On Mon, Aug 17, 2020 at 12:36:51PM -0700, Kees Cook wrote:
> On Mon, Aug 17, 2020 at 11:08:54AM +0200, David Sterba wrote:
> > On Sat, Aug 15, 2020 at 10:09:24AM -0700, Kees Cook wrote:
> > > +static inline bool __must_check __must_check_overflow(bool overflow)
> > > +{
> > > + return unlikely(overflow);
> >
> > How does the 'unlikely' hint propagate through return? It is in a static
> > inline so compiler has complete information in order to use it, but I'm
> > curious if it actually does.
>
> It may not -- it depends on how the compiler decides to deal with it. :)
In theory yes, in practice, the compilers will ignore this macro.
And if you success to force compiler to use this macro, it won't give
any real performance advantage.
We (RDMA) tried very hard to see any performance gain by instrumenting
code with likely/unlikely in our performance critical data path both in
user space and kernel. The performance results were statistically equal.
If you are interested, we had a very intense discussion about it when
likely/unlikely can still be usable (hint random input).
https://lore.kernel.org/linux-rdma/20200807160956.GO4432@unreal
Thanks
>
> > In case the hint gets dropped, the fix would probably be
> >
> > #define check_add_overflow(a, b, d) unlikely(__must_check_overflow(({ \
> > typeof(a) __a = (a); \
> > typeof(b) __b = (b); \
> > typeof(d) __d = (d); \
> > (void) (&__a == &__b); \
> > (void) (&__a == __d); \
> > __builtin_add_overflow(__a, __b, __d); \
> > })))
>
> Unfortunately not, as the unlikely() ends up eating the __must_check
> attribute. :(
>
> --
> Kees Cook
