On Wed, 2022-03-09 at 13:30 -0800, Andrew Pinski wrote:
> On Wed, Mar 9, 2022 at 1:25 PM David Malcolm via Gcc
> <gcc@gcc.gnu.org> wrote:
> > 
> > We gained __attribute__ ((access, ...)) in GCC 10:
> >   
> > https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html
> > which identifies one of the pointer/reference arguments of a
> > function
> > as being accessed according to an access-mode: read_only,
> > read_write,
> > write_only, or none.
> > 
> > We also have __attribute__ ((nonnull)) to indicate that a function
> > argument (or all of them) must be non-NULL.
> > 
> > There doesn't seem to be a relationship between these in the
> > implementation, but it strikes me that almost anywhere that a user
> > might use the "access" attribute, that parameter is probably going
> > to
> > be required to be nonnull - though perhaps there are cases where
> > APIs
> > check for NULL and reject them gracefully?
> 
> No, I think they are separate. The access just says these access
> attributes are read only, write only, read-write or don't access what
> the pointer points to; it does not say they have to be read or
> written
> to.
> I think it is a bad idea to connect the two ideas because you could
> have some cases where an argument is optional but is only read from;
> or is only written to (there are many in GCC sources even).

Thanks for the clarification...

> 
> Thanks,
> Andrew Pinski
> 
> > 
> > Might we want to somehow make __attribute__ ((access, ...)) imply
> > __attribute__ ((nonnull))?  (for non "none" access modes, perhaps?)
> > 
> > If so, one place to implement this might be in tree.cc's
> > get_nonnull_args, and have it add to the bitmap any arguments that
> > have an appropriate access attribute.
> > 
> > get_nonnull_args is used in various places:
> > - validating builtins
> > - in ranger_cache::block_apply_nonnull
> > - by -Wnonnull (in pass_post_ipa_warn::execute)
> > - by -Wanalyzer-possible-null-argument and -Wanalyzer-null-
> > argument;
> > I'm tracking the failure of these last two to make use of
> > __attribute__
> > ((access)) in PR analyzer/104860.
> > 
> > So do we:
> > 
> > (a) leave it up to the user, requiring them to specify
> > __attribute__
> > ((nonnull)) in addition to  __attribute__ ((access, ...))

...so that's (a) then.

I think it might be more user-friendly to be explicit about this in the
documentation, maybe something like the attached?

(not yet fully tested, but seems to build)

Dave




> > 
> > (b) leave it up to the individual sites in GCC that currently make
> > use
> > of get_nonnull_args to add logic for handling   __attribute__
> > ((access,
> > ...))
> > 
> > (c) extend get_nonnull_args
> > 
> > ?
> > 
> > Thoughts?
> > Dave
> > 
>