Improve app pattern performance gaps - #1296
Merged
Merged
Conversation
TheHypnoo
marked this pull request as draft
May 21, 2026 20:34
Member
Author
Post-review fixes (7 commits + fmt)Applied the items raised in the self-review above:
Before / after — full app-pattern matrixAbsolute numbers on the post-fix sweep are noticeably higher across all three runtimes (Bun and Node are 1.5-3× slower too) — that's machine noise / thermal on the rerun host, not code. The honest comparison is the perry/bun ratio column:
Summary: 5 better, 5 same, 1 worse, 0 in the slow bucket (≥2×) — same as the PR baseline. Trade-off:
|
TheHypnoo
added a commit
that referenced
this pull request
May 21, 2026
) `crates/perry-codegen/src/collectors/escape_news.rs` was 2034 lines and failed the 2,000-line file-size CI gate. Splitting it into topical sibling modules under `collectors/` — no behavior change; function bodies move verbatim and the same public/internal API is re-exported from `collectors/mod.rs`. Moves: - `this_as_value.rs` — `class_uses_this_as_value`, `class_chain_extends_builtin_error`, `stmts_use_this_as_value`, `expr_uses_this_as_value`. - `class_accessors.rs` — `is_class_getter`, `is_class_setter`. - `escape_check.rs` — `find_new_candidates`, `check_escapes_in_stmts`, `check_escapes_in_expr`. - `local_refs.rs` — `expr_contains_local_get`, `mark_all_candidate_refs_in_expr`. `escape_news.rs` keeps the escape-collection entry points (`collect_non_escaping_news`, `collect_non_escaping_new_used_fields` and the two private helpers) plus the trailing `MAX_SCALAR_ARRAY_LEN` const. Final line counts: escape_news 693, escape_check 870, this_as_value 337, local_refs 114, class_accessors 50 — all comfortably under the 2,000 gate.
…is shared When the same pending promise is fed into multiple Promise.all([...]) calls (legal JS), only the first registered state was being drained on settle because promise_all_take_handler used swap_remove on the first match. The second Promise.all would never resolve. Replace the take handler with promise_all_take_all_handlers, which drains every PROMISE_ALL_STATES entry keyed on the settling promise. Both js_promise_resolve and js_promise_reject in then.rs now enqueue one PromiseAll task per drained state. Adds a unit test in promise/combinators.rs (promise_all_with_shared_pending_input_resolves_both) that registers two Promise.all calls sharing a pending input and verifies both settle after the shared promise resolves.
…d path With GC_THRESHOLD_INITIAL_BYTES now equal to GC_TRIGGER_ABSOLUTE_CEILING (both 128 MB after app-pattern tuning), the C4b-delta-tune hard cap on next_trigger supersedes any adaptive doubling. Add an inline note above the doubling branch explaining the cap and preserving the 1 GB GC_THRESHOLD_MAX_BYTES bound as an intentional escape hatch for future threshold raises or env-driven bisection.
The length<=64 fast path skips the generational write barrier and only records layout metadata, on the assumption that the freshly-allocated result array stays in the nursery for the life of the loop. Document the invariant explicitly so a future GC policy change that tenures aggressively (or PERRY_GC_FORCE_EVACUATE stress runs) doesn't silently break the remembered set for this code path.
TypeScript classes can declare `get` accessors with arbitrary side
effects, and the discard-map optimization has no general way to prove
that a PropertyGet on the callback parameter targets a plain data slot.
Remove PropertyGet from the discard_pure_expr arms — anonymous-shape
constructions still flow through the Expr::New { __AnonShape_… } arm,
which is the intended target of the optimization.
Per ES2015 §12.2.5, `[...x]` invokes GetIterator(x) which throws a
TypeError when x is null or undefined. The single-spread fast path
previously routed straight through js_array_clone, which silently
returns an empty array for those inputs (kept for back-compat with
Array.from's not-iterable behavior).
Add a dedicated runtime entry point js_array_clone_for_spread that
inspects the NaN-box tag bits on the raw boxed value and throws a
TypeError("<receiver> is not iterable") for null/undefined before
forwarding to js_array_clone. Update the codegen single-spread arm
to call it. Array.from's behavior is unchanged.
The (target_has_slots, slot) tuple where target_has_slots was always literally true was a noisy leftover from an earlier shape of this branch. Collapse it to a plain Option<slot> lookup.
- escape_news.rs: clarify why the trailing `_ => {}` arm is kept after
the explicit Expr-variant list — new allocating variants should be
added explicitly above rather than relying on the fallthrough.
- json/parser.rs: explain the 96-bytes/object pre-size heuristic for
arrays-of-objects (empirical average for small JSON objects with a
handful of short string keys), plus the 16..16_384 clamp rationale.
) `crates/perry-codegen/src/collectors/escape_news.rs` was 2034 lines and failed the 2,000-line file-size CI gate. Splitting it into topical sibling modules under `collectors/` — no behavior change; function bodies move verbatim and the same public/internal API is re-exported from `collectors/mod.rs`. Moves: - `this_as_value.rs` — `class_uses_this_as_value`, `class_chain_extends_builtin_error`, `stmts_use_this_as_value`, `expr_uses_this_as_value`. - `class_accessors.rs` — `is_class_getter`, `is_class_setter`. - `escape_check.rs` — `find_new_candidates`, `check_escapes_in_stmts`, `check_escapes_in_expr`. - `local_refs.rs` — `expr_contains_local_get`, `mark_all_candidate_refs_in_expr`. `escape_news.rs` keeps the escape-collection entry points (`collect_non_escaping_news`, `collect_non_escaping_new_used_fields` and the two private helpers) plus the trailing `MAX_SCALAR_ARRAY_LEN` const. Final line counts: escape_news 693, escape_check 870, this_as_value 337, local_refs 114, class_accessors 50 — all comfortably under the 2,000 gate.
TheHypnoo
force-pushed
the
feat/performance-app-pattern-gaps
branch
from
May 22, 2026 07:07
dc4571f to
328ec5d
Compare
TheHypnoo
marked this pull request as ready for review
May 22, 2026 07:07
proggeramlug
added a commit
that referenced
this pull request
May 22, 2026
…sweep (#1414) Rolls up 26 PRs that merged to main post-v0.5.1023 without version bumps: - node:crypto gap-fixes (#1386 #1393 #1394 #1402 #1405): randomInt, timingSafeEqual, getHashes/getCiphers, sha224/sha384, base64 digest, Buffer hash input, no-arg digest() → Buffer, pbkdf2Sync digest arg, scryptSync. - node:perf_hooks (#1321 + #1328 #1342 coverage): performance + User Timing + PerformanceObserver native impl, granular node-suite + edge-case coverage. - #1090 GC checkpoint runtime work (#1324). - #1311 geisterhand on iOS (#1316 #1383 #1384 #1385). - #1312 process.env.X (unset) is nullish undefined (#1314). - #1319 thread-safety hardening for cross-thread runtime statics. - #1322 exact-head GC evidence packet. - #1323 wasm timers dispatch through mem_call bridge (#1329). - #1317 node:timers/promises shadow-segfault fix (#1326). - #1330 node:process suite (#1331). - #1292 bcrypt.hash() returns String (#1307). - #1293 fastify .json()/.body external-fastify dispatch (#1308). - #1296 app pattern performance gaps. - #1297 diagnostics_channel parity. - #1301 iOS App Groups capability (#1313). - #1318 #1325 os/methods/modern-methods static dispatch. - #1315 expanded Node parity test coverage. - #1382 ui-ios stdlib pump for async fetch. - #1392 ui-wasm reactive state + setText (#1404).
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
1 participant
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.
Summary
This PR closes the slow app-pattern performance gaps found during the benchmark audit. It adds targeted runtime and codegen optimizations for JSON parsing, scalar-replaced object literals, discarded
Array.mapwork,Promise.all, async-step microtasks, buffer transcode allocation pressure, and array/string hot paths.What changed
JSON.parsepaths with hot shape reuse, parse-owned fast stores, key-cache improvements, lazy tape selection, stack-backed small object fields, and decimal fast paths.Array.mapresults and skips pure discarded map callbacks that only build unused anonymous objects.Promise.allby replacing per-input closures with a directTask::PromiseAllpath and GC-scanned state.Before / after
Initial focused baseline:
json_parse_1mbobject_deep_clonepromise_all_chainsbuffer_transcodeFull app-pattern matrix after this PR:
buffer_transcodedate_format_parsejson_parse_1mbjson_stringify_1mbmap_1mobject_deep_clonepromise_all_chainsregex_replacestring_concat_csvstring_split_map_joinstring_template_interpNo app-pattern kernel remains in the
slowbucket.Remaining gaps
promise_all_chainsis still borderline; the remaining cost appears to be async state-machine and microtask overhead rather thanPromise.allitself.string_concat_csvis just over the OK threshold and should be a focused follow-up for mixed string concat chains.date_format_parsestill reports a tiny checksum mismatch against Bun/Node; that is a correctness follow-up, not a performance blocker for this PR.Validation