Re: emacs-26 9e59de9: Use GCALIGNED properly for GCC

[Paul Eggert]

Eli Zaretskii wrote:
> Paul, maybe we should simply go back a few notches and not mark
> main_thread as GCALIGNED, except on that single platform which needed
> it (i.e. via #ifdef)?  Then we could return all the other symbols to
> their previous syntax.
>
> Or maybe we should move GCALIGNED to yet another position, like at the
> beginning or the end of the whole construct.  Like this:
>
>     GCALIGNED struct foo FOO;
> or
>     struct foo FOO GCALIGNED;
>
> Not sure what else can we do, but this seems to be an exceptionally
> fragile feature, at least with GCC 7 that many people are using.

I've investigated this further, and neither solution is likely to work. It is 
merely the luck of the draw that we have seen the bug only on that platform so 
far, and a similar bug is likely to occur on other platforms. Neither GCALIGNED 
syntax works in general (see GCC bug 82914) and we've seen examples of failures 
with either syntax.

I have thought of a solution that fixes the problem by dropping use of GCALIGNED 
and instead using classic C unions along with 'char alignas (8)'. The idea is to 
gcalign a 'struct foo' this way:

   union gcaligned_foo { struct foo s; char alignas (8) gcaligned; };

Something like this should work on all platforms that Emacs ports to. I plan to 
work on a first cut tomorrow.