Re: PR 58958: wrong aliasing info

[Marc Glisse <marc dot glisse at inria dot fr>]

On Mon, 4 Nov 2013, Richard Biener wrote:

> On Fri, Nov 1, 2013 at 11:39 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
> > Hello,
> >
> > the issue was described in the PR and the message linked from there.
> > ao_ref_init_from_ptr_and_size calls get_ref_base_and_extent, which may
> > detect an array_ref of variable index, but ao_ref_init_from_ptr_and_size
> > never learns of it and uses the offset+size as if they were meaningful.
>
> Well...  it certainly learns of it, but it chooses to ignore...
>
>   if (TREE_CODE (ptr) == ADDR_EXPR)
> -    ref->base = get_ref_base_and_extent (TREE_OPERAND (ptr, 0),
> -                                        &ref->offset, &t1, &t2);
> +    {
> +      ref->base = get_addr_base_and_unit_offset_1 (TREE_OPERAND (ptr, 0),
> +                                                  &t, 0, &safe);
> +      ref->offset = BITS_PER_UNIT * t;
> +    }
>
> safe == (t1 != -1 && t1 == t2)

I'll try that... (I need to think whether that's sufficient to be safe)

> note that ao_ref_init_from_ptr_and_size gets the size fed in as argument
> so I fail to see why it matters at all ...?  That is, if you feed in a wrong
> size then it's the callers error.

The caller is feeding the right size. The issue is that 
get_ref_base_and_extent cannot determine the offset as a constant. 
Normally, get_ref_base_and_extent then gives you a safe combination of 
offset+maxsize to cover the whole decl. Here, we don't want to use the 
size determined by get_ref_base_and_extent, we know better, but that also 
means we have to handle the case where the offset couldn't be determined 
as a constant.

--
Marc Glisse