SUZUKI Sosuke 9484218ba4 perf(buffer): move Buffer.slice/subarray to native C++ with int32 fast path (#26819) ...

## Summary

Move `Buffer.slice()` / `Buffer.subarray()` from a JS builtin to a
native C++ implementation, eliminating the `adjustOffset` closure
allocation and JS→C++ constructor overhead on every call. Additionally,
add an int32 fast path that skips `toNumber()` (which can invoke
`valueOf`/`Symbol.toPrimitive`) when arguments are already int32—the
common case for calls like `buf.slice(0, 10)`.

## Changes

- **`src/bun.js/bindings/JSBuffer.cpp`**: Add
`jsBufferPrototypeFunction_sliceBody` with `adjustSliceOffsetInt32` /
`adjustSliceOffsetDouble` helpers. Update prototype hash table entries
from `BuiltinGeneratorType` to `NativeFunctionType` for both `slice` and
`subarray`.
- **`src/js/builtins/JSBufferPrototype.ts`**: Remove the JS `slice`
function (was lines 667–687).
- **`bench/snippets/buffer-slice.mjs`**: Add mitata benchmark.

## Benchmark (Apple M4 Max)

| Benchmark | Before (v1.3.8) | After | Speedup |
|---|---|---|---|
| `Buffer(64).slice()` | 27.19 ns | **14.56 ns** | **1.87x** |
| `Buffer(1024).slice()` | 27.84 ns | **14.62 ns** | **1.90x** |
| `Buffer(1M).slice()` | 29.20 ns | **14.89 ns** | **1.96x** |
| `Buffer(64).slice(10)` | 30.26 ns | **16.01 ns** | **1.89x** |
| `Buffer(1024).slice(10, 100)` | 30.92 ns | **18.32 ns** | **1.69x** |
| `Buffer(1024).slice(-100, -10)` | 28.82 ns | **17.37 ns** | **1.66x**
|
| `Buffer(1024).subarray(10, 100)` | 28.67 ns | **16.32 ns** | **1.76x**
|

**~1.7–1.9x faster** across all cases. All 449 buffer tests pass.

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

2026-02-09 01:46:33 -08:00

124 KiB

Raw Permalink Blame History

View Raw