https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63791
Marcus Kool <marcus.kool at urlfilterdb dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|RESOLVED |REOPENED
Resolution|INVALID |---
--- Comment #6 from Marcus Kool <marcus.kool at urlfilterdb dot com> ---
(In reply to H.J. Lu from comment #5)
> (In reply to Marcus Kool from comment #2)
> >
> > To resume, gcc 4.8.4 and gcc 4.9.2 produce code that can be optimised
> > further, and gcc 5.1.0 produces even slower code which means that the
> > implementation of *_set1_epi8() is slower/much-slower than that it can be.
>
> That is done on purpose. Add -mtune=intel will get what you want.
I do not understand why with the "-O3 -mavx2" flags gcc "on purpose" produces 4
instructions and the compiler flag "-mtune=intel" is mandatory to improve this.
The -mavx2 flags says that the platform is AVX2 (i.e. Haswell and better).
I read the man page of 5.1.0 and must say that I am frustrated with the
explanation given for the flag "-mtune=intel". I always thought that -O3 was
the way to tell the compiler "do your best to make the code run fast" and up to
gcc 4.9.2, "-mavx2" was good enough to say to generate code for AVX2.
Resuming, for gcc previous to 5.x, "-O3 -mavx2" did the trick and starting with
5.x "-O3 -mavx2 -mtune=intel" does the trick. And the additional caveat is
that the behaviour of "-mtune=intel" can change in the future.
What is the reasoning behind this ?
