On 23 November 2015 at 12:49, Joe Stringer <joestrin...@nicira.com> wrote: > On 23 November 2015 at 10:06, Ben Pfaff <b...@ovn.org> wrote: >> When Joe added these types I assumed that he used the unconventional >> prototypes for hton128() and ntoh128() because the return value >> convention was inefficient. If GCC and Clang actually optimize the use >> of a return value in some kind of sensible way then I agree that the >> usual convention is nicer. >> >> Joe, did you have another reason? > > This was mostly done based on an assumption that this was more > optimal, rather than actually digging into the compiled code and > seeing that it was generated differently. > > Looking now, before this patch vs. after on my 64-bit system.. > > GCC-4.9: hton128/ntoh128 require one less MOV with this patch, but > calling conventions (in format_u128) require 3 extra MOV (+4 MOV, -1 > LEA) for format_u128(). > Clang-3.7: hton128/ntoh128 are roughly equivalent, although with this > patch they use some MOVUPS/MOVAPS instructions for 128-bit moves. > Calling conventions seem to require as much as 6-12 (!) extra MOVs > however, details below.
<snip> Woops, looks like I missed compiler optimizations. They look much closer together with --O2, so I think this patch makes sense to apply without reservation. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
