-findirect-dispatch performance

Mon Sep 29 23:25:00 GMT 2003

On Tuesday, Sep 30, 2003, at 04:47 Pacific/Auckland, Andrew Haley wrote:
> Now that the gc bitmap generation bug is fixed, I've been measuring
> the performance difference with -findirect-dispatch.
>> Bottom line: it seems to me that, as far as I can measure, there is no
> perforance difference.

Excellent ;-). This is consistent with what we saw measuring 
-findirect-dispatch for virtual calls.
> The difference between these figures is less than the measurement
> noise. So, I'd have to put my machine into single user mode to
> determine which was fastest.

oprofile could be useful here. It can measure the actual cycles used by 
the app, as report by the CPU's performance counters. I'm not sure if 
its 100% accurate, but should be much more so than using timers.
> Now, the embedded CaffeineMark is not the best Java benchmark program,
> so if anyone can suggest anything better I'm interested. However, on
> the basis of these timings I think we can agree that this is good
> enough.

Yes. I guess it depends how many cross-class field accesses the 
benchmark is doing, though.
Are you planning to check this in soon?
Regards
Bryce.