POC: Perf/handlebars v2 parser

[johanrd]

POC from claude: Three parsers compared

• main — current @handlebars/parser (Jison-generated) + simple-html-tokenizer
• v2-parser (johanrd/ember.js @ perf/handlebars-v2-parser) — claude-iterated recursive descent JS parser replacing only the Jison HBS layer, keeping simple-html-tokenizer untouched, Where Jison's generated lexer tests up to 40 regexes per token and slices the input string on every match, the v2-parser uses an index-based cursor with indexOf('{{') for content scanning and charCodeAt dispatch for mustache classification — no string copies, no regex gauntlet. That's why it's ~1.8x faster at the HBS layer even though it's doing the same parse.
• rust/wasm — [FEATURE rust-parser] Rust/WASM template parser using pest.rs emberjs/ember.js#21313

Benchmarks run with Node 24, warmed JIT, on the full ember-template-compiler precompile() path (so this includes preprocess() → ASTv2 normalization → opcode encoding → wire format — the whole thing).

Parse only (ms/call)

template	main (Jison)	v2-parser (index-based cursor parser)	rust/wasm (this PR)
small (25 chars)	0.024ms	0.015ms	0.038ms
medium (352 chars)	0.177ms	0.102ms	0.478ms
real-world (1505 chars)	0.605ms	0.343ms	3.857ms
large (3520 chars)	1.688ms	0.942ms	22.960ms

Full pipeline results (ms/call)

template	chars	main (Jison)	v2-parser	rust/wasm (#21313))
small	25	0.047ms	0.038ms	0.067ms
medium	352	0.492ms	0.397ms	0.779ms
real-world	1494	1.832ms	1.577ms	4.947ms
large (10x medium)	3520	5.095ms	4.667ms	27.107ms

Parse vs compile split (medium template)

phase	main (Jison)	v2-parser	rust/wasm
preprocess() only	0.175ms (40%)	0.093ms (26%)	0.480ms (66%)
compile only (shared)	0.262ms (60%)	0.266ms (74%)	0.250ms (34%)
total	0.438ms	0.358ms	0.730ms

The compile step (ASTv2 normalization + opcode encoding) costs the same ~0.25ms in all three — it's identical code. Only the parse phase differs.

What does this show?

The compile step (ASTv2 normalization + opcode encoding) costs ~0.25ms in all three — identical code. The gap is entirely in preprocess(), and it compounds: rust/wasm's JSON bridge (serde_json::to_string → JSON.parse() → convertLocations() walk) is O(AST size), so the gap widens with template complexity (2x at medium → 5.8x at large).

The single-pass architecture is a real win in theory — the current pipeline genuinely scans HTML twice (@handlebars/parser treats it as opaque content, then simple-html-tokenizer re-tokenizes it via tokenizePart()).

[github-actions]

📊 Package size report   -0%↓

File	Before (Size / Brotli)	After (Size / Brotli)
Total (Includes all files)	17.1 MB / 3.2 MB	-0%↓17.1 MB / -0%↓3.2 MB
Tarball size	3.9 MB	-0%↓3.9 MB

_{🤖 This report was automatically generated by pkg-size-action}