Re: __sync_lock_test_and_set on ARM

[David Daney <ddaney at avtrex dot com>]

Andrew Haley wrote:
> Phil Endecott writes:
>  > David Daney wrote:
>  > > Andrew Haley wrote:
>  > >> Phil Endecott writes:
>  > >>  > Andrew Haley wrote:
>  > >>  > > this stuff really should be done by the compiler. 
>  > >>  > 
>  > >>  > Yes.  I've filed a bug asking for a __sync_lock_test_and_set builtin:
>  > >>  > 
>  > >>  > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33413
>  > >> 
>  > >> Surer, but the problem is that for most of the things we want to do
>  > >> (lightweight locks, for example) __sync_lock_test_and_set() doesn't
>  > >> really do what we need: we need compare_and_swap().  That's why the
>  > >> kernel helper is so useful, because it's robust even if we are on
>  > >> pre-ARMv6 hardware.
>  > >
>  > > Probably the enhancement request should be expanded to include *all* the 
>  > > __sync_* atomic memory primitives  which includes compare_and_swap.
>  > 
>  > No, because none of the others can be implemented without kernel 
>  > support or other magic.
>  
> Yeah, I think that's right.  Even in the case of post-v6 -- where it
> actually is possible to do this safely in userland -- if it's
> necessary to execute a bunch of instructions you might as well call a
> subroutine.

There is nothing that says that __sync_compar_and_swap() cannot expand 
to a libcall.  If a libcall is the more efficient, then it *should* be 
generated.

The reason I suggest it is that I would like to see as much knowledge as 
possible about how to generate optimal code put into the compiler. 
Ideally a GCC user would not have to write target specific assembly for 
this.

David Daney