Repository navigation
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
⏭️ Performance benchmarks were skippedPerf benchmarks (CodSpeed) are opt-in on pull requests — they no longer run on every push. They always run automatically after merge to To run them on this PR, do any of the following, then push a commit (or re-run CI):
|
Binary size checks passed✅ 7 passed
Generated by |
This was referenced Jul 21, 2026
…roject load; parallel LSP sweep Three fixes for the "12 threads = 1.18x" cold-check scaling wall, diagnosed by comparing against ty/ruff's salsa usage (whose driver pattern ours already matches): - Prime wave: compute every file's PPIR semantic index across workers BEFORE fanning out per-file checks. Whole-package aggregate queries (package_items / namespace_items) fold over every file's index, and every file's check demands them - on a cold database the first worker to claim the aggregate memo computed parse+lowering+indexing for the entire project inline on one thread while all other workers parked on that memo's sync slot. With per-file indexes primed the aggregate fold is cheap and workers stay busy. Also replaces the old first-file serial warm-up, which warmed the builtin package (compiler2_all_files orders builtins first), not the user package. - Bulk project load: add_or_update_files registers all sources with ONE project-file-list write; the per-file path cloned and re-set the whole files Vec and bumped the salsa revision once per file (O(files^2) copies during initial load). Source reads now happen across worker threads, preserving discovery order so FileId assignment and diagnostic ordering are unchanged. - The LSP diagnostics candidate now checks files through the shared order-preserving parallel checker (under the held source guard - the same exclusion the serial loop relied on), so every per-edit path inherits the parallelism and the prime wave. check_files_parallel returns per-file results in input order so the LSP can zip them back to documents; the flat collector delegates. Measured (Apple Silicon, 12 cores): - cold pure check, 295k LOC: 7.4s -> 2.8s wall (1.2 -> 2.4+ cores) - cold check+seed, 295k LOC: 10.5s -> 4.4s; 59k LOC: ~0.9s - LSP edit->diagnostics: 250 files 464 -> 172ms; 500 files 1.7s -> 176ms; 1000 files 5.5s -> 209ms (now ~flat in project size) - LSP load test: burst storms + sustained churn at 250/1000 files, zero crashes, stable RSS, convergence after every storm Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01V4oee6wrCHzXH6mM2h8eJo
hellovai
force-pushed
the
hellovai/compiler-parallel-load-lsp
branch
from
July 22, 2026 03:17
c2da018 to
f214619
Compare
Contributor
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
Author
meefs
pushed a commit
to meefs/baml
that referenced
this pull request
Jul 23, 2026
…itor latency (BoundaryML#4105) Consolidated stack (supersedes BoundaryML#4098, BoundaryML#4100, BoundaryML#4101 — their commits are all contained here, rebased onto current canary). Seven commits, reviewable independently, in dependency order. ## What this does A ground-up performance pass on the compiler driven by head-to-head benchmarking against tsgo (native TypeScript) and `go build` on semantically identical generated corpora (1.5k–295k LOC), followed by root-cause profiling of every gap found. ### 1. `perf(compiler2)`: kill the two cold-compile quadratics; parallelize the cache-path check - The package-wide Jinja type environment was rebuilt inside every `check_file` call (O(N²)) — and twice per compile via `get_bytecode`'s error gate. Now a memoized salsa query (`jinja_base_types`) + skipped entirely for files with no prompts. - `ProjectDatabase` handle clones deep-copied two N-entry PathBuf maps; the parallel drivers clone ~1.55N handles per compile (quadratic CPU + peak RSS). Now Arc-wrapped, O(1). - `baml check`/`run` routed through a serial diagnostics collector while the parallel one sat unreachable next to it. Now shared. - The LSP proves the project error-free once (`get_bytecode_unchecked`), and its diagnostics candidate now includes package-level diagnostics — cross-file name conflicts previously **never reached the editor at all**. ### 2. `feat(cli)`: `baml pack` uses the bytecode cache; `baml check` seeds it Pack was fully cold on every invocation (no cache participation at all); check reused but never wrote, so check-only workflows never warmed. Now one warm cache serves check → run → test → pack. ### 3–4. `build` + `fix(ci)`: hardware SHA-2 (`sha2/asm`, gated off MSVC), doc-link fix The warm path hashes ~hundreds of MB of unit payloads per run; it was on the software SHA fallback. ### 5. `perf(compiler2)`: prime wave, bulk project load, parallel LSP sweep Cold parallel check scaled 1.18× on 12 cores: every file's check demands whole-package aggregates, and the first worker to claim that memo computed parse+index for the entire project inline while all other workers parked on the sync slot (diagnosed against ty/ruff's salsa usage; interventionally verified). Per-file indexes are now primed across workers before fan-out. Project load stops re-writing the file list per file (O(files²)) and reads sources in parallel (FileId order preserved). The LSP candidate uses the shared order-preserving parallel checker. ### 6. `perf(cache)`: warm no-op checks stop re-proving the cache A warm no-op check cost the same as a full recompute: the throws-gate safety compare re-parsed every file, and seeding read+hashed 78MB of unit payloads for metadata. Now: the gate is skipped when the dirty partition is provably empty (a value compared against itself), throws fragments live verbatim in the manifest so a no-op check opens **zero** unit payloads, and units load in parallel (manifest order preserved) only when a delta means relink will need them. Both verify oracles (`BAML_CACHE_VERIFY` full compare + always-on 1-in-32 sampled re-derive) now check the manifest copy — exactly what is served. Design was adversarially reviewed; the unsound variant (fact-space throws compare) was rejected and every real edit still takes the full runtime-space gate. ### 7. `refactor(cli)`: one shared `ProjectSession` for every project-loading command check/run/test/pack/describe/grep/generate each hand-rolled a drifting subset of the same preamble — the root cause of "pack forgot the cache" and "describe never primes." One implementation; commands declare `CacheUse::{ReadWrite, ReadWriteTests, ReadOnly, Off}`. Introspection installs seeds only on a no-delta plan (provable tautology) and derives honestly otherwise. ## Measured impact (Apple Silicon, 12 cores, hyperfine) | | before | after | |---|--:|--:| | cold check, 295k LOC | 21–46 s (126 s CPU, 5.3 GB RSS) | **6.6 s** (~13 s CPU, 2.7 GB) | | cold check, 59k LOC | 3.3 s | **1.24 s** | | warm no-op check, 295k / 59k | 7.2 s / 3.2 s | **2.0–2.6 s / 0.4–0.6 s** | | `run` after check (59k / 295k) | full compile | **0.12 s / 0.65 s** (cache hit) | | `pack` after check (59k / 295k) | 1.7 s / 40.5 s (always cold) | **0.19 s / 0.87 s** (cache hit) | | LSP edit→diagnostics (250/500/1000 files) | 0.46 / 1.7 / 5.5 s | **179 / 192 / 216 ms** (flat) | | `describe`, 295k / 59k LOC | 6.3 s / 716 ms | **2.6 s / 402 ms** | Cold compile scaling is now ~linear in project size (was quadratic: 5× code cost 33× CPU), verified to 8000 modules. ## Verification - Diagnostics byte-identical to the pre-change binary (error-heavy corpus, cold + warm) at each stage. - Cache soundness: `BAML_CACHE_VERIFY=1` honest-recompile compare (no-op + after-edit) and forced `BAML_CACHE_SAMPLED_VERIFY=1` across edit sequences — all pass; 43 bytecode-cache oracle tests updated and green. - LSP load-tested: 50-edit multi-file burst storms + sustained ~17 edits/s churn at 250/1000 files — zero crashes, stable RSS, convergence after every storm. - Full stack green in CI (doc, Windows incl. MSVC sha2 gating, size gate, all SDK matrices). ## Deliberate trade-offs - Cold `check` at mid-size projects (~15–30k LOC) is ~10% slower than before: it now emits + stores once to warm the cache; every later check/run/pack is served warm. - A warm *edit* check re-emits and stores (that is what buys the instant run/pack); at 295k LOC the edit path remains ~6–9 s — the known remaining floor, tracked separately. - A unit file deleted out from under a warm cache is not detected on a no-op run; the next compile that needs it falls back to a byte-identical full compile. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01V4oee6wrCHzXH6mM2h8eJo <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * CLI commands now share a consistent, cache-aware project session workflow, including lenient read-only sessions for inspection when configuration is missing. * **Performance** * Faster project loading and diagnostics checking via parallel per-file processing. * Improved bytecode-cache reuse by avoiding unnecessary payload hashing/loading, including a fast “no-delta” warm path. * **Bug Fixes** * Fixed callable-throws fragment seeding and verification for warm “no-delta” runs. * Improved package-level diagnostics handling during compilation and error-gating behavior. * **Chores** * Improved build robustness for cryptography acceleration on MSVC targets. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Claude Fable 5 <[email protected]>
This branch was previously 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
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.
Stacked on #4098. Fixes the "12 threads = 1.18× speedup" wall, diagnosed by comparing our salsa usage against ty/ruff's (our driver pattern was already identical to theirs — the problem was one query shape).
What
package_items/namespace_items) fold over every file's index; on a cold db the first worker to claim that memo computed parse+lower+index for the entire project inline on one thread while all other workers parked on the memo's sync slot. Also retires the old first-file warm-up, which warmed the builtin package, not the user package.add_or_update_fileswrites the project file list once (the per-file path re-cloned the whole Vec and bumped the salsa revision per file — O(files²) during load); source reads are parallel with discovery order preserved (FileId assignment and diagnostic ordering unchanged).End-user impact (measured)
Editor latency is now ~flat in project size.
Verification
baml_project/bex_project/baml_cli.Not in this PR: ty-style snapshot+cancellation for the LSP (blocked on the
panic = "abort"release profile — cancellation requires unwinding; needs a product decision), and the warm-edit floor (throws-gate + unit loads; separate reviewed design).🤖 Generated with Claude Code
https://claude.ai/code/session_01V4oee6wrCHzXH6mM2h8eJo