svn commit: r334545 - in head/sys: contrib/zstd/lib/freebsd kern netinet/libalias sys

[Mateusz Guzik]

On Wed, Jun 6, 2018 at 12:52 AM, Gleb Smirnoff <glebius at freebsd.org> wrote:

> On Tue, Jun 05, 2018 at 06:35:37PM +0200, Mateusz Guzik wrote:
> M> > On 06/02/2018 17:20, Mateusz Guzik wrote:
> M> > > +#ifdef _KERNEL
> M> > > +#define      malloc(size, type, flags) ({
>   \
> M> > > +     void *_malloc_item;
>    \
> M> > > +     size_t _size = (size);
>   \
> M> > > +     if (__builtin_constant_p(size) && __builtin_constant_p(flags)
> &&\
> M> > > +         ((flags) & M_ZERO)) {
>    \
> M> > > +             _malloc_item = malloc(_size, type, (flags) &~
> M_ZERO);  \
> M> > > +             if (((flags) & M_WAITOK) || _malloc_item != NULL)
>    \
> M> > > +                     bzero(_malloc_item, _size);
>    \
> M> > > +     } else {
>   \
> M> > > +             _malloc_item = malloc(_size, type, flags);
>   \
> M> > > +     }
>    \
> M> > > +     _malloc_item;
>    \
> M> > > +})
> M> > > +#endif
> M> >
> M> > Mateusz,
> M> >
> M> > Thank you for this and for all of your performance work.  It is all
> very
> M> > interesting stuff.
> M> >
> M> >
> M> Thank you for the kind words. It is positive feedback like this which
> M> keeps me going!
>
> Btw, what was the point of checking flags || result? Most places in kernel
> ignore flags and just test result regerdless of M_WAITOK/M_NOWAIT.
>
> The result is already in a register, why do you think checking for absense
> of M_WAITOK is faster that checking for !NULL _malloc_item?
>

This part is only reachable if flags are known at compilation time. If they
contain
M_WAITOK, the flag check will get elided along (we know for a fact it
passes)
and subsequently the NULL check will be short circuited, iow for known
M_WAITOK|M_ZERO flags this is:

_malloc_item = malloc(_size, type, flags & ~ M_ZERO);
bzero(_malloc_item, _size);

-- 
Mateusz Guzik <mjguzik gmail.com>