On Sat, Aug 14, 2004 at 12:51:06PM +0000, Christoph_Egger@gmx.de wrote:
> Hi!
> 
> 
> All applications that do the following crashes on the sparc64
> port (I don't know, if the crash happens on other ports, too):
> 
> 
> #include <stdio.h>
> 
> 
> struct foo_t {
>         int a;
>         unsigned char buf[8192];
> };
> 
> struct bar_t {
>         int i1;
>         int i2;
> 
>         double d1;
>         double d2;
> };
> 
> 
> int main(void)
> {
>         int a = 0;
>         struct foo_t f;
>         struct bar_t *b;
> 
>         memset(&f, 0, sizeof(f));
>         f.buf[2] = 2;
> 
>         b = (struct bar_t *)(f.buf + 2);
>         a = b->i2;  // <-- accessing b causes segfault
> 
>         printf("a: %i\n", a);
>         return 0;
> }
> 
> 
> I have tested it on other OS's (Darwin/ppc32, Solaris/sparc64)
> and this small app does NOT crash there.

This is not a bug in the VM, it's a bug in your program. When you do
         b = (struct bar_t *)(f.buf + 2);
b is not 32-bits-aligned, and this will cause a trap on sparc64.
testing on ppc32 is irrelevant, because alignement issues are
architecture-dependants.
To understand how it works on Solaris/sparc64 you should look closely at
what is happening, because on solaris the kernel may emulate this 
instruction by itself (as NetSD does on the alpha).
It may also be the compiler optimising this differently, or detecting that
the access won't be aligigned and correcting this by itself.

In any cases, doing such nonaligned access is non-portable and should not
be used.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--