Re: [patch] Removal of redundant prefetches during loop unrolling

[Paolo Bonzini <paolo dot bonzini at lu dot unisi dot ch>]

> > Also can you comment on what happens with tree complete unrolling and
> > someone having a __buitin_prefetch there?

> 2) It is doubtful that it would be very profitable to have
>    __buitin_prefetch in a loop that iterates so few times it gets
>    completely unrolled.  There must be enough time for a prefetch to
>    finish, which is almost impossible in this type of loop, unless
>    it contains a function call in that a lot of time is spent (and
>    in this case, leaving the redundant prefetches should not make much
>    harm).

Or if the completely-unrolled loop is the innermost of two loops, and 
the outermost loop that is not unrolled.  (I think that in this case the 
user could put his prefetch in the outermost loop, since he cares enough 
to write prefetches manually).

Paolo