[V8 perf] Loop callback changing is throttling

首先克隆链接的复现代码，并在 Node.js v24 下运行 demo-1.ts 和 demo-2.ts，比较切换 callback 时的耗时。调查报告中的 V8 多态 inline cache 行为，并针对 BOX 和 LANCZOS 情况验证结果。当更改 callback 不再导致重复的 downsample2D 调用相较于预期耗时变慢时，即视为完成。

Version

v24.20.0

Platform

Windows 10

Subsystem

No response

What steps will reproduce the bug?

• Clone the repo https://github.com/AlttiRi/loop-callback-throttling-bug
• Install deps ("sharp")
• Run demo-1.ts and demo-2.ts

How often does it reproduce? Is there a required condition?

Always

What is the expected behavior? Why is that the expected behavior?

• BOX (~5 secs) ✅
• BOX (~5 secs) ✅
• LANCZOS (~8 secs) ✅
• LANCZOS (~8 secs) ✅

What do you see instead?

• BOX (~5 secs) ✅
• BOX (~5 secs) ✅
• LANCZOS (~30 secs) ❌
• LANCZOS (~30 secs) ❌

Additional information

I have downsample1D_V1 function (it's very unoptimized, but it works, and it's perfect for the demonstration), which executes the passed callback const weight = fn(x); billions of times (3_107_687_040 times per 1 downsample2D call).

There are 2 types of the passed callback function — lanczos3 and box:

function lanczos3(x: number): number {
  if (-3 <= x && x < 3) {
    return sinc(x) * sinc(x / 3);
  }
  return 0;
}

function sinc(x: number): number {
  if (x === 0) {
    return 1;
  }
  x = Math.PI * x;
  return Math.sin(x) / x;
}


function box(x: number): number {
  if (x > -0.5 && x <= 0.5) {
    return 1;
  }
  return 0;
}

I call downsample2D multiple times with the follow callback params (demo-1.ts):

• BOX (~5 secs) ✅
• BOX (~5 secs) ✅
• LANCZOS (~30 secs) ❌
• LANCZOS (~30 secs) ❌

The first 2 calls work with the expected speed, for ~5 second.
But when the callback is changed, downsample2D works multiple times slower than expected, for ~30 seconds.

In reverse order I also face the throttling (demo-2.ts):

• LANCZOS (~8 secs) ✅
• LANCZOS (~8 secs) ✅
• BOX (~30 secs) ❌
• BOX (~30 secs) ❌

The problem appears on both Node.js (24) and Deno (2.9).

Bun (1.3/1.4) works fine:

• BOX (~5 secs) ✅
• BOX (~5 secs) ✅
• LANCZOS (~8 secs) ✅
• LANCZOS (~8 secs) ✅

Related keywords: performance v8 polymorphic inline cache