587255 - Make Math.random faster

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Mandelin [:dmandelin]]

We are currently 2x slower than JSC on this microbenchmark:

  function f() {
    for (var i = 0; i < 10000000; ++i) {
      Math.random(1);
    }
  }

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

I did the comparison again with the body of Math.random mostly commented out on both sides, and it seems we are about the same speed then. So, the difference is probably in the RNG itself.

JM wanted SS win: 2 ms

Comment 1

[Ryan VanderMeulen [:RyanVM]]

Interp: 1470
TM: 1476
JM: 606
JM+TI: 582
d8: 249

About 2x slower than v8. I don't have a JSC shell available to test against.

Comment 2

[Florian Bender]

This should depend on Bug 774364 or be closed.

Comment 3

[Hugh Nougher [:Hughman]]

I'm guessing this might as well be closed?

Comment 4

[Tooru Fujisawa [:arai]]

Tested with modified code.

code:
  var n = 0;
  function f() {
    for (var i = 0; i < 100000000; ++i) {
      n += Math.random();  // use return value, to avoid getting optimized away
    }
  }
  var t0 = new Date;
  f();
  console.log(new Date - t0);

results:

x86_64
  Firefox 40      : 790
  Firefox Nightly : 450
  Chrome          : 1122 (4404 on 2nd run, not sure why...)
  Chrome Canary   : 1494 (5397)
  Safari          : 1366
  WebKit          : 963

x86
  Firefox 40      : 1549
  Firefox Nightly : 782

I think it's fast enough :)