perf: optimize parser hot paths for ~18% faster compilation

[MathiasWP]

Summary

Optimizes the Svelte compiler's parser phase — the hot path that runs on every file compilation during development (HMR). The changes are surgical micro-optimizations to 7 files in packages/svelte/src/compiler/phases/1-parse/, with no architectural changes.

Changes

• Parser.match() — slice → startsWith: Avoids creating a new substring on every multi-char match call (called thousands of times per compilation)
• Parser.match_regex() — slice → sticky flag: Uses lastIndex + y flag instead of slicing the entire template string on every regex match. All 14 regex patterns passed to match_regex/read across 4 files updated with y flag.
• Parser.allow_whitespace() / require_whitespace() — regex → charCode: Replaces regex_whitespace.test() with numeric charCode comparisons in a tight loop (called hundreds of times per compilation). Covers the full \s character set with a fast path for common ASCII whitespace.
• Text parser — string concatenation → slice: Replaced data += template[index++] loop (O(n²)) with tracking start position + single slice() call (O(n))
• read_sequence — string concatenation → slice: Same pattern — derive raw from template.slice() in flush() instead of char-by-char concatenation
• is_valid_element_name() — inline regexes → module-level constants: Extracted 2 regex literals to module scope
• Dynamic new RegExp() → indexOf: Replaced parser.read_until(new RegExp(...)) with template.indexOf() for script/style closing tags
• match_bracket — Array.includes → Set.has: Pre-computes close bracket values as a Set for O(1) lookups

Benchmark

A benchmark script was written that:

1. Compiles 5 real test components of varying complexity (a11y validators, form bindings, component migration output, rich selects)
2. Compiles 2 synthetic stress-test templates (500 elements and 2000 elements with expressions and entities)
3. Runs 50 warmup iterations for parse, 20 for compile (JIT warmup)
4. Forces GC between benchmarks via --expose-gc
5. Measures both parse-only and full compilation (parse + analyze + transform + codegen)

Parse-only results

Component	Before	After	Speedup
a11y-role-supports-aria-props (371 lines)	2.410ms	2.129ms	-11.7%
form-default-value-spread (199 lines)	1.726ms	1.360ms	-21.2%
svelte-component (251 lines)	1.925ms	1.648ms	-14.4%
select-with-rich-content (157 lines)	0.795ms	0.636ms	-20.0%
rich-select (173 lines)	0.807ms	0.637ms	-21.1%
synthetic-large (500 elements)	2.560ms	1.547ms	-39.6%
synthetic-huge (2000 elements)	136.849ms	114.115ms	-16.6%
Total	147.071ms	122.072ms	-17.0%

Full compile results

Component	Before	After	Speedup
a11y-role-supports-aria-props	25.407ms	21.990ms	-13.4%
form-default-value-spread	10.711ms	9.422ms	-12.0%
svelte-component	10.304ms	8.481ms	-17.7%
select-with-rich-content	5.986ms	4.694ms	-21.6%
rich-select	6.010ms	5.059ms	-15.8%
synthetic-large (500 elements)	30.511ms	25.196ms	-17.4%
synthetic-huge (2000 elements)	932.801ms	764.794ms	-18.0%
Total	1021.729ms	839.637ms	-17.8%

Benchmarked on macOS, Node.js, Apple Silicon. The parse improvements carry through to end-to-end compile time since the parser is a significant portion of the compilation pipeline. Text-heavy templates benefit the most (up to 40% faster parse).

[changeset-bot]

🦋 Changeset detected

Latest commit: 020a683

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
svelte	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[Rich-Harris]

this is awesome!

[MathiasWP]

this is awesome!

I’ll do more investigation tomorrow, i have some findings that may be very interesting