Re: SIGFPE and -mieee
Kelledin <kelledin+DAXP@skarpsey.dyndns.org> writes:
> Also, one other detail I forgot: what about EV6 and higher? My
> understanding is that on Alpha processors >=EV6, -mieee is
> either unnecessary, or the extra code it introduces comes at
> negligible cost.
The latter. The trap barriers introduced by -mieee are discarded as
nops very early in the pipeline, since the ev6 has precise traps.
Also, it can do some more denormal arithmetic in hardware. Example:
double x = 1.0;
for (int i = 0; i < 1000000000; ++i)
x *= 2;
no -mieee SIGFPE
-mieee on ev5 165.81 seconds (inf arithmetic emulated by kernel)
-mieee on ev6 1.03 seconds
If you use -mcpu=ev6, -mieee should be totally free since gcc doesn't
insert any trap barriers then.
--
Falk
Reply to: