tools/nolibc: i386: Fix a stack misalign bug on _start

The ABI mandates that the %esp register must be a multiple of 16 when
executing a 'call' instruction.

Commit 2ab44633 ("tools/nolibc: i386: shrink _start with _start_c")
simplified the _start function, but it didn't take care of the %esp
alignment, causing SIGSEGV on SSE and AVX programs that use aligned move
instruction (e.g., movdqa, movaps, and vmovdqa).

The 'and $-16, %esp' aligns the %esp at a multiple of 16. Then 'push
%eax' will subtract the %esp by 4; thus, it breaks the 16-byte
alignment. Make sure the %esp is correctly aligned after the push by
subtracting 12 before the push.

Extra:
Add 'add $12, %esp' before the 'and $-16, %esp' to avoid over-estimating
for particular cases as suggested by Willy.

A test program to validate the %esp alignment on _start can be found at:

   https://lore.kernel.org/lkml/ZOoindMFj1UKqo+s@biznet-home.integral.gnuweeb.org



[ Thomas: trim Fixes tag commit id ]

Cc: Zhangjin Wu <falcon@tinylab.org>
Fixes: 2ab44633 ("tools/nolibc: i386: shrink _start with _start_c")
Reported-by: Nicholas Rosenberg <inori@vnlx.org>
Acked-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Ammar Faizi <ammarfaizi2@gnuweeb.org>
Reviewed-by: Alviro Iskandar Setiawan <alviro.iskandar@gnuweeb.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>