Re: trouble spots for atomic access

[Rich Felker <dalias@libc.org>]

On Wed, May 20, 2015 at 01:05:19PM +0200, Szabolcs Nagy wrote:
> * Jens Gustedt <jens.gustedt@inria.fr> [2015-05-20 00:47:44 +0200]:
> > Am Dienstag, den 19.05.2015, 18:07 -0400 schrieb Rich Felker:
> > > On Tue, May 19, 2015 at 03:57:00PM +0200, Jens Gustedt wrote:
> > > >  - pthread_once_t should always be volatile
> > > >  - pthread_spinlock_t should always be volatile
> > > 
> > > These are C++ ABI changes. I'd like to make the change but I'm
> > > concerned it might break things.
> > 
> > Both are broken as they are now, if you fall into a compiler that
> > "knows" that the object itself isn't volatile qualified, and by that
> > excuse takes the liberty to optimize out loads. For both types there
> > is one point where there is a cast to (volatile int*) followed by a
> > load, that might not do what we want.
> > 
> > (For pthread_once_t, this on line 43 in pthread_once.c)
> > 
> > I think the safest would be to make the data types volatile. If that
> > is not possible, do the load with some new function "a_load_rel" that
> > is guaranteed to be volatile, atomic and with memory_order relaxed.
> 
> fwiw i scanned a significant portion of debian packages with nm -CD
> for volatile int* usage and didnt find any case that was related to
> posix types (a few libraries dealing with atomics had their own
> volatile int*)
> 
> (found 6553 packages with c++ dependencies in stable, scanned them all)
> 
> so changing the types wouldnt break too many musl binaries i think

This sounds promising, but I'm still a bit worried about making the
change. I wonder if there's other data that could support the safety
of doing it.

Rich