Performance regression 4.3 vs 4.4/4.5

[Salvatore Filippone]

Il giorno gio, 19/11/2009 alle 06.17 -0800, Tim Prince ha scritto:
> Salvatore Filippone wrote:
> > To expand on my previous message, I have built a self contained test
> > demonstrating the problem. 
> 
> Do you get similar results if you use a sane method to obtain the 
> operands, particular if you try the same data in each case?

I am not entirely sure what you mean. The subroutine is called in an
iteration loop, where the  outer code makes adjustments to the values in
X according to the outcome of the calculation, so if you want me to try
calling the routine multiple times with the same input, it would not fit
the original usage. As you may have guessed, in the original code (and
in the first version I posted) there is a cache array, so that the
effect is present when going beyond the cache array size, which is why I
took it out entirely of the self-contained case. 
And, yes, in the original code the arrays are initialized based on some
geometry information (not random numbers), but the effect is the same
(once you take out the cache array, that is )


Salvatore