https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71716
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |jakub at gcc dot gnu.org
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 38811
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38811&action=edit
gcc7-pr71716.patch
Untested fix.
Beyond this, it would be nice if at least the C FE _Atomic cas loop lowering is
improved (for the lock-free case only?), so that the old var does not have
floating point type, but instead corresponding integral type. The thing is,
otherwise if one is unlucky enough, the VIEW_CONVERT_EXPR in the IL from
__int128 to long double can be expanded through XFmode store to uninitialized
memory slot, which leaves lots of bits uninitialized. If those uninitialized
bits don't match the one in the *addr, then the CAS will fail the first time.
