When this code is compiled by clang, it optimizes two byte writes into
one word write, like this:

>        fake_grow_down[0] = 'c';
>        *(fake_grow_down - 1) = 'b';
> 401b60:       66 41 c7 46 ff 62 63    movw   $0x6362,-0x1(%r14)

This is incorrect, as the stack is supposed to grow page by page,
so we need to touch one page then another, i.e. the order is important.

To fix, let's use volatile pointer. After this change, the following
(correct) code is generated:

>        *p-- = 'c';
>   401b60:       41 c6 06 63             movb   $0x63,(%r14)
>        *p = 'b';
>   401b64:       41 c6 46 ff 62          movb   $0x62,-0x1(%r14)

[v2: same fix for another similar place]

Cc: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: Kir Kolyshkin <kir@openvz.org>
Acked-by: Andrei Vagin <avagin@virtuozzo.com>
Signed-off-by: Pavel Emelyanov <xemul@virtuozzo.com>