Repository navigation
Faster front end, [In] parser fix, layout pins, diagnostic helpers - #156
Merged
Merged
Conversation
A for head's initializer is parsed with the grammar's [~In] so that `in`
can start a for-in, but every bracketed production inside it --
parentheses, array and object literals, arguments, member brackets,
template substitutions, function bodies -- takes [+In] again. The parser
cleared `allow_in` for the whole initializer, so `for (var a = ('k' in o);
...)` and `for (var f = function () { return k in o; }; ...)` were syntax
errors. The parser now records each token's bracket depth and withholds
`in` only at the depth the initializer started at; an unbracketed arrow
body still inherits the restriction.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each operand of a binary expression descended through ten one-function-per-level parsers (`||` down to `*`), each returning the expression by value and scanning its own operator list, before the one level that owned the operator saw it. One loop now takes an operand once and binds each operator by its precedence, building the same left-associative trees. Parsing imaging-darkroom (a 1.8 MB literal) takes 21% fewer cycles; crypto-md5's parse 12% fewer. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every identifier the compiler resolves looks its name up in the lexical scope maps and the local-slot map, which used the standard SipHash. They now use the engine's `NameMap`/`NameSet` (FxHash), like the runtime's name tables. Parsing and compiling cdjs takes about 10% fewer cycles. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Converting host UTF-8 to the engine's representation pushed every scalar through the surrogate-escape check, and function source text for `Function.prototype.toString` is converted this way for every function a script declares, nested ones included. Only a scalar at or above U+10000 -- a UTF-8 sequence led by 0xF0 or above -- can fall in the sentinel range, so text without one is copied as it is. cdjs's parse and compile run 9% fewer instructions. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Finishing each compiled body gathers the names it and its nested closures read and write, cloning every name (and every nested closure's whole cached list) into a BTreeSet<String>, once per store instruction. The sets now borrow the names and the lists are cloned once, at the end. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Whether LLVM inlined vm_numeric_leaf::math_binary into the typed-loop executor flipped with unrelated edits (here, front-end changes in other modules): inlined, it grew the executor by 176 bytes and re-rolled its code, and access-nsieve ran 4% and heterogeneous_property_read 5% more cycles at identical instructions. It is now out of line, as it was, and the order file is re-pinned. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Both are hot typed-executor callees (ai-astar's equality and named reads) that floated in the unordered tail; they now own slots after the other callees. With every callee fixed, a scan of the executor's offset put 0x200 best across the canaries (string_key_map_churn 0.98, heterogeneous_property_read 1.00, imaging-desaturate 0.97 against the 0x0 placement's 1.00, 1.06, 0.97). Byte-identical binaries at different paths still measure ai-astar 1-2% apart, which bounds what a scan can resolve there. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Three diagnostics this session's work kept rebuilding by hand in /tmp: `tools.benchmark.front_end` measures lexing, parsing and compiling of each external bundle (wrapped in a function never called, process start subtracted) against QuickJS-NG or another build, with each case's share of its whole run; `tools.benchmark.bundles` writes the corpus bundles to target/bundles for profiling; `scripts/layout-scan.sh` pins the typed-loop executor at each offset, relinks to a fixed point and screens the layout canaries. None touches the hashed measurement protocol. docs/performance-workflow.md lists them with the existing counter screen, which is the whole-corpus A/B to use instead of ad-hoc loops. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A binary's first process after a build or a pause runs slower, which inflated the empty-script baseline past small front ends (crypto-md5 read negative). One run is now discarded, and the least of the pairs estimates each cost. Also records the perf23 units and stack run in T033. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
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
string_from_utf8_scalars(function source text for every function); borrowed name sets in bytecode finalization.ininside brackets of aforinitializer (for (var a = ('k' in o); ...), a function body, arguments, templates) was a SyntaxError; it is withheld only at the initializer's own bracket depth now.math_binarykept out of the typed executor (a codegen-unit shift had inlined it: access-nsieve +4% at equal instructions),boxed_equalityandget_namedpinned, executor offset re-scanned to 0x200.tools.benchmark.front_end(front-end cost vs NG or another build),tools.benchmark.bundles(corpus bundles on disk),scripts/layout-scan.sh(executor placement scan); documented in docs/performance-workflow.md and AGENTS.md. None touches the hashed measurement protocol.Evidence
Formal stack run 3babb93 vs main f2b21ab (30 blocks,
target/comparison/perf23-3babb930): external geomean 0.994 vs main, 0.772 vs QuickJS-NG; gaussian-blur 0.952, imaging-darkroom 0.963, tinderbox 0.965. Worst: ai-astar 1.023 (two byte-identical binaries at different paths differ by 1-2% on it), regexp-dna 1.015. Sentinels 1.003.Verification
cargo test -p qjs-parser,cargo test -p qjs-runtime(2246), new tests for precedence/associativity and the for-initializerincases (expected values from V8).language/statements/for,for-in,expressions/in,exponentiation,coalesce,logical-and: 0 gaps.check.shvia pre-push; tools unit tests.🤖 Generated with Claude Code