Re: Linux getdents.c is not aliasing safe

[Daniel Jacobowitz <dan at debian dot org>]

On Wed, Oct 30, 2002 at 12:03:12PM +0100, Jakub Jelinek wrote:
> On Wed, Oct 30, 2002 at 10:52:30AM +0100, Wolfram Gloger wrote:
> > > > [That is: two pointers to a buffer, of different types; walk through
> > > > the buffer, reading one object of the first type and writing out an
> > > > object of the second lesser-or-equal sized type.]
> > > 
> > > No, there's no standard C way to do that.  An object can only have one
> > > type at a time.  You have to use another buffer as an intermediate.
> > 
> > AFAICS, the code already has intermediate buffers:
> > 
> >           dp = (DIRENT_TYPE *)buf;
> >           kdp = (struct kernel_dirent64 *) kbuf;
> > 
> >               uint64_t d_ino = kdp->d_ino;
> >               int64_t d_off = kdp->d_off;
> >               unsigned char d_type = kdp->d_type;
> > 
> >               DIRENT_SET_DP_INO (dp, d_ino);
> >               dp->d_off = d_off;
> > 
> > So, why not just use a char* ptr to write out the data using memcpy(),
> > do away with dp and e.g. change the last line into:
> > 
> > 	memcpy(buf + offsetof(DIRENT_TYPE, d_off), &d_off, sizeof d_off);
> > 
> > (the DIRENT_SET_DP_INO line of course needs to be handled a little
> > differently).  Since char* can alias anything, that would remain
> > portable C.
> 
> But that intermediate buffer is in registers (on sane arches), while using
> memcpy would mean at least with current compilers memory.
> glibc requires gcc (even 3.2+) to compile anyway, so IMHO just using the
> union as drow posted seems best to me.

If my patch is OK, could someone commit it, please?

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer