTSC) + TSC |
| 449 | : TSC - Top.EntryTSC; |
| 450 | TopNode->CallCount++; |
| 451 | TopNode->CumulativeLocalTime += LocalTime - CumulativeTreeTime; |
| 452 | CumulativeTreeTime += LocalTime; |
| 453 | ShadowStack.trim(Elements: 1); |
| 454 | |
| 455 | // TODO: Update the histogram for the node. |
| 456 | if (TopNode->FId == FId) |
| 457 | break; |
| 458 | } |
| 459 | } |
| 460 | |
| 461 | const RootArray &getRoots() const XRAY_NEVER_INSTRUMENT { return Roots; } |
| 462 | |
| 463 | // The deepCopyInto operation will update the provided FunctionCallTrie by |
| 464 | // re-creating the contents of this particular FunctionCallTrie in the other |
| 465 | // FunctionCallTrie. It will do this using a Depth First Traversal from the |
| 466 | // roots, and while doing so recreating the traversal in the provided |
| 467 | // FunctionCallTrie. |
| 468 | // |
| 469 | // This operation will *not* destroy the state in `O`, and thus may cause some |
| 470 | // duplicate entries in `O` if it is not empty. |
| 471 | // |
| 472 | // This function is *not* thread-safe, and may require external |
| 473 | // synchronisation of both "this" and |O|. |
| 474 | // |
| 475 | // This function must *not* be called with a non-empty FunctionCallTrie |O|. |
| 476 | void deepCopyInto(FunctionCallTrie &O) const XRAY_NEVER_INSTRUMENT { |
| 477 | DCHECK(O.getRoots().empty()); |
| 478 | |
| 479 | // We then push the root into a stack, to use as the parent marker for new |
| 480 | // nodes we push in as we're traversing depth-first down the call tree. |
| 481 | struct NodeAndParent { |
| 482 | FunctionCallTrie::Node *Node; |
| 483 | FunctionCallTrie::Node *NewNode; |
| 484 | }; |
| 485 | using Stack = Array<NodeAndParent>; |
| 486 | |
| 487 | typename Stack::AllocatorType StackAllocator( |
| 488 | profilingFlags()->stack_allocator_max); |
| 489 | Stack DFSStack(StackAllocator); |
| 490 | |
| 491 | for (const auto Root : getRoots()) { |
| 492 | // Add a node in O for this root. |
| 493 | auto NewRoot = O.Nodes.AppendEmplace( |
| 494 | args: nullptr, args: NodeIdPairArray(*O.NodeIdPairAllocator), args&: Root->CallCount, |
| 495 | args&: Root->CumulativeLocalTime, args&: Root->FId); |
| 496 | |
| 497 | // Because we cannot allocate more memory we should bail out right away. |
| 498 | if (UNLIKELY(NewRoot == nullptr)) |
| 499 | return; |
| 500 | |
| 501 | if (UNLIKELY(O.Roots.Append(NewRoot) == nullptr)) |
| 502 | return; |
| 503 | |
| 504 | // TODO: Figure out what to do if we fail to allocate any more stack |
| 505 | // space. Maybe warn or report once? |
| 506 | if (DFSStack.AppendEmplace(args: Root, args&: NewRoot) == nullptr) |
| 507 | return; |
| 508 | while (!DFSStack.empty()) { |
| 509 | NodeAndParent NP = DFSStack.back(); |
| 510 | DCHECK_NE(NP.Node, nullptr); |
| 511 | DCHECK_NE(NP.NewNode, nullptr); |
| 512 | DFSStack.trim(Elements: 1); |
| 513 | for (const auto Callee : NP.Node->Callees) { |
| 514 | auto NewNode = O.Nodes.AppendEmplace( |
| 515 | args&: NP.NewNode, args: NodeIdPairArray(*O.NodeIdPairAllocator), |
| 516 | args&: Callee.NodePtr->CallCount, args&: Callee.NodePtr->CumulativeLocalTime, |
| 517 | args: Callee.FId); |
| 518 | if (UNLIKELY(NewNode == nullptr)) |
| 519 | return; |
| 520 | if (UNLIKELY(NP.NewNode->Callees.AppendEmplace(NewNode, Callee.FId) == |
| 521 | nullptr)) |
| 522 | return; |
| 523 | if (UNLIKELY(DFSStack.AppendEmplace(Callee.NodePtr, NewNode) == |
| 524 | nullptr)) |
| 525 | return; |
| 526 | } |
| 527 | } |
| 528 | } |
| 529 | } |
| 530 | |
| 531 | // The mergeInto operation will update the provided FunctionCallTrie by |
| 532 | // traversing the current trie's roots and updating (i.e. merging) the data in |
| 533 | // the nodes with the data in the target's nodes. If the node doesn't exist in |
| 534 | // the provided trie, we add a new one in the right position, and inherit the |
| 535 | // data from the original (current) trie, along with all its callees. |
| 536 | // |
| 537 | // This function is *not* thread-safe, and may require external |
| 538 | // synchronisation of both "this" and |O|. |
| 539 | void mergeInto(FunctionCallTrie &O) const XRAY_NEVER_INSTRUMENT { |
| 540 | struct NodeAndTarget { |
| 541 | FunctionCallTrie::Node *OrigNode; |
| 542 | FunctionCallTrie::Node *TargetNode; |
| 543 | }; |
| 544 | using Stack = Array<NodeAndTarget>; |
| 545 | typename Stack::AllocatorType StackAllocator( |
| 546 | profilingFlags()->stack_allocator_max); |
| 547 | Stack DFSStack(StackAllocator); |
| 548 | |
| 549 | for (const auto Root : getRoots()) { |
| 550 | Node *TargetRoot = nullptr; |
| 551 | auto R = O.Roots.find_element( |
| 552 | P: [&](const Node *Node) { return Node->FId == Root->FId; }); |
| 553 | if (R == nullptr) { |
| 554 | TargetRoot = O.Nodes.AppendEmplace( |
| 555 | args: nullptr, args: NodeIdPairArray(*O.NodeIdPairAllocator), args: 0u, args: 0u, |
| 556 | args&: Root->FId); |
| 557 | if (UNLIKELY(TargetRoot == nullptr)) |
| 558 | return; |
| 559 | |
| 560 | O.Roots.Append(E: TargetRoot); |
| 561 | } else { |
| 562 | TargetRoot = *R; |
| 563 | } |
| 564 | |
| 565 | DFSStack.AppendEmplace(args: Root, args&: TargetRoot); |
| 566 | while (!DFSStack.empty()) { |
| 567 | NodeAndTarget NT = DFSStack.back(); |
| 568 | DCHECK_NE(NT.OrigNode, nullptr); |
| 569 | DCHECK_NE(NT.TargetNode, nullptr); |
| 570 | DFSStack.trim(Elements: 1); |
| 571 | // TODO: Update the histogram as well when we have it ready. |
| 572 | NT.TargetNode->CallCount += NT.OrigNode->CallCount; |
| 573 | NT.TargetNode->CumulativeLocalTime += NT.OrigNode->CumulativeLocalTime; |
| 574 | for (const auto Callee : NT.OrigNode->Callees) { |
| 575 | auto TargetCallee = NT.TargetNode->Callees.find_element( |
| 576 | P: [&](const FunctionCallTrie::NodeIdPair &C) { |
| 577 | return C.FId == Callee.FId; |
| 578 | }); |
| 579 | if (TargetCallee == nullptr) { |
| 580 | auto NewTargetNode = O.Nodes.AppendEmplace( |
| 581 | args&: NT.TargetNode, args: NodeIdPairArray(*O.NodeIdPairAllocator), args: 0u, args: 0u, |
| 582 | args: Callee.FId); |
| 583 | |
| 584 | if (UNLIKELY(NewTargetNode == nullptr)) |
| 585 | return; |
| 586 | |
| 587 | TargetCallee = |
| 588 | NT.TargetNode->Callees.AppendEmplace(args&: NewTargetNode, args: Callee.FId); |
| 589 | } |
| 590 | DFSStack.AppendEmplace(args: Callee.NodePtr, args&: TargetCallee->NodePtr); |
| 591 | } |
| 592 | } |
| 593 | } |
| 594 | } |
| 595 | }; |
| 596 | |
| 597 | } // namespace __xray |
| 598 | |
| 599 | #endif // XRAY_FUNCTION_CALL_TRIE_H |
| 600 | |