mirror of
https://github.com/abhigyanpatwari/GitNexus.git
synced 2026-09-11 22:53:04 +00:00
2 commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Gergő Magyar
|
95f87fc12a
|
perf(ingestion): Linux-kernel-scale analysis — worker-pool parse + finalize O(n²) + scope-resolution memory wall (#1983) (#2038)
* fix(ingestion): reduce parse-phase memory for huge repos (#1983)
Stop retaining full parse-cache chunks in RAM alongside the merged graph,
slim on-disk shards, defer worker ParsedFile emission for scope-resolver
languages, and add GITNEXUS_DEBUG_HEAP probes for OOM diagnosis.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(ingestion): address #2038 tri-review findings (parse-phase memory)
Resolves the confirmed review findings on PR #2038:
- P1: thread exportedTypeMap through the sequential parse path
(processParsingSequential) so a no-worker run over a partially-warm
cache no longer silently drops the sequential-miss files' exported
types. Cache hits made exportedTypeMap.size > 0, suppressing the
end-of-loop buildExportedTypeMapFromGraph rebuild, but the sequential
path never populated the map. Regression test added (fails on the
pre-fix tree, passes after) plus a fully-sequential differential oracle.
- P2: saveParseCache builds its on-disk index from hashes actually
written/copied (writtenKeys), never a usedKeys hash whose shard write
or copy was skipped — no more phantom index entries.
- P2: add a unit test asserting SCOPE_RESOLUTION_LANGUAGES stays in sync
with SCOPE_RESOLVERS (asymmetric drift would lose a language's ParsedFile).
- Backfill cache coverage: loadParseCacheChunk missing/corrupt -> undefined,
pruneCache onDiskKeys branch, slim preserves nodes, saveParseCache
copy-evicted-shard round-trip.
- Cleanups: single-source heap-probe gating via isDebugHeapEnabled();
hoist the per-chunk mkdir in persistParseCacheChunk behind a
process-scoped Set; gate COBOL's unused worker-side ParsedFile
extraction (graph nodes still come from cobolPhase) while keeping
fileCount/progress unconditional.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor(ingestion): remove dead worker-side ParsedFile extraction
After #2038 gated worker `ParsedFile` emission behind `!isScopeResolutionLanguage(language)`, and with all 16 SupportedLanguages registered in SCOPE_RESOLVERS, that gate was structurally always true — the worker already produced no ParsedFiles and scope-resolution re-extracts each file from source on the main thread (run.ts). Remove the now-dead machinery:
- Drop both worker `extractParsedFile` call-sites (tree-sitter processFileGroup + the standalone-provider branch) and the `result.parsedFiles.push`. The standalone branch keeps fileCount/onFileProcessed per file. `result.parsedFiles` stays declared but empty (field removal deferred).
- Remove the now-orphaned `scopeSourceKind` var + `ScopeCaptureSourceKind`/`extractParsedFile`/`isScopeResolutionLanguage` imports.
- Delete the consumerless `migrated-languages.ts` (isScopeResolutionLanguage + SCOPE_RESOLUTION_LANGUAGES) and its drift-guard test — parse-worker was their only importer. Also improves AGENTS.md "shared ingestion code must not name languages" compliance.
`extractParsedFile` and the scope-extractor-bridge stay (scope-resolution/run.ts + Vue resolver use them). Behavior-preserving: worker-sequential-parity passes before and after; tsc/eslint clean; no baseline/golden drift.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor(ingestion): worker-pool-only parsing; remove sequential parser (#1983)
Completes the #1983 huge-repo parse-OOM effort by making the worker pool
GitNexus's sole parse path.
Parallel serialization (the perf core): workers serialize their ParsedFiles to
a disk store in parallel and stream them back to scope-resolution, so the main
thread no longer re-parses every file (the tree-sitter native-memory leak that
caused the OOM). Adds chunk merge-pipelining + work-proportional chunk sizing so
the pool stays saturated.
Remove the sequential parser: `--workers 0`, `GITNEXUS_WORKER_POOL_SIZE=0`, and
`skipWorkers` now hard-error (no silent degrade — #1741); the small-repo
threshold no longer selects an in-process path; pool creation stays lazy /
cache-miss-gated so warm all-hit runs never spawn workers.
Worker-path parity fixes — removing sequential surfaced two pre-existing gaps
that tiny-fixture tests had masked by running below the worker threshold, both
fixed by carrying per-file metadata as DATA across the worker boundary (never
re-parsing on the main thread, preserving the OOM fix):
- C++: templateConstraints wired into worker node identity (SFINAE overload
disambiguation) + ADL / inline-namespace capture side-channel serialized
onto the ParsedFile.
- Kotlin: companion-scope side-channel serialized the same way (companion /
static dispatch).
Validation: tsc + build clean; full suite green (10,190 pass — the only
deterministic failures were the now-fixed C++/Kotlin worker-path gaps; the 2
remaining full-run failures are pre-existing load flakiness, green in
isolation); cpp-pipeline benchmark stays linear on a 1-worker pool.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(ingestion): wire C static-linkage side-channel + ADL O(1) collect + tri-review cleanups (#1983)
Follow-up to the worker-pool-only refactor, from a tri-review of the parse path.
- C static-linkage side-channel (P1): cProvider had no collect/applyCaptureSideChannel,
so on the now-sole worker path C `static` file-local marks were lost across the worker
boundary -> false cross-file CALLS edges + over-broad #include wildcard visibility on
every C analysis (the Linux kernel is C). Mirror the C++/Kotlin wiring: serialize
`staticNames` per file onto ParsedFile.captureSideChannel and restore it on the main
thread (no re-parse). + a worker-path regression test (the existing c-static-isolation
fixture passed vacuously — its collision resolves via #include before the global
free-call fallback ever consults static-linkage).
- captureSideChannel `kind` discriminant: add `kind:'cpp'`/`kind:'c'` tags + guards
(Kotlin already had one) now that C/C++/Kotlin share the single generic field.
- Perf: collectCppAdlSideChannel scanned the whole argInfoBySite/noAdlSites maps per file
(O(F^2) per sub-batch, ~100M parseSiteKey calls at kernel scale). Add per-filePath
lockstep indexes -> O(1) collect; serialized snapshot byte-identical.
- Cleanups: inline the one-line processParsingWithWorkers wrapper into processParsing;
drop the always-empty WorkerExtractedData.calls/assignments/constructorBindings fields;
remove the voided astCache param from processParsing; refresh stale "sequential
fallback" JSDoc.
Validation: tsc + build clean; cpp 297/297, c 8/8 (incl. the new worker-path
static-linkage guard), typescript + parsedfile-store green; cpp ADL benchmark stays linear.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* perf(scope-resolution): index C/C++ #include resolution in finalize (O(n²)→O(n))
Kernel-scale C/C++ analysis ground in finalizeScopeModel because three
per-#include operations each did a full O(F) scan with no index — the
finalize O(n²) that surfaced once the #1983 parse-phase OOM was fixed:
- expand{C,Cpp}WildcardNames: parsedFiles.find() per wildcard edge → O(R·F)
- resolveImportTarget: new Set(allFilePaths) rebuilt per #include
- resolveCImportTarget: suffix-match scanned all workspace paths
Each is replaced with a WeakMap-per-pass index keyed on the stable
parsedFiles/allFilePaths references that scope-resolution run.ts passes
once per pass:
- Map<ScopeId,ParsedFile> for wildcard expansion (c/static-linkage.ts +
cpp/file-local-linkage.ts)
- memoized augmented header set (c/scope-resolver.ts + cpp/scope-resolver.ts)
- basename-bucketed suffix index in resolveCImportTarget (c/import-target.ts),
shared by C and C++ since resolveCppImportTarget delegates to it
Collapses the C/C++ finalize from O(R·F) to O(R+F). Pure-perf, byte-identical
edge output: 962 targeted tests green (490 C + 472 C/C++ scope-resolution);
the basename index preserves the exact endsWith('/'+target) match and the
fewest-path-components-then-lexicographic tie-break.
The kernel's ~25-30k .h headers are classified C++, so both providers must
be fixed. Proven on the Linux kernel: the C finalize completed
(sr-post-finalize lang=c → sr-end lang=c), which the pre-fix run never
reached in 16+ min of grinding.
Build-independent follow-ups (separate from this finalize fix), documented
for later: emitFreeCallFallback same-name buckets (emit phase),
buildGraphNodeLookup + precount global setup, the ParsedFile store-load,
the dart/go/ruby expand-wildcards .find siblings, and the ~26GB
scope-resolution memory floor (full kernel completion needs >~40GB RAM).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(bench): regenerate C scope-capture baseline for the #1983 c-static-linkage-worker fixture
bench/scope-capture/measure.mjs fingerprints emitCScopeCaptures over the
lang-resolution/c-* fixture corpus. The #1983 PR added the
c-static-linkage-worker fixture (caller.c/lib.c/lib.h/local.c — the
worker-path static-linkage side-channel test) but did not regenerate the C
baseline, so `--check` has been red on this branch (main, lacking the
fixture, still matches 0de009b).
Pure fixture-corpus drift — no c/captures.ts or query change branch-vs-main,
existing fixtures' captures byte-identical (c-captures.test.ts 45/45),
scaling stays linear (~0.97). Regenerated: 0de009b -> 39f3a83. Bench now
PASS (14 languages). Unrelated to the finalize O(n²) fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* perf(scope-resolution): lower kernel-scale resident memory floor + setup cost
Reduce the scope-resolution resident-memory floor and setup throughput on
huge repos (Linux kernel), the wall that remains after #1983 (parse OOM) and
the finalize O(n^2) fix (
|
||
|
Gergő Magyar
|
f1b8438388
|
perf(cpp): index ADL candidates once instead of per-site rescans (#1990)
* perf(cpp): index ADL candidates once instead of per-site rescans C++ scope-resolution `emit` dominated large-repo analysis (~6.76h on a 5,969-file repo — ~70% of the total run). `pickCppAdlCandidates` ran once per unresolved ADL-eligible call site and each time: - rescanned every parsed file (rebuilding a per-file scope map per call), - scanned every workspace def (`findCppClassDefBySimpleName`), and - used an O(scopes²) child-scope walk for hidden friends. That is O(unresolved sites × files); with hundreds of thousands of unresolved C++ sites the emit phase went super-linear. `resolve` (registry lookup) was only 3.5s — the cost was entirely in fallback edge emission. Build an `AdlCandidateIndex` once per run (lazy, guarded by `parsedFiles` identity, reset in `clearCppAdlState`) and query it per site: - `classDefsBySimple` — preserves `defs.byId` order so first-match / ambiguous semantics are identical to the legacy linear scan. - `nsCandidates` — namespace-owned callables, with inline-namespace transparency. - `friendCandidates` — hidden-friend + class-member callables; a parent→children scope index replaces the O(scopes²) walk. - `nsFunctionsByQName` / `nsFunctionsBySimple` — function-reference ADL path. A monotonic `seqByNodeId` (file-major; namespace defs before friend/member defs within a file) lets the per-site query merge candidates across associated namespaces, dedup by nodeId, and sort — reproducing the exact legacy candidate set and order. Per-site cost drops from O(sites × files) to O(associated namespaces); the emit phase goes from linear-in-sites to flat. Benchmark (files=80): emit at 1000 sites 232ms → 9ms, 2000 sites flat at 17ms; the eliminated term scales with file count, so the speedup is ~1000×+ on the real 5,969-file repo. Behavior is unchanged: synthetic candidate output is byte-identical before/after, all 270 C++ integration resolver tests and 4/4 resolver-parity-expected-failures pass, and tsc + eslint are clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(cpp): correct ADL state-lifecycle and cache-guard comments The header lifecycle block listed three module-level maps and named clearFileLocalNames as the reset caller; both became inaccurate when the candidate index was added. Enumerate all five state pieces, name the real caller (loadResolutionConfig), and document that ensureAdlIndex's staleness guard keys on parsedFiles identity while the index also depends on scopes and classToNamespaceQualifiedName. Addresses PR #1990 tri-review (U1, U3). Doc-only; no behavior change. * test(cpp): guard the ADL seq-coverage invariant in dev/test pickCppAdlCandidates sorts merged candidates by seqByNodeId with a `?? 0` fallback. That fallback is unreachable today (every bucketed def is seq-assigned in the same build block), but a future regression could break it and silently collapse two seq-0 candidates, dropping a CALLS edge with no error. Add validateAdlSeqCoverage and run it from buildAdlIndex under the resolver's opt-in validation gate (NODE_ENV!=production && VALIDATE_SEMANTIC_MODEL!=0), so a broken invariant throws loudly in dev/CI instead. Production behavior and the hot path are unchanged. Unit-tested; 270/270 cpp integration tests pass with the guard active. Addresses PR #1990 tri-review (U2). * test(cpp): parity fixture for ADL hidden-friend + namespace-callable merge pickCppAdlCandidates merges friendCandidates (hidden friends of associated classes) and nsCandidates (namespace-owned callables) for a single associated namespace. The byte-identical-parity claim rested only on an uncommitted harness. Add a fixture that reaches one callable through each bucket — combine only via a hidden friend, process only via a namespace member — so dropping either bucket from the merge fails the suite. Candidate order is not observable (narrowing resolves a unique survivor or suppresses), so the guard is on the set. Addresses PR #1990 tri-review (U4). * test(cpp): add ADL emit-scaling benchmark Guards the PR #1990 optimization against reintroducing the O(sites x files) ADL candidate scan. Generates many UNRESOLVED ADL sites (class-typed arg + a callee declared nowhere) and co-scales files and sites with N, so the old cost is O(N^2) and the new cost O(N). Isolates the scope-resolution emit ms from parse-dominated wall time via the logger test destination (capture verified) and asserts the end-to-end emit ratio stays under fileRatio^1.5. Gated by GITNEXUS_BENCH=1; runs build-free (workerPoolSize: 0). Addresses the benchmark request alongside PR #1990 (U5). * test(cpp): add cpp pipeline file-count benchmark Fills the one missing per-language pipeline benchmark (cobol/csharp/go/php/ ruby/rust already have one); modeled on cobol-pipeline-benchmark.test.ts. Generates synthetic C++ with constant per-file work and constant header fan-out, sweeps file count through the full pipeline, and guards linearity with a coarse time-ratio bound plus a deterministic node-ratio bound (the non-flaky guard against reintroducing O(fileCount^2) work). Gated by GITNEXUS_BENCH=1; runs build-free (workerPoolSize: 0). Addresses the benchmark request alongside PR #1990 (U6). * style(cpp): prettier-format adl benchmark * test(cpp): rebaseline scope-capture fingerprint for new ADL fixture The U4 parity fixture (cpp-adl-ns-plus-hidden-friend-same-name) lives under test/fixtures/lang-resolution/cpp-*, so its lib.h + app.cpp join the cpp scope-capture bench corpus (bench/scope-capture/measure.mjs). That is pure fixture-corpus growth — no scope-extractor change, existing fixtures' captures byte-identical — so the cpp fingerprint legitimately drifts (fixture_count 265->267). Rebaseline cpp to match, as #1965/#1975 did for earlier fixture additions. Verified: --check PASS for all 14 languages. Addresses PR #1990 tri-review (U4 follow-on). --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |