JavaScript `reduce` 性能

Question

我最近花了一些时间来研究传感器(函数式编程中的工具，旨在提高性能而不损失代码的可读性/灵 active )，当我开始测试它们的实际速度时，我得到了一些非常令人失望的结果。考虑:

const inc = x => x + 1;

const isEven = x => x % 2 === 0;

// simplest, shortest way I would be comfortable with if performance wasn't an issue

const mapFilter = xs => xs.filter(isEven).map(inc);

// transducers way

// function composition
const compose = (...fns) => x => fns.reduceRight((y, f) => f(y), x);

const map = f => step => (a, c) => step(a, f(c));
const filter = p => step => (a, c) => (p(c) ? step(a, c) : a);

// basic reducer for building array
const build = (acc, x) => {
  acc.push(x);
  return acc;
};

// transducer, it doesn't create intermediate arrays hence should theoretically be faster 
const transducers = xs =>
  xs.reduce(compose(filter(isEven), map(inc))(build), []);

// native loop for comparison
const nativeLoop = data => {
  const result = [];
  const l = data.length;
  for (let i = 0; i < l; i++) {
    const x = data[i];
    if (isEven(x)) result.push(inc(x));
  }
  return result;
};

const data = Array(1000).fill(1);

const base = ["simplest, chained map and filter", () => mapFilter(data)];
const alternative = ["composed transducers", () => transducers(data)];
const alternative2 = ["native loop", () => nativeLoop(data)];

/* console.log(Benchmark) */
console.log("Running benchmarks....");

const suite = new Benchmark.Suite();
suite
  .add(...base)
  .add(...alternative)
  .add(...alternative2)
  .on("cycle", function(event) {
  console.log(String(event.target));
})
  .on("complete", function() {
  console.log("Fastest is " + this.filter("fastest").map("name").join(", "));
})
// run async
  .run({ async: true });

<script src="https://cdnjs.cloudflare.com/ajax/libs/lodash.js/4.17.15/lodash.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/benchmark/2.1.4/benchmark.min.js"></script>

我希望表演的顺序是

native 循环 > 传感器 > 链式映射/过滤器

同时，除了比其他任何方法都快得多的原生方法之外，令我非常惊讶的是，reduce/transduce 方法比使用 map/filter 和创建中间数组要慢得多(更慢，就像 Chrome 中的一个数量级)。有人可以向我解释一下这个结果的由来吗？

Answer 1

您的基准测试是错误的，因为您在每次运行时都构建了一个新的传感器链。

const inc = x => x + 1;

const isEven = x => x % 2 === 0;

// simplest, shortest way I would be comfortable with if performance wasn't an issue

const mapFilter = xs => xs.filter(isEven).map(inc);

// transducers way

// function composition
const compose = (...fns) => x => fns.reduceRight((y, f) => f(y), x);

const map = f => step => (a, c) => step(a, f(c));
const filter = p => step => (a, c) => (p(c) ? step(a, c) : a);

// basic reducer for building array
const build = (acc, x) => {
  acc.push(x);
  return acc;
};

// transducer, it doesn't create intermediate arrays hence should theoretically be faster
const reducer = compose(filter(isEven), map(inc))(build);
const transducers = xs => xs.reduce(reducer, []);

// native loop for comparison
const nativeLoop = data => {
  const result = [];
  const l = data.length;
  for (let i = 0; i < l; i++) {
    const x = data[i];
    if (isEven(x)) result.push(inc(x));
  }
  return result;
};

const data = Array(1000).fill(1);

const base = ["simplest, chained map and filter", () => mapFilter(data)];
const alternative = ["composed transducers", () => transducers(data)];
const alternative2 = ["native loop", () => nativeLoop(data)];

/* console.log(Benchmark) */
console.log("Running benchmarks....");

const suite = new Benchmark.Suite();
suite
  .add(...base)
  .add(...alternative)
  .add(...alternative2)
  .on("cycle", function(event) {
  console.log(String(event.target));
})
  .on("complete", function() {
  console.log("Fastest is " + this.filter("fastest").map("name").join(", "));
})
// run async
  .run({ async: true });

<script src="https://cdnjs.cloudflare.com/ajax/libs/lodash.js/4.17.15/lodash.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/benchmark/2.1.4/benchmark.min.js"></script>

如您所见，转换器确实比链式 map 和 filter 方法更快。