[PATCH] Tweaks for constant-equality-comparisons in BPI

[Hal Finkel]

----- Original Message -----
> 
> On 28 October 2013 14:25, Hal Finkel < hfinkel at anl.gov > wrote:
> 
> 
> 
> 
> 
> - On the other hand, if there is any validity to it, we're currently
> doing exactly the opposite of what we should be doing. Maybe doing
> exactly the opposite is the wrong approach as well, because the
> overall bias is very weak, and the best thing to do is really to do
> nothing
> 
> 
> I think that's a good question that needs answering.
> 
> 
> But the fact that your test-suite results didn't show any radical
> change in behaviour (not that it's good for it, but), I don't think
> that either choice will make a massive change on user code. It might
> be great for benchmark A or B, but in general, it'd get lost in
> noise. In a nutshell, to change the behaviour, we need a good
> reason, either that the current is broken, or that the new is
> better, but we don't know that yet.
> 

Agreed (this was my motivation for separating the patches). To be fair, I think that a lot of the current implementation is intuition based, so making more intuition-based changes where the data is ambiguous would not necessarily be out of place. However, I agree that we should not make changes without evidence of a positive benefit.

> 
> 
> 
> 
> If you could, I'd appreciate it. I'll also provide an update with the
> BG/Q timings.
> 
> 
> 
> I'll check the test-suite and others on my Chromebook and will get
> back to you.

Thanks!

 -Hal

> 
> 
> cheers,
> --renato

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory