643565 - Analyze v8-crypto performance

Status: RESOLVED INVALID

(Core :: JavaScript Engine, defect)

Description

[David Mandelin [:dmandelin]]

We do fairly well with tracing, but v8/nocs is 2x faster than our mjit. Why?

Comment 1

[David Mandelin [:dmandelin]]

Attachment: Profiler output

Comment 2

[David Mandelin [:dmandelin]]

The 2x difference given in comment 0 is for 64-bit, which means that it actually compares SM/x64 with v8/nocs/x86. What it means is that the code we generate for 64-bit is not as good as 32-bit, but we already knew that.

On 32-bit Linux both ways, we are only 12% slower than v8/nocs here. I suspect most of that difference is captured by this microbenchmark:

    function f() {
        for (var i = 0; i < 6600000; ++i) {
            var n = 12;
            while (--n) {
            }
        }
    }

    var t0 = new Date;
    f();
    print(new Date - t0);

v8/nocs is 4x faster than js -a -m on this. If I take out the inner loop and instead run the outer loop 12x as many iterations, they are 2x faster. So, we have problems with our basic loop performance, which seem to get worse with nested loops. We know that |i++| and such are deoptimized; it might be just that, and maybe the problem bites us harder with nested loops.

Comment 3

[Till Schneidereit [:till]]

We're pretty much on par with v8 for the microbenchmark here, and more recent analysis is happening in bug 768572.

Filter on jm-cleanup.