This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: double alignmwent/benchmark data


>coming back to the "benchmark" program posted recently, I wanted to add some
>more datapoints, showing how irregular fp performance currently is:

Thanks for the info!  However, having tried out -mstack-align-double
for myself and found some questionable results, I'm not quite sure
it currently represents what we can actually achieve via this
approach until I look into it further.  (For example, it didn't seem
to help much with my `align' program, which has the express purpose
of trying to uncover cases where double-precision stuff is unaligned.)

I've given up on this issue for 1.1, but still want to look into it
further soon, to see if I can help provide a more complete solution,
or at least more info.  For example, I'm concerned g77 itself still
isn't handling alignment issues properly, though I don't really
have any evidence for that right now.  I'd like to step through the
relevant code when -mstack-align-double is specified and see what
g77 is actually doing (and what the back-end info, e.g. the TYPE_DECLs,
actually look like).

In the meantime, I suggest nobody take the measurements involving
-mstack-align-double (or presumably the other new options introduced
by the patch that's been floating around, from pgcc I guess) too
seriously in terms of concluding one way or another whether it
helps or hurts.  -malign-double seems to work fine, though -- it
has the generally-intended effects on egcs 1.1, last I checked,
according to my `align' program.

        tq vm, (burley)


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]