From: Linus Torvalds
> Sent: 18 May 2021 15:56
>
> On Tue, May 18, 2021 at 12:27 AM Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
> > >
> > > I wonder if the kernel should do the same, or whether there are still cases
> > > where memcpy() isn't compiled optimally. armv6/7 used to be one such case, but
> > > it was fixed in gcc 6.
> >
> > It would have to be memmove(), not memcpy() in this case, right?
>
> No, it would simply be something like
>
> #define __get_unaligned_t(type, ptr) \
> ({ type __val; memcpy(&__val, ptr, sizeof(type)); __val; })
You still need something to ensure that gcc can't assume that
'ptr' has an aligned type.
If there is an 'int *ptr' visible in the call chain no amount
of (void *) casts will make gcc forget the alignment.
So the memcpy() will get converted to an aligned load-store pair.
(This has always caused grief on sparc.)
A cast though (long) might be enough, as might a cast to a __packed
struct pointer type.
Using a union of the two pointer types might be ok - but might
generate a store/load to stack.
An alternative is an asm statement with input and output of different
pointer types but using the same register for both.
That ought to force the compile to forget any tracked type
and value.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
