[Bug tree-optimization/54004] New: __atomic_always_lock_free has a fault implementation in which the target has no saying

[hp at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54004

             Bug #: 54004
           Summary: __atomic_always_lock_free has a fault implementation
                    in which the target has no saying
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassig...@gcc.gnu.org
        ReportedBy: h...@gcc.gnu.org
        Depends on: 54003


Created attachment 27819
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27819
Conceptual patch - note comments

Basically, fold_builtin_atomic_always_lock_free compares alignment of the type
of the atomic data to that of the normal alignment for a mode.  A better
default is to compare to the *natural* alignment.  For most targets, the two
are the same. If that's approach is not ok, then a target hook is needed. 
Conceptual patch (i.e. untested, uncompiled) for the first approach attached. 
Note that this suggestion is counter to the documented dogma of "presence of
library or atomic pattern implies lock-free-ness", but that already has
inconsistencies, see linked PR.

Note that this likely means we need a libgcc implementation for
__atomic_is_lock_free or at least document that the target needs to supply one.