Igor Schein on Mon, 10 Jan 2000 11:13:32 -0500

[Date Prev] [Date Next] [Thread Prev] [Thread Next] [Date Index] [Thread Index]

Re: optimization flags

On Fri, Jan 07, 2000 at 09:45:41PM -0500, Ilya Zakharevich wrote:
> > And finally, -frunroll-loops and -funroll-all-loops could also be
> > beneficial, but I will need to verify that.
> How do you do it?  I found Ultra5 absolutely useless for benchmarking,
> the speed of the same program on the same "unused" machine varies up
> to 90% when phase of sun changes.  Sun's engineers say that this is
> due to "cache coloring", which may be switched on and off in PROM (but
> switching it off decreases performance).  Does somebody know what they
> are talking about?
> Ilya

Not really, but I don't have this problem on Ultra60.  Binary compiled
without -funroll-loops consistently outperformed binary compile with
-funroll-loops, so we shouldn't use this flag.

To summarize, here's the proposed combination of flags to use in Configure
for Solaris:

-O3 -DGCC_INLINE -mv8 -fomit-frame-pointer

Similar goes for AlphaLinux:

-O3 -DGCC_INLINE -fomit-frame-pointer

Now for IntelLinux it's more compilated, because supposedly there're
optimizations levels as high as -O9 ( nowhere documented ).  So let it
stay as it is now, until that clears up.