[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot
                   |                            |gnu.org, jakub at gcc dot
                   |                            |gnu.org, rsandifo at gcc
                   |                            |dot gnu.org

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-07-04 08:10:19 UTC ---
I think the problem is either that the MEM has the MEM_NOTRAP_P flag set on it:
(mem/f/i:SI (symbol_ref/i:SI ("__dso_handle") <var_decl 0x7ffff13ce320
__dso_handle>) [0 S4 A32])
That makes ifcvt believe it can try to use a conditional move with that as an
operand, and as there is no conditional move with MEM operand on this target,
it is forced into register first.

MEM_NOTRAP_P is set in set_mem_attributes_minus_bitpos:
  /* Note whether this expression can trap.  */
  MEM_NOTRAP_P (ref) = !tree_could_trap_p (t);
where t is VAR_DECL for __dso_handle with DECL_WEAK.

Or the problem might be that noce_try_cmove_arith should be using
  else if (may_trap_or_fault_p (a) || may_trap_or_fault_p (b))
    return FALSE;
instead of
  else if (may_trap_p (a) || may_trap_p (b))
    return FALSE;

Eric/Richard, what do you think?  The comments seem to be fuzzy to me, some
comments say that MEM_NOTRAP_P only apply to the position where the MEM is
originally used (and in this case the __dso_handle var is read only guarded
with if (__dso_handle != NULL), on the other side I wonder if we ever clear
MEM_NOTRAP_P on unguarded DECL_WEAK var references.