[Bug target/71991] Inconsistency for __attribute__ ((__always_inline__)) among LTO and non-LTO compilation

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71991

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

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

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Jan Hubicka from comment #5)
> Created attachment 43798 [details]
> proposed fix
> 
> 
> this patch simply while-lists some transitions of target flags for always
> inline functions. It is ugly but I can't think of anything else which would
> look safe.
> Martin, you mentioned there was packages broken by this. Perhaps we can try
> rebuild with this patch to see if all of the real world issues are solved?
> If not, i guess we will need to decide case-by-case what is safe and what
> not :(

I think that patch is quite reasonable, with a couple of small nits LGTM
s/alwyas/always/
put = on the next line in bool always_inline =
and perhaps add int safe_mask = always_inline ? always_inline_safe_mask : 0;
and use unconditionally
(caller_opts->x_target_flags & ~safe_mask)
!= (callee_opts->x_target_flags & ~safe_mask)

maybe use unsigned HOST_WIDE_INT instead of int for the masks, so when the
masks grow beyond 32 bits we don't suddenly misbehave (or add gcc_chec
king_assert that sizeof (callee_opts->x_target_flags) == sizeof
(always_inline_safe_mask)

The xen case should just be fixed in xen IMHO.