Re: [PATCH] Add stdatomic.h for clang>=3.1 and gcc>=4.1

[Rich Felker <dalias@libc.org>]

On Sun, Nov 23, 2014 at 09:49:03AM +0100, Jens Gustedt wrote:
> Hello,
> 
> Am Samstag, den 22.11.2014, 20:43 -0500 schrieb Rich Felker:
> > On Sun, Nov 23, 2014 at 02:31:35AM +0100, Jens Gustedt wrote:
> > > Actually, I think a specially cooked synchronization tool would be
> > > better. Something that combines an atomic pointer (to point to the
> > > object) with a futex living on it for the waiting. This would probably
> > > be a bit more challenging to implement, but here we really have an
> > > interest to have the fast path really fast, just one CAS of the
> > > pointer.
> > 
> > I don't get what you mean. To access an atomic object larger than the
> > hardware supports, you have to hold a lock for the whole interval of
> > reading/writing.
> 
> No, why do you think that? If you implement access to a critical
> resource through a mutex, you only need one mutex and not several
> ones. The association to the whole range of the resource is only
> logical.

I think my statement was just unclear. I don't mean that the lock has
to be associated with the whole interval in memory. I mean the lock
has to be held for the whole interval in time. This makes it clear
that a spinlock is not appropriate even if you don't have the deadlock
issue to worry about from priority scheduling.

> Thinking of it, there might be an unintended loophole in the
> standard, due to a difference in _Atomic as a qualifier and as
> specifier. The qualifier version seems to permit to be applied to a
> struct that itself contains other _Atomic types. This then would not
> work with the table approach. I'll investigate.

Accessing individual members of atomic structs is UB last I checked.
IMO it's rather stupid that the standard allowed atomic structs
anyway, but we're stuck with them.

> > This is O(n) in the size of the object.
> 
> This would be prohibitive, indeed. Luckily we don't need that, so only
> the copy operation is O(n), and not the lock.

O(n) time the lock is held, not O(n) time obtaining locks. Sorry.

> > I don't see
> > where your ideas about pointers and CAS are coming in.

I'm still confused about this.

> > > > > What has all of this to do with VLA? I am lost.
> > > > 
> > > > The operands of __typeof__ and sizeof get evaluated when they have VLA
> > > > type. I think this is the problem.
> > > 
> > > ah, ok
> > > 
> > > No, this isn't a problem, I think. Arrays aren't allowed to be subject
> > > of an _Atomic qualification (arrays are never qualified
> > > themselves). For _Atomic type, the standard explicitly excludes
> > > arrays. So arrays in general and VLA in particular should never be
> > > passed as such into any of these generic functions, only pointers to
> > > atomic objects can.
> > 
> > Is a pointer to a variably modified type considered variably modified?
> 
> yes
> 
> > If so maybe these are affected too...
> 
> no, the pointers that can be passed to the atomic "functions" are
> always pointers to atomic objects, so they can't be arrays (and so
> VLA) themselves, nor can an atomic struct containt a VLA or a pointer
> to VLA.

_Atomic int (*pmat)[n];

Then &pmat is a pointer to a valid atomic type, an atomic pointer to a
VLA type.

Rich