calls with stack_size > 16 pop argument values into x3, overwriting
the number of 16-byte stack octwords. This breaks the checks for popping
x4..x7 and potentially unbalances sp (based on the vaue of args[16..23]).

Use a scratch register (x9) for this count so its lifetime does not
conflict with preparing the parameter/result registers.
Signed-off-by: Kevin Puetz <PuetzKevinA@JohnDeere.com>
Signed-off-by: Alexandre Julliard <julliard@winehq.org>