V8引擎中的Ignition和TurboFan
学习思考
Post on: May 29, 2023
Last edited: 2023-10-8
Views
type
Post
status
Published
date
May 29, 2023
slug
javascript-v8-turbo
summary
V8引擎的即时编译（JIT）策略显著提高了从JavaScript生成可执行代码的速率（编译器TurboFan的功劳），我们有必要对这个机制有基本的认识，从而避免写出一些影响性能的代码。
tags
思考
开发
前端
category
学习思考
icon
password
Property
Oct 8, 2023 09:02 AM
前端性能优化中常被我们忽略的一点是在引擎层面JavaScript被解析和执行耗费的时间。对于V8引擎来说，JavaScript被解析器解析为抽象语法树（AST）后，会进入一个解释器+编译器的组合管道执行，解释器会输出字节码，而编译器直接输出机器代码（https://v8.dev/blog/launching-ignition-and-turbofan）。
形象的V8引擎点火器 + 涡轮风扇的组合
形象的V8引擎点火器 + 涡轮风扇的组合
V8引擎的这种即时编译（JIT）策略显著提高了从JavaScript生成可执行代码的速率（编译器TurboFan的功劳），我们有必要对这个机制有基本的认识，从而避免写出一些影响性能的代码。

Inigition和TurboFan

notion image
上图阐述了一个JavaScript文件如何在浏览器中被执行的过程，首先从远程服务器上获取这个文件（由于网络的延迟不是小问题，我们的优化经常在这一步，例如减少JavaScript bundle的大小以提升首屏加载的速度），获取到完整的源文件后，浏览器会解析这个文件，在内存中生成AST数据结构，交给Ignition加上TurboFan的管道执行。
Ignition（点火器）顾名思义是先开始工作的，它生成字节码（并非CPU直接执行的机器码），执行速率相对没有那么快。这时候TurboFan编译器开始工作了，它会根据程序执行的历史来进行一些假定，试图对高频率执行的部分代码直接编译为机器码以供快速执行。

小实验

借助user timing API（https://developer.mozilla.org/zh-CN/docs/Web/API/Performance_API/User_timing），我们可以测量程序运行的时间。下面我们来看一个例子，重复执行某一段代码：
import { performance } from "perf_hooks";

let iterations = 1e8;

const a = 1;
const b = 2;

const add = (x: any, y: any) => x + y;

performance.mark("start");

while (iterations--) {
  add(a, b);
}

performance.mark("end");

performance.measure("My Special Benchmark", "start", "end");

const [measure] = performance.getEntriesByName("My Special Benchmark");
console.log(measure);
这是一段由TypeScript编写的Nodejs程序，我们将它编译为JavaScript：
"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
const perf_hooks_1 = require("perf_hooks");
let iterations = 1e8;
const a = 1;
const b = 2;
const add = (x, y) => x + y;
perf_hooks_1.performance.mark("start");
while (iterations--) {
    add(a, b);
}
perf_hooks_1.performance.mark("end");
perf_hooks_1.performance.measure("My Special Benchmark", "start", "end");
const [measure] = perf_hooks_1.performance.getEntriesByName("My Special Benchmark");
console.log(measure);
使用Nodejs执行上面的代码：
[marking 0x09c7828d3389 <JSFunction (sfi = 0x323b1849b881)> for optimized recompilation, reason: hot and stable]
[marking 0x09c7828d4699 <JSFunction add (sfi = 0x323b1849b939)> for optimized recompilation, reason: small function]
[compiling method 0x09c7828d4699 <JSFunction add (sfi = 0x323b1849b939)> (target TURBOFAN) using TurboFan]
[compiling method 0x09c7828d3389 <JSFunction (sfi = 0x323b1849b881)> (target TURBOFAN) using TurboFan OSR]
[optimizing 0x09c7828d3389 <JSFunction (sfi = 0x323b1849b881)> (target TURBOFAN) - took 0.083, 0.417, 0.000 ms]
[optimizing 0x09c7828d4699 <JSFunction add (sfi = 0x323b1849b939)> (target TURBOFAN) - took 0.125, 0.333, 0.041 ms]
[completed optimizing 0x09c7828d4699 <JSFunction add (sfi = 0x323b1849b939)> (target TURBOFAN)]
PerformanceMeasure {
  name: 'My Special Benchmark',
  entryType: 'measure',
  startTime: 27.794666051864624,
  duration: 49.99266695976257,
  detail: null
}
node —trace-opt dist/hello.js
可以看到TurboFan进行了编译优化。
不妨试试修改代码：
import { performance } from "perf_hooks";

let iterations = 1e8;

const a = 1;
const b = 2;

const add = (x: any, y: any) => x + y;

performance.mark("start");

while (iterations--) {
  add(a, b);
}

iterations = 1e8;

add("foo", "bar");

while (iterations--) {
  add(a, b);
}

performance.mark("end");

performance.measure("My Special Benchmark", "start", "end");

const [measure] = performance.getEntriesByName("My Special Benchmark");
console.log(measure);
将这段代码运行一下：
PerformanceMeasure {
  name: 'My Special Benchmark',
  entryType: 'measure',
  startTime: 26.50658416748047,
  duration: 265.9936249256134,
  detail: null
}
注释掉add(”foo”, “bar”)行：
PerformanceMeasure {
  name: 'My Special Benchmark',
  entryType: 'measure',
  startTime: 27.4902081489563,
  duration: 101.20350003242493,
  detail: null
}
可以看到时间显著减少了。
这里发生的事情事实上是因为中间对add的调用否定了TurboFan对这个函数签名的假定，它的类型假定失效，此时就会取消之前所做的编译优化，而等待接来下的循环执行了一段时间后才继续优化。
通过—trace-deopt参数可以查看这个过程：
[marking 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)> for optimized recompilation, reason: hot and stable]
[marking 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> for optimized recompilation, reason: small function]
[compiling method 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> (target TURBOFAN) using TurboFan]
[compiling method 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)> (target TURBOFAN) using TurboFan OSR]
[optimizing 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)> (target TURBOFAN) - took 0.083, 0.542, 0.000 ms]
[optimizing 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> (target TURBOFAN) - took 0.125, 0.292, 0.000 ms]
[completed optimizing 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> (target TURBOFAN)]
[bailout (kind: deopt-soft, reason: Insufficient type feedback for call): begin. deoptimizing 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)>, opt id 1, bytecode offset 102, deopt exit 6, FP to SP delta 192, caller SP 0x00016aeea420, pc 0x00012884c8ac]
[bailout (kind: deopt-eager, reason: not a Smi): begin. deoptimizing 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)>, opt id 0, bytecode offset 2, deopt exit 0, FP to SP delta 32, caller SP 0x00016aeea328, pc 0x00012884ca54]
[marking 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> for optimized recompilation, reason: small function]
[compiling method 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> (target TURBOFAN) using TurboFan]
[marking 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)> for optimized recompilation, reason: hot and stable]
[optimizing 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> (target TURBOFAN) - took 0.083, 0.208, 0.042 ms]
[completed optimizing 0x0abcdcc14701 <JSFunction add (sfi = 0x3092d43b1ae1)> (target TURBOFAN)]
[compiling method 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)> (target TURBOFAN) using TurboFan OSR]
[optimizing 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)> (target TURBOFAN) - took 0.125, 0.333, 0.000 ms]
[bailout (kind: deopt-soft, reason: Insufficient type feedback for generic named access): begin. deoptimizing 0x0abcdcc133f1 <JSFunction (sfi = 0x3092d43b1a29)>, opt id 3, bytecode offset 127, deopt exit 4, FP to SP delta 192, caller SP 0x00016aeea420, pc 0x00012884ce24]
PerformanceMeasure {
  name: 'My Special Benchmark',
  entryType: 'measure',
  startTime: 28.826040983200073,
  duration: 268.1276671886444,
  detail: null
}
如果我们在代码中加入V8指令，禁用TurboFan编译优化：
"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
const perf_hooks_1 = require("perf_hooks");
let iterations = 1e8;
const a = 1;
const b = 2;
const add = (x, y) => x + y;
perf_hooks_1.performance.mark("start");
%NeverOptimizeFunction(add);
while (iterations--) {
    add(a, b);
}
iterations = 1e8;
// add("foo", "bar");
while (iterations--) {
    add(a, b);
}
perf_hooks_1.performance.mark("end");
perf_hooks_1.performance.measure("My Special Benchmark", "start", "end");
const [measure] = perf_hooks_1.performance.getEntriesByName("My Special Benchmark");
console.log(measure);
PerformanceMeasure {
  name: 'My Special Benchmark',
  entryType: 'measure',
  startTime: 25.05691695213318,
  duration: 885.5741250514984,
  detail: null
}
由此可见，如果完全不适用TurboFan进行编译优化，单纯通过解释器执行字节码速度将慢得多。
到这里我们应该对TurboFan所做的编译优化有一个基本了解了。
 
Relate Posts
2023-10-08
函数柯里化
2023-05-27
重新学习TypeScript（基础）
2022-11-12
浏览器缓存
2022-08-01
使用ESLint、Prettier统一团队代码
2022-04-01
弹性盒子布局基础
2022-07-14
Intersection Observer API
函数柯里化重新学习TypeScript（基础）
目录
0%
Louis K
一个游手好闲的懒人