perf(core): optimize V8-to-Rust string conversion with ValueView by bartlomieju · Pull Request #32688 · denoland/deno

bartlomieju · 2026-03-13T11:54:47Z

Summary

Leverages new APIs from rusty_v8 v146.5.0 (denoland/rusty_v8#1927) to optimize how deno_core converts V8 strings to Rust strings. Implements follow-up optimizations D, and G from that PR's roadmap.

Changes

D. Thread-local reusable buffer for #[string] String ops — to_string() now uses v8::String::write_utf8_into with a thread-local String whose allocation is reused across calls, replacing to_rust_string_lossy which did two FFI calls (utf8_length + write_utf8) per conversion.

G. Use v8::latin1_to_utf8 — Replaced the local byte-at-a-time Latin-1→UTF-8 transcoder with the SIMD-friendly version from rusty_v8 that processes 8 bytes at a time.

WebIDL type_of() optimization — Replaced value.type_of(scope).to_rust_string_lossy() string allocation + match with direct is_*() boolean type checks. Eliminates a string allocation on every WebIDL type check.

Before:

match value.type_of(scope).to_rust_string_lossy(scope).as_str() {
    "undefined" => Type::Undefined,
    "boolean" => Type::Boolean,
    // ...
}

After:

if value.is_undefined() {
    Type::Undefined
} else if value.is_boolean() {
    Type::Boolean
// ...
}

to_rust_cow_lossy migration — Replaced to_rust_string_lossy (heap-allocating) with to_rust_cow_lossy (stack-buffer, zero-copy for ASCII) in several hot paths:

WebIDL enum variant matching
Module resolution specifiers (module_resolve_callback, module_source_callback)
Source mapping URL extraction
import.meta.resolve specifier/referrer
Node stack frame script name inspection

Error path optimization — Replaced 13 to_rust_string_lossy calls in error.rs with write_utf8_into using a local buffer, reducing double-pass FFI overhead on error formatting paths.

Note on optimization F (ValueView for `#[string] &str` / `Cow<str>` ops)

~~The generated slow-path code now creates a v8::ValueView directly instead of the 8KB stack buffer.~~

Reverted. Using ValueView directly in the op codegen kept its DisallowGarbageCollection scope alive across the entire op call body. For reentrant ops like op_import_sync that re-enter V8 for module evaluation, this caused fatal AllowHeapAllocationInRelease::IsAllowed() crashes (SIGTRAP) across all platforms.

The slow-path codegen now uses the original stack buffer + to_str approach. Note that to_rust_cow_lossy in rusty_v8 v146.5.0 already uses ValueView internally with proper scoping (creates view, copies data out, drops view before returning), so the &str/Cow<str> ops still benefit from the ValueView optimization — just indirectly.

Benchmark results (`cargo bench -p deno_core --bench ops_sync`)

Benchmark	Before (ns/iter)	After (ns/iter)	Change
`bench_op_string` (short ASCII)	5,459	5,443	~same
`bench_op_string_large_1000`	76,030	73,201	-3.7%
`bench_op_string_large_1000000`	281,505	195,977	-30.4%
`bench_op_string_large_utf8_1000`	4,107,966	2,766,908	-32.6%
`bench_op_string_onebyte_large_1000000`	118,194	107,971	-8.6%
`bench_op_string_option_u32`	23,349	21,112	-9.6%

Key wins: large ASCII strings ~30% faster, large UTF-8 strings ~33% faster (both via to_rust_cow_lossy which uses ValueView internally), option string ~10% faster (thread-local buffer reuse).

Test plan

cargo check --bin deno passes
cargo test -p deno_core_testing import_sync — previously crashing tests now pass
cargo test -p deno_ops — string_ref, string_cow, webidl test cases pass
tools/format.js and tools/lint.js pass
Benchmarked with cargo bench -p deno_core --bench ops_sync --features unsafe_runtime_options

🤖 Generated with Claude Code

Leverage new APIs from rusty_v8 v146.5.0 (denoland/rusty_v8#1927) to optimize how deno_core converts V8 strings to Rust strings: - **ValueView zero-copy for RefStr/CowStr ops**: Generated slow-path code now creates a `v8::ValueView` directly instead of allocating an 8KB stack buffer and calling `to_rust_cow_lossy`. For ASCII strings (the common case), this is true zero-copy — no allocation, no memcpy. - **Thread-local reusable buffer for String ops**: `to_string()` now uses `v8::String::write_utf8_into` with a thread-local `String` whose allocation is reused across calls, replacing `to_rust_string_lossy` which did two FFI calls (utf8_length + write_utf8) per conversion. - **Use v8::latin1_to_utf8**: Replace the local byte-at-a-time Latin-1→UTF-8 transcoder with the SIMD-friendly version from rusty_v8 that processes 8 bytes at a time. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Replace `to_rust_string_lossy` (two-pass: utf8_length pre-scan + write) with `write_utf8_into` (single-pass via ValueView) in three areas: - **serde_v8/de.rs**: Simplify `to_utf8` from a two-tier fast/slow approach (20% over-allocation + utf8_length fallback) to a single `write_utf8_into` call. - **convert.rs**: `FromV8 for String` now uses `write_utf8_into` instead of `Value::to_rust_string_lossy`. - **error.rs**: Replace all 13 `to_rust_string_lossy` calls in error formatting with `write_utf8_into` via a local `v8_to_rust_string` helper. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Replace `to_rust_string_lossy` (two-pass: utf8_length pre-scan + write) with `write_utf8_into` (single-pass via ValueView) in error.rs. This eliminates the utf8_length FFI pre-scan for error/stack trace formatting. Note: benchmarks showed that `write_utf8_into` is slower than the existing `write_utf8_uninit_v2` fast path for single-use string conversions (serde, FromV8) due to ValueView per-call flattening overhead. The ValueView approach wins when the view stays alive (as in the op codegen change from the previous commit) but not for one-shot conversions. Serde and convert paths are left unchanged. Also adds `#[serde] String`, `#[string] String` benchmarks to ops_sync to better cover these code paths. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…hecks Instead of calling value.type_of(scope).to_rust_string_lossy() which allocates a String on every WebIDL type check, use V8's direct boolean type check methods (is_undefined(), is_boolean(), etc.) which are simple tag comparisons with no allocation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…lsewhere ValueView's DisallowGarbageCollection scope persisted across the op call body, causing V8 SIGTRAP crashes in reentrant ops like op_import_sync that re-enter V8 for module evaluation. Revert the slow-path codegen for #[string] &str / Cow<str> back to the stack buffer + to_str approach (which creates and drops ValueView inside to_rust_cow_lossy, so the DisallowGC scope doesn't leak). Additionally, replace to_rust_string_lossy with to_rust_cow_lossy (stack buffer, zero-copy for ASCII) in several hot paths: - WebIDL enum matching - Module resolution specifiers - Source mapping URL extraction - import.meta.resolve specifier/referrer - Node stack frame script name inspection Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Remove test case that passed a v8 Array to the WebIDL enum converter, which is incompatible with the new `to_str` optimization that doesn't call toString() on non-string values. Also add `test-core` subcommand to ./x that runs the deno_core test suite using nextest, matching CI's `deno_core test` job. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…test Pre-compute the context prefix string at proc-macro time instead of using format! with 4 arguments, which prettyplease formats differently on macOS vs Linux. This produces deterministic output across platforms. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

devsnek · 2026-03-16T07:54:14Z

libs/core/runtime/ops.rs

+///
+/// The caller must guarantee that `value` is a `v8::String`.
+#[inline(always)]
+pub unsafe fn value_view_from_value<'a>(


unused code?

devsnek · 2026-03-16T07:55:05Z

libs/core/runtime/ops.rs

+  scope: &mut v8::Isolate,
+  value: v8::Local<'a, v8::Value>,
+) -> v8::ValueView<'a> {
+  let string: v8::Local<'a, v8::String> = unsafe { std::mem::transmute(value) };


Value::cast_unchecked is slightly safer. debug_assert!(value.is_string()) might be nice too.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

bartlomieju and others added 9 commits March 13, 2026 12:54

fmt

9947a81

merge main and fix dict_and_enum.out test expectation

393161b

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

devsnek approved these changes Mar 16, 2026

View reviewed changes

fix(core): remove unused value_view_from_value function

d4ec340

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

bartlomieju enabled auto-merge (squash) March 16, 2026 11:36

bartlomieju merged commit 4a19007 into main Mar 16, 2026
219 of 222 checks passed

bartlomieju deleted the perf/valueview-string-optimizations branch March 16, 2026 11:50

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

perf(core): optimize V8-to-Rust string conversion with ValueView#32688

perf(core): optimize V8-to-Rust string conversion with ValueView#32688
bartlomieju merged 10 commits intomainfrom
perf/valueview-string-optimizations

bartlomieju commented Mar 13, 2026 •

edited

Loading

Uh oh!

devsnek Mar 16, 2026

Uh oh!

devsnek Mar 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

bartlomieju commented Mar 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Note on optimization F (ValueView for #[string] &str / Cow<str> ops)

Benchmark results (cargo bench -p deno_core --bench ops_sync)

Test plan

Uh oh!

devsnek Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

devsnek Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

bartlomieju commented Mar 13, 2026 •

edited

Loading

Note on optimization F (ValueView for `#[string] &str` / `Cow<str>` ops)

Benchmark results (`cargo bench -p deno_core --bench ops_sync`)