Re: equating symbols to undefined

["H. J. Lu" <hjl at lucon dot org>]

On Sun, Oct 16, 2005 at 11:43:29AM +0200, Jan Beulich wrote:
> >>> "H. J. Lu" <hjl@lucon.org> 10/14/05 9:31 PM >>>
> >On Fri, Oct 14, 2005 at 12:27:42PM -0700, H. J. Lu wrote:
> >> On Fri, Oct 14, 2005 at 11:58:43AM -0700, H. J. Lu wrote:
> >> > On Fri, Oct 14, 2005 at 11:54:22AM -0700, H. J. Lu wrote:
> >> > > On Fri, Oct 14, 2005 at 08:25:29PM +0200, Jan Beulich wrote:
> >> > > > In what way does the sample sent not meet this? It matches
> exactly the
> >> > > > real situation that I having the problem with, only that I
> added the
> >> > > > __asm__(".global ..."); to show that making the symbol global
> doesn't
> >> > > > help. I specifically made that part contitional on a manifest
> constant,
> >> > > > so you can easily compare the two and see that making the
> symbol global
> >> > > > introduces a new problem, whereas not making the symbol global
> causes an
> >> > > 
> >> > > Can you tell me what the new problem is? I got
> >> > > 
> >> > > [hjl@gnu-13 alias]$ make
> >> > > cc  -c -o alias.o -save-temps alias.c
> >> > > cc  -c -o main.o -save-temps main.c
> >> > > cc  -c -o normal.o -save-temps normal.c
> >> > > cc  -c -o user.o -save-temps user.c
> >> > > cc -o tst alias.o main.o normal.o user.o
> >> > > ./tst
> >> > > copymem(0x500c00, 0x400700, 400)
> >> > > copymem(0x500c00, 0x400700, 400)
> >> > > 
> >> > > I expect "make" fails if there is a problem.
> >> > > 
> >> > 
> >> > Never mind. I figured it out.
> >> 
> >> Why does alias.c have to include alias.h? Specially, why does alias.c
> >> have to have
> >> 
> >> __asm__(".equiv memcpy, copymem");
> >> 
> >
> >It should be
> >
> >__asm__(".equiv memcpy, copymem");
> >__asm__(".global memcpy");
> >
> >I don't think they have to be in alias.c.
> 
> alias.c has to include alias.h according to good programming practice,
> and suppressing the __asm__-s (which indeed aren't necessary for
> alias.c, but also [should] do no harm) would only clutter the header
> file. Besides this or anything else, this problem is what I'd call a
> classical regression (because it obviously worked and there was no
> notion anywhere that this should not work).

I don't think the __asm__ statement belongs to alias.c. It shouldn't be
there to begin with. The old behavior silently makes a local symbol
global. I don't think it is a good idea. Can someone check out other
assemblers to see what they do?

> 
> And yes, as agreed to before, .alias would be a lot better suited for
> this purpose, but we don't have it (yet).

I can dig out my old patch and make it generic.


H.J.