https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85740
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|middle-end |target
--- Comment #11 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So GCC heurstics for figuring out a[i] > m if it is likely or unlikely is
always going to be "wrong". My bet if the test vector was always incrementing,
the version without the __builtin_expect would be faster.
x86_64 has a loop cache which is why the version without the __builtin_expect
seems faster.
