The ARM version of __div64_32() encapsulates a call to __do_div64 with
non-standard argument passing. In particular, __n is a 64-bit input
argument assigned to r0-r1 and __rem is an output argument sharing half
of that r0-r1 register pair.

With __n being an input argument, the compiler is in its right to
presume that r0-r1 would still hold the value of __n past the inline
assembly statement. Normally, the compiler would have assigned non
overlapping registers to __n and __rem if the value for __n is needed
again.

However, here we enforce our own register assignment and gcc fails to
notice the conflict. In practice this doesn't cause any problem as __n
is considered dead after the asm statement and *n is overwritten.
However this is not always guaranteed and clang rightfully complains.

Let's fix it properly by making __n into an input-output variable. This
makes it clear that those registers representing __n have been modified.
Then we can extract __rem as the high part of __n with plain C code.

This asm constraint "abuse" was likely relied upon back when gcc didn't
handle 64-bit values optimally. Turns out that gcc is now able to
optimize things and produces the same code with this patch applied.
Reported-by: Antony Yu <swpenim@gmail.com>
Signed-off-by: Nicolas Pitre <nico@fluxnic.net>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Tested-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>