src: avoid copying source string in TextEncoder.encode

[anonrig]

Summary

EncodeUtf8String, which backs TextEncoder.prototype.encode(), copied the entire source string out of the V8 heap into a MaybeStackBuffer (via WriteOneByteV2/WriteV2) before encoding — and allocated on the heap for strings larger than the stack buffer (> 4096 chars). The sibling EncodeInto already avoids this by reading the flat content directly through v8::String::ValueView.

This reads the flat content via ValueView instead. Because ValueView carries a DisallowGarbageCollection scope, the backing store cannot be allocated while the view is alive, so the view is used in two short scopes:

1. Size pass — validate the encoding and compute the exact UTF-8 output length.
2. Allocate the exactly-sized backing store (GC now permitted).
3. Encode pass — re-acquire the view (flattening is cached on the string, so this is cheap) and convert directly into the backing store.

The rare unpaired-surrogate path still copies into a mutable buffer, since to_well_formed_utf16 runs in place. Output buffer sizing and WHATWG replacement semantics are unchanged.

Benchmarks

benchmark/util/text-encoder.js (op=encode, n=1e6, 12 interleaved runs each via compare.js):

type	len=256	len=1024	len=8192
ascii	+14.1%	+23.5%	+43.9%
one-byte (latin1)	+14.4%	+22.2%	+12.3%
two-byte (utf-16)	+16.7%	+20.4%	+15.5%

len=32 uses the unchanged small-string path (within noise). As a control, the untouched encodeInto path stayed flat (−2.0%..+0.5%) across all 12 configurations, confirming the harness is unbiased.

Verification

• Builds clean (full relink).
• All encoding / TextEncoder / TextDecoder parallel tests pass.
• Functional spot-checks across ASCII, Latin1, BMP, valid surrogate pairs, and lone/reversed surrogates match Buffer.from(s, 'utf8').

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/74106/