On Nov 20, 2007 12:43 PM, Vegard Nossum <vegard.nossum@xxxxxxxxx> wrote:
> On Nov 20, 2007 12:16 PM, Vegard Nossum <vegard.nossum@xxxxxxxxx> wrote:
> > Here, offsetof(struct socket, type) = 0x24, like the one used in the
> > reads/writes. The type here is short, on 386 that's 16 bits. So why is
> > gcc later reading 32 bits off the same address, is that really legal?
> > Shouldn't that really have been a MOVZWL? Or did I miss something
> > obvious?
>
> I will add that compiling the file in question without optimisations
> (it was -Os), it does indeed produce a MOVZWL instruction instead. I
> am trying to construct a minimal test-case now. Help is still
> appreciated.
For the curious, here is the minimal example program:
struct a {
int dummy;
short x;
};
struct b {
unsigned short x;
};
extern void dummy(struct b *b);
void sock_init_data(struct a *a, struct b *b)
{
dummy(b);
if (a) {
asm("/* THIS FIELD IS ONLY 16 BITS */");
b->x = a->x;
}
}
As you see, with -Os, this produces:
/* THIS FIELD IS ONLY 16 BITS */
movl 4(%ebx), %eax
movw %ax, (%esi)
While the no-optimizations code looks like this:
movl 8(%ebp), %eax
movzwl 4(%eax), %eax
movl %eax, %edx
movl 12(%ebp), %eax
movw %dx, (%eax)
The first one is clearly loading a 32-bit value from an address that
may not have more than 16 bits written to it.
So, is it legal for the compiler to do that?
Vegard
--
To unsubscribe from this list: send an email with
"unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx
Please read the FAQ at http://kernelnewbies.org/FAQ
