gcc tickets for sparse attributes

gcc tickets for sparse attributes

[H. Peter Anvin]

Hi,

I have filed gcc tickets asking for direct support in gcc for some
sparse extensions that we use heavily in the kernel:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59850
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59851

However, I would also like support for the context extensions, but I'm
not knowledgeable enough to describe the semantics accurately.  Would
anyone be willing to file a ticket describing how the context extension
works well enough that it could be implemented?

	-hpa

Re: gcc tickets for sparse attributes

[H. Peter Anvin]

On 01/16/2014 09:27 PM, H. Peter Anvin wrote:
> Hi,
> 
> I have filed gcc tickets asking for direct support in gcc for some
> sparse extensions that we use heavily in the kernel:
> 
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59850
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59851
> 

Add:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59852

	-hpa

Re: gcc tickets for sparse attributes

[Johannes Berg]

On Thu, 2014-01-16 at 21:27 -0800, H. Peter Anvin wrote:

> However, I would also like support for the context extensions, but I'm
> not knowledgeable enough to describe the semantics accurately.  Would
> anyone be willing to file a ticket describing how the context extension
> works well enough that it could be implemented?

IMHO the context extension doesn't work well enough in sparse to
document and implement as is. It would be much better if it actually was
able to differentiate between contexts, rather than treating each one
the same. This would avoid the problem that locking one lock and
unlocking another (in the kernel's __acquire/ __release mechanism) could
still result in a warning.

johannes

Re: gcc tickets for sparse attributes

[Josh Triplett]

On Thu, Jan 16, 2014 at 09:27:08PM -0800, H. Peter Anvin wrote:
> I have filed gcc tickets asking for direct support in gcc for some
> sparse extensions that we use heavily in the kernel:
> 
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59850
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59851
> 
> However, I would also like support for the context extensions, but I'm
> not knowledgeable enough to describe the semantics accurately.  Would
> anyone be willing to file a ticket describing how the context extension
> works well enough that it could be implemented?

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59856

Also:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59855

- Josh Triplett

Re: gcc tickets for sparse attributes

[Josh Triplett]

On Fri, Jan 17, 2014 at 09:33:16AM +0100, Johannes Berg wrote:
> On Thu, 2014-01-16 at 21:27 -0800, H. Peter Anvin wrote:
> 
> > However, I would also like support for the context extensions, but I'm
> > not knowledgeable enough to describe the semantics accurately.  Would
> > anyone be willing to file a ticket describing how the context extension
> > works well enough that it could be implemented?
> 
> IMHO the context extension doesn't work well enough in sparse to
> document and implement as is. It would be much better if it actually was
> able to differentiate between contexts, rather than treating each one
> the same.

That would certainly be nice, but that's something actually much more
easily done in GCC than in Sparse, given the types of information GCC
already has available to implement features like alias analysis.

In any case, the spec I wrote up assumes a distinction between contexts,
but allows for an initial implementation like Sparse's that ignores the
distinction.

> This would avoid the problem that locking one lock and
> unlocking another (in the kernel's __acquire/ __release mechanism) could
> still result in a warning.

That would actually *not* produce a warning, though it should.  In
general, I *think* an implementation like Sparse's that ignores the
distinction between locks should produce false negatives but not false
positives.

- Josh Triplett

Re: gcc tickets for sparse attributes

[Johannes Berg]

> > IMHO the context extension doesn't work well enough in sparse to
> > document and implement as is. It would be much better if it actually was
> > able to differentiate between contexts, rather than treating each one
> > the same.
> 
> That would certainly be nice, but that's something actually much more
> easily done in GCC than in Sparse, given the types of information GCC
> already has available to implement features like alias analysis.

Right.

> In any case, the spec I wrote up assumes a distinction between contexts,
> but allows for an initial implementation like Sparse's that ignores the
> distinction.

Ok cool. :)

> > This would avoid the problem that locking one lock and
> > unlocking another (in the kernel's __acquire/ __release mechanism) could
> > still result in a warning.
> 
> That would actually *not* produce a warning, though it should.  In
> general, I *think* an implementation like Sparse's that ignores the
> distinction between locks should produce false negatives but not false
> positives.

Right, it doesn't report one now.

johannes