On Wed, Jan 21, 2004 at 02:47:17PM +0100, Maciej W. Rozycki wrote:
> On Tue, 20 Jan 2004, Daniel Jacobowitz wrote:
>
> > No, I'm pretty sure Pavel's right.
> >
> > -#ifdef __MIPSEB__
> > case 1: these.sig[0] = these32.sig[0] | (((long)these32.sig[1]) << 32);
> > -#endif
> > -#ifdef __MIPSEL__
> > - case 1: these.sig[0] = these32.sig[1] | (((long)these32.sig[0]) << 32);
> > -#endif
> >
> > Consider a 64-bit sigset. 32-bit userland, 64-bit kernel. Here's a
> > userland sigset with signal 33 set, only, on a little endian target.
> > Word 1, least significant bit, right?
>
> Right, but...
>
> > byte address in memory
> > 1 2 3 4 5 6 7 8
> > val 0 0 0 0 0 0 0 1
>
> ... this is incorrect -- it would be right for big-endian; word #1, bit #1
> for little-endian is:
>
> byte address in memory
> 1 2 3 4 5 6 7 8
> val 0 0 0 0 1 0 0 0
>
>
> > Obviously, as a 64-bit integer the sigset looks different. There it's
> > supposed to be 1 << (33 - 1).
> > val 0 0 0 1 0 0 0 0
>
> Again, for little-endian it should actually be:
>
> val 0 0 0 0 1 0 0 0
>
> i.e. the whole operation is actually a no-op, except that the 64-bit
> vector is assured to be properly aligned for doubleword accesses.
Re-reading what I wrote, the above was actually supposed to be a
big-endian example. D'oh! If you pretend I wrote "big endian" up at
the top, then it makes sense.
> As a side note -- that's the reason certain C code portability problems
> related to the width of the machine word only get actually discovered when
> problematic software is run on a big-endian processor. I've been hit by
> this property once -- I was porting a 16-bit program and it appeared to
> run just fine on both a 32-bit (i386) and a 64-bit (Alpha) little-endian
> CPU, but when run on a 32-bit big-endian one (SPARC) I discovered a few
> more bits to be cleaned up.
>
> > So the correct algorithm to convert a userspace sigset to a kernel
> > sigset is to shift the second word left 32 bits, and leave the first
> > word right aligned, and or them together. Which is what using the
> > __MIPSEB__ case does.
>
> But this conclusion is of course right.
>
> Maciej
>
> --
> + Maciej W. Rozycki, Technical University of Gdansk, Poland +
> +--------------------------------------------------------------+
> + e-mail: macro@ds2.pg.gda.pl, PGP key available +
>
--
Daniel Jacobowitz
MontaVista Software Debian GNU/Linux Developer
|