This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch] rtlopt branch merge part 6 - speed improvement of branch predictions
> >I wonder if the following would be better. Thoughts?
I have measured compilation time of combine.c on athlonXP 1600,
branch prediction with real.c took 0.23 s,
with sreal.c 0.11 s.
On platforms which do not have multiply and divide instruction
it probably runs slower.
> Does this ensure identical fp math for this on all
> platforms?
>
> I think it doesn't, even if all the world uses VAX^H^H^HIEEE
> math; like, denormal handling isn't always identical.
For less than 32-bit machines the precision would be smaller than
32-bit, for other machines which are not 32-bit nor 64-bit
the precission is higher than 32-bit.
It should not get denormal numbers (which would be rounded to 0) at all,
because the maximum and minimum exponent is pretty high (+- INT_MAX / 4)
and thus it should not be reached.
Did I answer your questions?
Josef