This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: gcc compile-time performance
- From: Richard Earnshaw <rearnsha at arm dot com>
- To: Richard Henderson <rth at redhat dot com>
- Cc: Jan Hubicka <jh at suse dot cz>, Robert Dewar <dewar at gnat dot com>, Richard dot Earnshaw at arm dot com, gcc at gcc dot gnu dot org
- Date: Mon, 20 May 2002 18:22:43 +0100
- Subject: Re: gcc compile-time performance
- Organization: ARM Ltd.
- Reply-to: Richard dot Earnshaw at arm dot com
> On Mon, May 20, 2002 at 05:08:41PM +0200, Jan Hubicka wrote:
> > Yes, that is alternative. Will it be considerably faster than the current
> > FP emulation library doing the same?
>
> Maybe.
>
> Probably less so if real.c is rewritten so that it can use 32/64
> arithmetic instead of 16/32. This gets more work done per cycle.
> IIRC, Roman Lechtchinsky is looking into that because alpha-unicosmk
> doesn't have a 16 bit type.
>
> Another alternative is to use a completely different, more efficient
> soft fp implementation.
There's still a lot of fiddling with NaNs and DeNorms etc in that code
which just isn't needed in this case. Checking for those corner cases can
considerably slow things down.