[PATCH] Look at restrict disambiguation in tree-ssa-alias.c unconditionally (PR tree-optimization/50522)

[Jakub Jelinek]

On Tue, Oct 04, 2011 at 11:55:17AM +0200, Richard Guenther wrote:
> > On Tue, Oct 04, 2011 at 11:01:27AM +0200, Richard Guenther wrote:
> > > > > > void foo (int *p)
> > > > > > {    
> > > > > >   int * __restrict p1 = p;
> > > > > >   int * __restrict p2 = p + 32;
> > > > > >   int *q;
> > > > > >   int i;
> > > > > >   for (i = 0; i < 32; ++i)
> > > > > >     p1[i] = p2[i];
> > > > > >   p = p1;
> > > > > >   q = p2 - 31;
> > > > > >   for (i = 0; i < 32; ++i)
> > > > > >     p[i] = q[i];
> > > > > > }
> > > > > > 
> > 
> > > In the above first loop the restrict pointers p1 and p2 access
> > > distinct object pieces.  The second loop uses non-restrict qualified
> > > pointers p and q (that are based on the restrict variants p1 and p2
> > > though) to access overlapping pieces.  Is the second loop invalid
> > > because p and q are based on p1 and p2 even though they are not
> > > restrict qualified?
> > 
> > IMHO yes.  The standard doesn't seem to talk about the accesses being done
> > through the restricted pointer, but about accesses that are based on
> > the restricted pointer, and as soon as you access in the associated block
> > (here the foo function) some object through an lvalue whose address is
> > based on some restricted pointer and the value is modified by any means,
> > then all accesses to that object need to be done through something
> > based on that restricted pointer.
> 
> So when I change the above to
> 
>  /*p = p;*/
>  q = (p + 32) - 31;

void foo (int *p)
{    
  int * __restrict p1 = p;
  int * __restrict p2 = p + 32;
  int *q;
  int i;
  for (i = 0; i < 32; ++i)
    p1[i] = p2[i];
  q = (p + 32) - 31;
  for (i = 0; i < 32; ++i)
    p[i] = q[i];
}

> then the code will be valid?  When I obfuscate that enough I
> can get GCC CSEing p + 32 and thus effectively q will look
> like it is based on p2.

The above is still invalid. p[0] through p[31] is modified and
is accessed both through lvalue whose address is based on p1 (p1[i])
and through lvalues whose address is not based on p1 (p[i] and
q[i] (the latter only for p[0] through p[30])).  If you take
the first loop out, it would be valid though.

	Jakub