[Bug tree-optimization/77689] Missing vectorization lead to huge performance loss
morandidodo at gmail dot com
gcc-bugzilla@gcc.gnu.org
Thu Sep 22 13:55:00 GMT 2016
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77689
--- Comment #8 from morandidodo at gmail dot com ---
(In reply to Uroš Bizjak from comment #7)
> (In reply to morandidodo from comment #6)
>
> > However, I just made some tests, and it seems that the performance lost is
> > particularly related to std::log. If something else is being used, like
> > std::sqrt for example, the code compiled with gcc is still slower, but less.
>
> Adding -mveclibabi=svml should vectorize calls to these functions, but you'd
> have to link with SVML vector library then.
>
> icc is doing this by default:
>
> 400e0f: e8 ac 16 00 00 callq 4024c0 <__svml_log4>
>
> but gcc doesn't:
>
> 40078b: e8 00 ff ff ff callq 400690 <log@plt>
Actually, I tried it, but it does not seem to work. Indeed, using
-mveclibabi=svml -lsvml -I(path-to-intel-lib-dir) does not seem to change the
call address. If you have got any advice, it is appreciated ;)
More information about the Gcc-bugs
mailing list