[RFC2 RESEND PATCH] [AARCH64] ILP32: support stat syscall family

[Joseph Myers]

On Wed, 6 Jul 2016, Yury Norov wrote:

>  - I didn't move DECLARE_TIMESPEC() to non-installed file because it's 
> used in stat.h,
>    and should be known on used side. If it's wrong please hint me how to 
>    do this trick.

First: why is it used in an installed header at all?  The public API uses 
struct timespec, not struct __timespec.  If you need to reinterpret the 
bits of struct timespec in some other way, because the kernel ABI is 
different from the userspace ABI, that is purely a matter for the relevant 
function implementations and should not affect the installed header.

Second: any interface in an installed header that does not start with 
_[_A-Z] is a public interface.  This means that (a) it must be permitted 
to appear in that header by the relevant standards, depending on the 
feature test macros in use, (b) we must have made a deliberate decision as 
a community that this is an interface that we wish to support for our 
users, indefinitely, (c) it must normally be documented and have testsuite 
coverage, and (d) it must be system-independent unless there is a definite 
reason for a system-specific interface.  So any macros, typedefs etc. in 
installed headers that do not meet all those requirements must be in the 
implementation namespace _[_A-Z]*.  This applies to all such macros added 
to installed headers by this patch.  That includes statfs_word_t; although 
POSIX reserves *_t, we should not have any such types in public headers, 
outside the _[_A-Z]* namespace, unless they are documented public 
interfaces meeting all of (a)-(d), so you should not use statfs_word_t as 
a type name.

> +/* Declare structure field that has different size
> +   in 32- and 64-bit ABIs with paddings where needed,
> +   so final layout becomes identical.  */
> +#if __WORDSIZE == 32
> +# if __BYTE_ORDER == __LITTLE_ENDIAN
> +#  define __type3264(type, name) type name; type name##_pad
> +# else
> +#  define __type3264(type, name) type name##_pad; type name
> +# endif

I don't think you want name##_pad; if name is a public interface, then 
this would make name##_pad a public interface as well, with all the 
corresponding requirements.  __##name##_pad would be better.  And why 
isn't this using alignment attributes as in 
sysdeps/unix/sysv/linux/generic/bits/stat.h?  I wonder if this needs to be 
separate from __field64 there or if there could be some common macro?

-- 
Joseph S. Myers
joseph@codesourcery.com