[PATCH] Update to new generic semaphore algorithm v2

[Maciej W. Rozycki]

On Thu, 22 Jan 2015, Torvald Riegel wrote:

> > In file included from ../sysdeps/nptl/lowlevellock.h:22:0,
> >                  from ../nptl/descr.h:30,
> >                  from ../sysdeps/mips/nptl/tls.h:73,
> >                  from ../include/errno.h:27,
> >                  from sem_waitcommon.c:20,
> >                  from sem_wait.c:20:
> > sem_waitcommon.c: In function ???~@~X__sem_wait_cleanup???~@~Y:
> > sem_waitcommon.c:190:47: error: left shift count >= width of type
> > [-Werror=shift-count-overflow]
> >    atomic_fetch_add_relaxed (&sem->data, -(1UL << SEM_NWAITERS_SHIFT));
> >                                                ^
> > ../include/atomic.h:617:31: note: in definition of macro
> > ???~@~Xatomic_fetch_add_relaxed???~@~Y
> >    __atomic_fetch_add ((mem), (operand), __ATOMIC_RELAXED); })
> >                                ^
> > sem_waitcommon.c: In function ???~@~X__new_sem_wait_slow???~@~Y:
> > sem_waitcommon.c:267:11: error: left shift count >= width of type
> > [-Werror=shift-count-overflow]
> >        1UL << SEM_NWAITERS_SHIFT);
> 
> Do you use an LP64 data model?  "1UL" is 64b if LP64, and
> SEM_NWAITERS_SHIFT is 32.  You could try 1ULL instead to see whether
> that makes a difference.
> The semaphore code uses the 64b version, because atomic.h thinks 64b
> atomic ops are available.
> 
> Which compiler do you use?

 N32 MIPS is an ILP32 ABI with 64-bit registers.  So the `long long' type 
and its derivatives use native 64-bit operations (that are atomic if 
required) with no performance penalty, but the traditional (C89) C data 
types are limited to 32 bits only.  I gather this is GCC; I'd expect that 
behaviour with GCC anyway.

  Maciej