James Padolsey ：Partial loop “unrolling”

 

Ps：本人不甚解，这样做的意义。

Posted: 13 Nov 2009 01:31 AM PST

In Thomas Fuchs’ latest JavaScript performance presentation he talks about the speed gains that can be experienced by using “unrolled” loops.

A conventional loop:

for (var i = 0; i < 10; ++i) {
    doFoo(i);
}

The “unrolled” version of that loop:

var i = 0;
doFoo(i++); doFoo(i++); doFoo(i++); doFoo(i++); doFoo(i++);
doFoo(i++); doFoo(i++); doFoo(i++); doFoo(i++); doFoo(i++);

A partially unrolled version:

for (var i = 0; i < 10; ) {
    doFoo(i++);
    doFoo(i++);
    doFoo(i++);
    doFoo(i++);
    doFoo(i++);
}

Interestingly, speed gains can be experienced dependent on the loop size, albeit marginal at best. I thought about ways to build this into a clever forEach function and came up with something that ‘pre-compiles’ functions containing partially unrolled loops. Have a look:

var forEach = (function() {
 
    var fns = [],
        callers = "true",
        numberFn = 10,
        i = 1;
 
    for ( ; i <= numberFn; ++i ) {
        callers += "&&f(a[++i])!==false";
        fns[i] = new Function("a", "f", "l", "var i=0;while (i<l) {"+callers+"}");
    }
 
    return function( array, fn ) {
 
        var len = array.length,
            n = numberFn, i;
 
        while (i = n--) {
            if ( len % i === 0 ) {
                return fns[i](array, fn, len);
            }
        }
 
    };
 
})();

This function will run one of 10 ‘pre-compiled’ functions on the passed array, dependent on the highest factor of the array’s length. I’m only creating 10 different functions in this example, you could create more.

If you were to pass an array with a length of 14, then fns[7] would be used, since 7 is the highest available factor (the highest number below 10 that 14 can be divided by, to gain a whole number). fns[7] looks something like this:

function anonymous(a, f, l) {
    var i = 0;
    while (i < l) {
        true &&
            f(a[++i]) !== false &&
            f(a[++i]) !== false &&
            f(a[++i]) !== false &&
            f(a[++i]) !== false &&
            f(a[++i]) !== false &&
            f(a[++i]) !== false &&
            f(a[++i]) !== false;
    }
}

The !== false part is used to create the effect of loop-breaking. Notice that the success of this boolean expression is depended upon to continue the chain of expressions (a && b && c)

I’ve only tested it briefly, and to be honest, there doesn’t seem to be a notable benefit. In IE, I can see a bit of improvement over the conventional forEach implementation but only if I’m using arrays with 1000+ lengths. I think this would only be useful in situations where you absolutely have to squeeze every inch of potential performance out of your app. Anyway, it’s still pretty interesting, I wonder what other fancy things can be created by using pre-compiled functions.