On Mon, Mar 19, 2018 at 11:23:42PM +0000, Al Viro wrote:

Benefits:
	* all SyS... wrappers (i.e. the thing that really ought to
go into syscall tables) have the same type.
	* we could have SYSCALL_DEFINE produce a trivial compat
wrapper, have explicit COMPAT_SYSCALL_DEFINE discard that thing
and populate the compat syscall table *entirely* with compat_SyS_...,
letting the linker sort it out.  That way we don't need to keep
track of what can use native and what needs compat in each compat
table on biarch.
	* s390 compat wrappers would disappear with that approach.
	* we could even stop generating sys_... aliases - if
syscall table is generated by slapping SyS_... or compat_SyS_...
on the name given there, we don't need to _have_ those sys_...
things at all.  All SyS_... would have the same type, so the pile
in syscalls.h would not be needed - we could generate the externs
at the same time we generate the syscall table.

And yes, it's a high-squick approach.  I know and I'm not saying
it's a good idea.  OTOH, to quote the motto of philosophers and
shell game operators, "there's something in it"...

FWIW, I have something that is almost reasonable on preprocessor side;
however, that has uncovered the following fun:
void f(unsigned long long);
void g(unsigned a, unsigned b)
{
        f((((unsigned long long)b)<<32)|a);
}

which does compile to "jump to f" on i386, ends up with the following
joy on arm:
        mov     r3, r1
        mov     r2, #0
        push    {r4, lr}
        orr     r2, r2, r0
        mov     r0, r2
        mov     r1, r3
        bl      f
        pop     {r4, lr}
        bx      lr
with gcc6; gcc7 is saner - there we have just
        mov     r2, #0
        orr     r0, r2, r0
        b       f

The former is
	r3 = r1
	r2 = 0
	r2 |= r0
	r0 = r2
	r1 = r3
The latter -
	r2 = 0
	r0 |= r2
which is better, but still bloody odd

And I'm afraid to check what e.g. 4.4 will do with that testcase...