608867 - Tune JIT heuristics for v8-crypto

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Cavan]

User-Agent:       Mozilla/5.0 (Windows NT 6.1; WOW64; rv:2.0b8pre) Gecko/20101101 Firefox/4.0b8pre
Build Identifier: 

I was looking at the AWFY individual graphs and noticed that -mjp is slower than both JM or TM alone. Isn't the goal of heuristics to be at least as fast as one of the JIT engines?

Reproducible: Always

Comment 1

[Cavan]

I should have checked this before but it does not seem to occur in win32 browser builds. With today's m-c win32 nightly I get these v8 crypto scores on my system.

-m -j -p: 5200
-m: 4300
-j: 5700

AWFY still shows -m -j -p being the worst so maybe it only occurs in the shell or only on OSX?

Comment 2

[David Mandelin [:dmandelin]]

Has anyone tried this in browser on OSX? It shouldn't block unless it affects the browser.

Comment 3

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

(In reply to comment #2)
> Has anyone tried this in browser on OSX? It shouldn't block unless it affects
> the browser.

I think there's a real problem here, where -mjp is slower than -m or -mj. It's a little hard to track down because it's only a problem on slower machines. I see it on a Core i5, but not an i7. And the effect on an i5 is less than what we see on AWFY.

The problem seems to be mostly fixed on my i5 by the patch in bug 609212. Once that patch lands, we can see what AWFY says.

Comment 4

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

To the extent that I can measure, this is fixed by the patch landed in bug 609212. Unfortunately, I can't figure out how to get individual scores from awfy2. However, here are the times I'm not getting:

-m: 2241ms
-j: 1925ms
-mj: 1900ms
-mjp: 1892ms

The old -mjp time was 1978ms.