In the last episode (May 25), Garrett Wollman said:
> Just for completeness, I did one final run of HINT with just `-O'
> specified (our usual default). `-O' results in significantly better
> integer performance than `-O4'. (Floating-point performance is just
> the opposite.)
>
> This suggests that compiling the world with `-O' levels higher than
> one is probably a bad idea. (The generated assembly is identical from
> `-O2' to `-O4'.) The `-O2' code appears to be less efficient at
> register allocation; about twice as much stack temporary space is
> required.
-O4 doesn't exist in egcs (or it didn't a month or so ago). According
to the source, -O2 enables all optimizations except -funroll-all-loops,
and all -O3 does is enable -funroll-all-loops.
I'd like to see separate runs, one with each -march= option (i386,
i486, i586, i686), so see if those many any difference at all.
-Dan
To Unsubscribe: send mail to majord...@freebsd.org
with "unsubscribe freebsd-current" in the body of the message
