Re: [RFC] bloody mess with __attribute__() syntax

[Josh Triplett <josh@freedesktop.org>]

Al Viro wrote:
> 	We have a fun problem and for a change it's not sparse fault.
> It's gcc folks' one.  Basically, __attribute__((...)) behaves in
> an idiotic way and it's an intentional (and documented) behaviour.
> In declaration of form
> 	T __attribute__((foo)) **v;
> the attribute applies to v, not to **v.  IOW, in that position it
> behaves (regardless of the nature of attribute) as storage class,
> not as a qualifier.  Even if the same attribute can be used in
> 	T * __attribute__((foo)) *v;
> where it will apply to *v.  Intended way to have it apply to **v is
> 	T (__attribute__((foo)) **v);
> 
> To put it mildly, that blows.  Note that qualifiers can *not* behave
> that way - direct declarator can not expand to (<qualifier> <something>).
> I.e. if you replace __attribute__((foo)) with qualifier in the
> above, you'll get invalid syntax.

Wow.  Insane.  So these all declare the same type:
__attribute__((foo)) T *v;
T __attribute__((foo)) *v;
T *__attribute__((foo)) v;
?  Specifically, they point to a foo-T, for convenient shooting?

> Now, that idiocy would be none of our concern, if not for the fact
> that noderef and address_space() are definitely supposed to imitate
> qualifiers.

context also represents a qualifier; the position of the qualifier should
determine things like whether you want to enforce the context when you access
a pointer or dereference a pointer.

> If anybody seriously suggests switching to syntax
> like
> 	int (__user *p);
> all over the place, well...

Definitely not an option.

> Note that gcc rules for __attribute__() (and that's the only source
> of rules we _have_ for the damn thing) clearly say that
> 	int __user *p;
> is the same thing as
> 	int *__user p;
>
> Now, we could declare gcc people responsible for that turd rejects
> of Vogon Construction Fleet and handle the damn thing sanely.
> The first part is clearly the right thing to do, but the second one...
> Can't do without breaking gccisms using __attribute__.  E.g.
> 	int (__attribute__((mode(__pointer__))) *p);
> is a gcc way to say "pointer to integer type equivalent to intptr_t" and
> 	int __attribute__((mode(__pointer__))) *p;
> is exactly the same thing as
> 	int *p;
> since the damn attribute applies to the entire type here (and is obviously
> a no-op).
>
> Frankly, I would rather add a new primitive (__qualifier__) mirroring the
> __attribute__, but acting like real qualifiers do.  And switched the
> noderef et.al. to it.

Something like that sounds vaguely reasonable.  It should allow the same set
of attributes, and just change what they apply to.  To use your example,
T __qualifier__((foo)) *v;
and
T (__attribute__((foo)) *v);
would mean the same thing.

> The only real alternative is to have __attribute__
> behaviour dependent on its guts and that's not feasible - remember that
> there can be more than one attribute in the list insider the damn thing.
> Besides, it's bloody disgusting.

Agreed.  Not an option, even if we *could* implement it.

- Josh Triplett

Attachment: signature.asc
Description: OpenPGP digital signature