On Thu, 15 Aug 2002, John M McIntosh wrote:

> >On Thursday 15 August 2002 12:43 pm, Joshua 'Schwa' Gargus wrote:
> >>  I recently built a VM with gcc-3.2, and had similar results.
> >
> >So what are you using for optimization flags?
> >
> >Perhaps the
> >-O2 -funroll-loops
> >
> >isn't enough...
> 
> Well I've opened a problem with Apple on this issue of how GCC 3.x 
> creates poorer code.

Taking this up with the people in charge of the Gnu Compiler Collection
(at RedHat now?) might produce faster results (or at least a "we know,
we're working on it...").

> Mind I don't see a quick solution. For powerPC 
> the optimizer creates very different code. It just runs slower.

You might try turning scheduling off entirely.  (Unscheduled code for PPC
in 2.9x is pretty close to what the common model wants, and 750s should
be able to fix most shortcomings in common model code on-chip.)  But
bad scheduling shouldn't produce a factor of 2 difference in speed...

> This might be true for all platforms, one would think that i386 code 
> should be faster.

I remember (from quite a while ago) that the big change in gcc3 was going
to be live range splitting.  Maybe localXX are being spilled and reloaded
(disaster!) by the new optimiser?  If these are spilled in the wrong
places then it could explain a factor of 2 in speed.  Worth taking a look
at.

> Also it wasn't clear if the gnuify step ran and the interp.c was altered?

This comes before the compiler and should be completely independent of its
version number (assuming "gcc -v" still prints "gcc" somewhere in its
output).

Ian