Re: Compatibility issue with 2.2.19pre7

[Jesse Pollard]

Andrea Arcangeli <[e-mail]>:
> On Tue, Jan 23, 2001 at 11:51:15PM -0800, Richard Henderson wrote:
> > On Thu, Jan 11, 2001 at 12:59:24AM +0100, Andrea Arcangeli wrote:
> > > What I said is that I can write this C code:
> > > 
> > > 	int x[2], * p = (int *) (((char *) &x)+1);
> > > 	main()
> > > 	{
> > > 		*p = 0;
> > > 	}
> > > 
> > > This is legal C code.
> > 
> > Err, no.  This is not "legal" by any stretch of the imagination.
> > This code has undefined behaviour.
> 
> I know this code has undefined behaviour at _runtime_. But I thought
> you were obliged to allow it to compile. That was my only point.
> 
> > We aren't even obliged to allow this to compile.
> 
> I'd love if you could forbid it to compile.

It would have to be implementation dependant/architecture dependant-
The only way to know is to realize that *p initializer value is invalid
is to know:

1. integer addresses MUST be aligned (say 4 or 8 byte boundaries)
2. the computation of (int *) (((char *) &x)+1) is a constant at
   compile time, and not link time (some linkers can do address offset
   calculations if "x" happens to be an external - ie. x +/- offset and
   offset is a constant)

Other systems would have to allow it to compile (and link). It still has
undefined behaviour, and is non-portable.

-------------------------------------------------------------------------
Jesse I Pollard, II
Email: [e-mail]

Any opinions expressed are solely my own.