Repository navigation
Conversation
Without the files field, npm falls back to .gitignore, which ignores java-binaries, so the published package shipped without the weaver JARs required by code/WeaverConfiguration.ts. Restore the allowlist, updated for the vitest/oxlint migration.
Load clang-dumper assets from the v16.0.5_3 release manifest and remove the legacy archive resource map. Update Clang 18 enum values, parser tolerance, bundled include/CUDA handling, and JS expectations needed by the new dumper.
Use entrypoints.txt from extracted clang-dumper include archives as the ordered system include roots. Throw when entrypoints.txt is missing instead of inferring roots from archive folder names.
Allow parser tests to select a local dumper without disturbing release cache assets. Restore strict resolution of referenced nodes so incomplete dumper output fails at its source.
Share the system-header threshold spelling between Clava's tool and plugin invocation paths.
Treat absolute tag-file values as build directories, use their local tool or plugin without release downloads or checksum validation, and force system-provided headers. Remove the superseded executable override and cover source parsing and local tool selection.
Add support for Clang's CXXParenListInitExpr node and preserve the original parenthesized argument list during code generation. Add an integration test covering aggregate initialization under C++20.
Key the in-process lock by the shared release cache lock file so concurrent libc and CUDA configurations cannot acquire different JVM locks for the same cache folder.
Store per-version cache locks beside the version directory so cleanup can delete the directory while holding the lock on Windows, Linux, and macOS. Update the cache cleanup test to use the shared lock-path helper.
Discover a system CUDA toolkit from portable environment, PATH, and platform-specific locations when bundled headers are unavailable. Continue with a single warning when no toolkit is found instead of rejecting local builds.
Coordinate cache initialization and stale cleanup with a temporary directory claim that is removed on release. Keep the claim outside version folders for Windows deletion and recover claims left by terminated processes.
Write the shared CUDA compatibility resource to a unique temporary file and install it with an atomic move, preventing concurrent Clava instances from reading a partial header.
Restore callinit and listinit as explicit enum values because the generated TypeScript API exposes raw strings without an enum conversion.
Add same-JVM, cross-JVM, multi-version, crash-recovery, stale-cleanup, and real release initialization coverage for the clang-dumper cache. Keep red diagnostic tests for reclaimed-lock release and cached last-used updates bypassing the cache lock.
Add a cross-JVM diagnostic proving stale cleanup can remove an executable still referenced by the in-process ClangFiles cache.
Use the immutable testing release with canonical macOS framework and subframework include paths. Preserve the local developer selector override unstaged.
Use the immutable testing release with portable resource-package fixtures and canonical framework includes. Preserve the developer selector override unstaged.
Skip built-in CUDA fixtures only when the Java resource installer reports no compatible host platform, and allow bounded cold resource setup. Record the RC8 nightly failures and the focused local CUDA/typecheck results without treating the prior CI run as green.
The setter integration uses high_resolution_clock, which the Windows standard library reports as steady_clock. Reuse the existing matching macOS output under a platform-neutral fixture name; all other expected output remains identical. Record the pre-fix CI result and leave fresh Windows verification pending.
Archive the validated 48-row Protobuf-only runtime cohort, separate 120-cycle memory diagnostics, exact drivers, RC8 asset provenance, raw sidecars, and the zero-observation preflight attempt. Point the report instructions at the final capture.
Archive resource-only timing and process traces comparing released RC8 consumer setup with the reviewed permission fix. Preserve historical suite measurements and document remaining release/toolchain differences without claiming full-suite runtime recovery.
Add isolated snapshot preparation, pinned runtime and memory collectors, strict saved-evidence validation, and private DraftLink retry support. Include cohort fixtures, CLI documentation, and fixture-based workflow and report tests. Preserve the frozen non-Protobuf controls.
Preserve the zero-observation JavaScript cohort preflight failure with its full logs and provenance. Archive separate 120-cycle RC8 memory diagnostics with source and executable hashes plus raw repeats. Keep the private report unchanged because runtime evidence did not validate.
Keep the accepted JavaScript test bytes in consumer-owned fixtures and verify every current source hash before staging them into the isolated Clava snapshot. Record current and staged per-file hashes alongside the two source revisions, validate them in capture and report paths, and show the overlay in the generated report.
Disable Python bytecode writes before the refresh workflow imports report helpers from the canonical Clava checkout. This prevents preparation from creating untracked cache files that trigger the checkout cleanliness gate.
Stage pinned JavaScript tests only inside a detached snapshot worktree and reject symlink paths, unknown source bytes, and post-stage changes. Require workload provenance in standalone reports and validate it against runtime evidence. Use Vitest's runner config loader for both phases, record the selected loader, and keep collectors from writing bytecode into canonical or snapshot source trees.
Prevent frozen JavaScript workload staging from changing files through hard links, and require standalone reports to carry every pinned test-source hash. Add regressions for canonical file preservation and missing or contradictory source manifests.
The protobuf path paid three costs the text path did not: - Every JVM re-hashed the cached 77 MB clang-dumper executable (~200 ms on the first App call), although CacheFiles.installFile already verifies the SHA-256 at download and publishes atomically. ccache's own compiler check covers the executable identity, so the ccache namespace now keys only on schema and toolchain. - ProtoAstReader kept several per-node HashSets to re-check producer invariants (dense-id exactness, NodeClass/Children coverage, End counters and raw bytes). Duplicates are still rejected through the existing maps; Header, End presence, payload contracts and top-level references are kept. - Opt-in wire metrics (PROTOBUF_METRIC, four ClangAstData timing keys) were threaded through the hot path. JFR covers this need. nas_bt.c steady-state App time drops ~6% (1117 -> 1049 ms) and the first App call in a JVM ~200 ms (835 -> 636 ms) in an in-JVM microbenchmark.
refresh_benchmark.py gains 'run' (prepare, collect, render and optional private publish in one call) and 'quick' (App phase only, direct and warm, one repeat, printed beside the frozen control medians). Workspace, release asset and resource cache paths now default to this machine's verified RC8 inputs. quick removes its snapshot worktrees and multi-gigabyte cache trees when done.
The parser no longer adds the dumper executable digest to its ccache namespace, so the warm-cache gate looked in a folder that is never written.
Write dumps uncompressed and set CCACHE_COMPRESS, the same cache policy as the FlatBuffers consumer, instead of streaming a Zstandard frame from clang-dumper and decoding it with zstd-jni. An in-JVM A/B on boolean.c, includes.cpp and NAS BT showed no difference between the two policies in cold or warm mode (within 1%), so the simpler one wins: it drops the zstd-jni dependency and the .zst dump path, and makes the two wire prototypes' cache paths identical for comparison.
three_way.py snapshots each transport's committed Clava ref, builds all of them against one shared specs-java-libs and lara-framework revision, and runs the rotated App and wall matrix in a single session with identical pinned JavaScript tests and per-call workload checks. render_comparison reuses the existing report charts with every transport from the same session and Before cache kept as the labelled historical baseline.
Every 'gradle test' re-ran the release resolver, its Python unit tests and the binding drift check (~2 s per Java suite command). Like the FlatBuffers build, run the drift and resolver tests under 'check' only, and let resolveProtoRelease be up to date when its tag, script and outputs are unchanged.
The FlatBuffers branch now targets Java 21; the overlay already compiles for the parser's own class version.
Record the validated RC8 runtime capture, separate fresh memory diagnostics, source provenance, corrected report history, and compressed raw sidecars.
Check pinned test counts for App and wall observations before rendering comparison rows. Add a regression test for failed App tests and clarify that differing revisions, native releases, header sources, and cache settings prevent isolating AST transport.
|
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Clava generated Protobuf bindings from a sibling schema checkout that could disagree with the selected dumper. Generation now uses the verified canonical schema and descriptor from
clang-dumper-release.tag, validates release toolchain compatibility before compilation, and checks consumer binding drift incheckand CI. Native Protobuf/protoc 28.3 and Java Protobuf 4.28.3 remain pinned.Compiled DataKeys, inheritance and checked conventions govern generated bindings. Unsupported emitted nodes fail with their class, ID and available location regardless of reachability; payload compatibility is checked against the actual class. Structured metadata and printers support the affected assembly, lambda, template, type and declaration edits. Unsupported dependent-friend forms reject explicitly. Bounded incremental framing/decompression, presence/reference checks, direct getters, the reader cache and external Zstd policy are preserved.
Cleanup covers syntax-only, rejection, cross-TU processing and repeated parsing. Resource selection, meaningful libc probes, framework search and bounded metadata caches use the selected schema, toolchain and verified executable. CUDA tests query support only when executed, and platform clock fixtures retain the setter transformation assertions. Packed Node 24 local/global CLI and API checks pass; npm 11 java-addon script approval is documented.
Validation:
v18.1.8_6-rc8passed all native platform jobs. Clean release-backed parsercheck: 253 tests, 249 passed, zero failures, four local-binary-only skips. All three release-specific checks, 10 resolver tests, binding drift, coverage andinstallDistpassed.360a41c5epassed Java and all six Node 24/25 Linux/macOS/Windows jobs. The Windows Node 24 runner connection loss and successful retry are retained as evidence.experiments/protobuf/validation/evidence/benchmark-20261008-rc8/.The local developer selector override is deliberately uncommitted. This PR is for review and RC testing, with no stable publication or master merge requested.