Re: ARRAY_REF patch

[Mark Mitchell <mark at codesourcery dot com>]

Daniel Berlin wrote:
> On Mon, 2005-08-29 at 07:03 -0700, Mark Mitchell wrote:
>
> > Daniel Berlin wrote:
> >
> > > On Sun, 2005-08-28 at 23:14 -0700, Mark Mitchell wrote: 
> >
> > > > In other words, does your patch result in better code? 
> > >
> > > No, not at the moment.
> >
> > Given that, I think we should see if we can get this ready to be checked 
> > in immediately after 4.1 branches.
> >
> >
> > > > On the technical side, my most basic question is about the 
> > > > representation you chose.  Why did you choose to have the first operand 
> > > > be a POINTER_TYPE rather than an ARRAY_TYPE? 
> > >
> > >
> > > Because that is what the user wrote, and all my attempts to transform
> > > the POINTER_TYPE into an ARRAY_TYPE acceptable by the middle end failed
> > > miserably.
> > > :)
> > >
> > >
> > > > I think a better tree 
> > > > representation would be to use the tree equivalent of:
> > > >
> > > >  *((T (*)[]) p)
> > >
> > >
> > > Honestly, I have no idea how to do this.
> >
> > I would think something like this, assuming that the user has a pointer 
> > expression "p":
> >
> >   elt_type = TREE_TYPE (p);
> >   array_type = build_array_type (elt_type, /*index_type=*/NULL_TREE);
> >   pointer_type = build_pointer_type (array_type);
> >   p = build1 (NOP_EXPR, pointer_type, p);
> >   p = build1 (INDIRECT_REF, array_type, p);
>
>
>
> So wait, i'm lost :)
> If i wanted to transform p[5] into this representation, what does the
> code to do so look like. 
> (I'm not quite clear on how this isn't just pointer arithmetic, since
> you are putting it back in INDIRECT_REF form)

After the code I wrote, add:

  expr = build2 (ARRAY_REF, elt_type, p, build_int_cst (5));

I don't promise to have all of this *exactly* right, but close enough...

The key point is to treat "p" as a pointer not a single element, but 
rather as a pointer to an array of elements.

--
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304