optimize TextDecoder for legacy encodings

[anonrig]

This should optimize performance of TextDecoder for legacy encodings (as reported by @ChALkeR)

[ask-bonk]

I'm Bonk, and I've done a quick review of your PR.

This PR optimizes TextDecoder for legacy encodings by reusing a Rust-side output buffer across decode calls instead of allocating a new Vec<u16> each time, returning a raw pointer+length pair through the CXX bridge instead of moving the Vec.

Issues

1. [Critical] Trailing semicolon makes decode() fail to compile

Line 194 of lib.rs has }; (with a trailing semicolon) on the final ffi::DecodeResult expression. In Rust, this turns the tail expression into a statement, changing the function's implicit return from ffi::DecodeResult to (). This is a compile error — the function signature requires -> ffi::DecodeResult. (The early return on the Malformed path at line 156 is correct — it uses return explicitly.)

2. [Low] Unbounded buffer growth

The state.output buffer doubles on OutputFull but is never shrunk. A single large decode permanently inflates the allocation for the lifetime of the decoder. For streaming use cases (TextDecoderStream), this is probably fine since decoders are short-lived. For long-lived decoders that occasionally see a large chunk, you may want to shrink the buffer if it exceeds some threshold (e.g., if state.output.capacity() > 4 * state.output.len() { state.output.shrink_to(...) }). Not blocking — just something to be aware of.

[ask-bonk]

I'm Bonk, and I've done a quick review of your PR.

The review has been posted on PR #6295. Here's a summary of what I found:

1. [Critical] Trailing semicolon on lib.rs:194 — The final ffi::DecodeResult { ... }; has a stray semicolon that turns the tail expression into a statement. This makes the function return () instead of ffi::DecodeResult, which is a compile error. I posted a suggestion to remove the semicolon.

2. [Low] Unbounded buffer growth — The reusable state.output buffer doubles on OutputFull but is never shrunk back. A single large decode permanently inflates memory for the decoder's lifetime. Not blocking, but noted as something to consider for long-lived decoders.

github run

[github-actions]

The generated output of @cloudflare/workers-types matches the snapshot in types/generated-snapshot 🎉