Re: [RFA] [tree-optimization/80576] Handle non-constant sizes in DSE

[Richard Biener <richard dot guenther at gmail dot com>]

On Mon, Sep 16, 2019 at 11:12 AM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Mon, Sep 9, 2019 at 10:10 PM Jeff Law <law@redhat.com> wrote:
> >
> > On 8/26/19 3:00 AM, Richard Biener wrote:
> > > On Fri, Aug 23, 2019 at 9:19 PM Jeff Law <law@redhat.com> wrote:
> > >>
> > >> On 8/22/19 4:46 AM, Richard Biener wrote:
> > >>>>> Also you seem to use this info to constrain optimization when you
> > >>>>> might remember that types of addresses do not carry such information...
> > >>>>> Thus it should be "trivially" possible to write a testcase that is miscompiled
> > >>>>> after your patch.  I also don't see this really exercised in the
> > >>>>> testcases you add?
> > >>>> Arggh.  You're absolutely correct.  I must be blocking out that entire
> > >>>> discussion from last summer due to the trama :-)
> > >>>>
> > >>>> If the destination is the address of a _DECL node, can we use the size
> > >>>> of the _DECL?
> > >>>
> > >>> Yes, but this should already happen for both invariant ones like &a.b.c
> > >>> and variant ones like &a.b[i].c in ao_ref_init_from_ptr_and_size.
> > >> I don't see that in ao_ref_init_from_ptr_and_size.  AFAICT if you don't
> > >> know the size when you call that routine (size == NULL), then you end up
> > >> with the ref->size and ref->max_size set to -1.
> > >>
> > >> Am I missing something here?
> > >
> > > Ah, of course.  ao_ref_from_ptr_and_size would need to be extended
> > > to constrain max_size.  So what I was
> > > saying is that ao_ref_init_from_ptr_and_size should get you
> > > a DECL ao_ref_base () from which you could constrain max_size with.
> > > Or rather ao_ref_from_ptr_and_size should be extended do that,
> > > mimicing what get_ref_base_and_extent does at the end in the
> > > if (DECL_P (exp)) case (mind flag_unconstrained_commons!).
> > So I was going to use get_ref_base_and_extent from within
> > ao_ref_init_from_ptr_and_size to capture these cases, but
> > get_ref_base_and_extent internally uses TYPE_SIZE to get the maximum
> > size of the referenced object.
> >
> > That likely represents a codegen bug waiting to happen.
>
> Yeah, you can't use get_ref_base_and_extent literally here.
>
> > I'll see if I can refactor just the bits we want so that we're not
> > duplicating anything.
>
> Not sure if that's too important.  But yes, splitting out
>
>   if (DECL_P (exp))
>     {
>       if (VAR_P (exp)
>           && ((flag_unconstrained_commons && DECL_COMMON (exp))
>               || (DECL_EXTERNAL (exp) && seen_variable_array_ref)))
>         {
>           tree sz_tree = TYPE_SIZE (TREE_TYPE (exp));
>           /* If size is unknown, or we have read to the end, assume there
>              may be more to the structure than we are told.  */
>           if (TREE_CODE (TREE_TYPE (exp)) == ARRAY_TYPE
>               || (seen_variable_array_ref
>                   && (sz_tree == NULL_TREE
>                       || !poly_int_tree_p (sz_tree)
>                       || maybe_eq (bit_offset + maxsize,
>                                    wi::to_poly_offset (sz_tree)))))
>             maxsize = -1;
>         }
>       /* If maxsize is unknown adjust it according to the size of the
>          base decl.  */
>       else if (!known_size_p (maxsize)
>                && DECL_SIZE (exp)
>                && poly_int_tree_p (DECL_SIZE (exp)))
>         maxsize = wi::to_poly_offset (DECL_SIZE (exp)) - bit_offset;
>     }
>   else if (CONSTANT_CLASS_P (exp))
>     {
>       /* If maxsize is unknown adjust it according to the size of the
>          base type constant.  */
>       if (!known_size_p (maxsize)
>           && TYPE_SIZE (TREE_TYPE (exp))
>           && poly_int_tree_p (TYPE_SIZE (TREE_TYPE (exp))))
>         maxsize = (wi::to_poly_offset (TYPE_SIZE (TREE_TYPE (exp)))
>                    - bit_offset);
>     }
>
> into a helper with just the computed offset as argument
> (plus maybe that seen_variable_array_ref which is meaningless
> for the address case or rather has to be assumed true(?)) should be possible.

Btw, for ao_ref_init_from_ptr_and_size we really want something like
get_ref_base_and_extent, just not literal.  Given struct { int i; int
b[4]; int j; }
and &a.b[i] get_addr_base_and_unit_offset currently returns NULL
but for ao_ref_init_from_ptr_and_size we like to see &a as base
(we're getting that already) plus offset = 32, max_size = 160
(including the 'j' member of the decl).  Note if it's &a->b[i] max_size
needs to be -1 since there may be more after the structure, we can
really only prune max_size based on an actual object size.

But getting ->offset better is of course independent on max_size pruning.

Richard.


> Richard.
>
> >
> > Jeff