Re: [PATCH 5/9] Come up with an abstraction.

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

On Wed, Aug 14, 2019 at 3:19 PM Martin Liška <mliska@suse.cz> wrote:
>
> On 8/14/19 3:04 PM, Richard Biener wrote:
> > On Mon, Aug 12, 2019 at 3:56 PM Martin Liška <mliska@suse.cz> wrote:
> >>
> >> On 8/12/19 2:43 PM, Richard Biener wrote:
> >>> On Mon, Aug 12, 2019 at 1:49 PM Martin Liška <mliska@suse.cz> wrote:
> >>>>
> >>>> On 8/12/19 1:40 PM, Richard Biener wrote:
> >>>>> On Mon, Aug 12, 2019 at 1:19 PM Martin Liška <mliska@suse.cz> wrote:
> >>>>>>
> >>>>>> On 8/8/19 5:55 PM, Michael Matz wrote:
> >>>>>>> Hi,
> >>>>>>>
> >>>>>>> On Mon, 10 Jun 2019, Martin Liska wrote:
> >>>>>>>
> >>>>>>>> 2019-07-24  Martin Liska  <mliska@suse.cz>
> >>>>>>>>
> >>>>>>>>      * fold-const.c (operand_equal_p): Rename to ...
> >>>>>>>>      (operand_compare::operand_equal_p): ... this.
> >>>>>>>>      (add_expr):  Rename to ...
> >>>>>>>>      (operand_compare::hash_operand): ... this.
> >>>>>>>>      (operand_compare::operand_equal_valueize): Likewise.
> >>>>>>>>      (operand_compare::hash_operand_valueize): Likewise.
> >>>>>>>>      * fold-const.h (operand_equal_p): Set default
> >>>>>>>>      value for last argument.
> >>>>>>>>      (class operand_compare): New.
> >>>>>>>
> >>>>>>> Hmpf.  A class without any data?  That doesn't sound like a good design.
> >>>>>>
> >>>>>> Yes, the base class (current operand_equal_p) does not have a data.
> >>>>>> But the ICF derive class has a data and e.g. func_checker::operand_equal_valueize
> >>>>>> will use m_label_bb_map.get (t1). Which are member data of class func_checker.
> >>>>>>
> >>>>>>> You seem to need it only to have the possibility of virtual functions,
> >>>>>>> i.e. fancy callbacks.  AFAICS you only have one derived class, i.e. a
> >>>>>>> simple distinction of two cases.  What do you think about encoding the
> >>>>>>> additional new (ICF) case in the (existing) 'flags' argument to
> >>>>>>> operand_equal_p (and in case the ICF flag is set simply call the
> >>>>>>> "callback" directly)?
> >>>>>>
> >>>>>> That's possible. I can add two more callbacks to the operand_equal_p function
> >>>>>> (hash_operand_valueize and operand_equal_valueize).
> >>>>>>
> >>>>>> Is Richi also supporting this approach?
> >>>>>
> >>>>> I still see no value in the abstraction since you invoke none of the
> >>>>> (virtual) methods from the base class operand_equal_p.
> >>>>
> >>>> I call operand_equal_valueize (and hash_operand) from operand_equal_p.
> >>>> These are then used in IPA ICF (patch 6/9).
> >>>
> >>> Ugh.  I see you call that after
> >>>
> >>>   if (TREE_CODE (arg0) != TREE_CODE (arg1))
> >>>     {
> >>> ...
> >>>         }
> >>>       else
> >>>         return false;
> >>>     }
> >>>
> >>> and also after
> >>>
> >>>   /* Check equality of integer constants before bailing out due to
> >>>      precision differences.  */
> >>>   if (TREE_CODE (arg0) == INTEGER_CST && TREE_CODE (arg1) == INTEGER_CST)
> >>>
> >>> which means for arg0 == SSA_NAME and arg1 == INTEGER_CST you return false
> >>> instead of valueizing arg0 to the possibly same or same "lose" value
> >>> and returning true.
> >>
> >> Yes. ICF does not allow to have anything where TREE_CODEs do not match.
> >>
> >>>
> >>> Also
> >>>
> >>> +  int val = operand_equal_valueize (arg0, arg1, flags);
> >>> +  if (val == 1)
> >>> +    return 1;
> >>> +  if (val == 0)
> >>> +    return 0;
> >>>
> >>> suggests that you pass in arbirtrary trees for "valueization" but it
> >>> isn't actually
> >>> valueization that is performed but instead it should do an alternate comparison
> >>> of arg0 and arg1 with valueization.  Why's this done this way instead of
> >>> sth like
> >>>
> >>>   if (TREE_CODE (arg0) == SSA_NAME)
> >>>    arg0 = operand_equal_valueize (arg0, flags);
> >>>  if (TREE_CODE (arg1) == SSA_NAME)
> >>>    arg1 = operand_equal_valueize (arg1, flags);
> >>
> >> Because I want to be given a pair of trees about which the function
> >> operand_equal_valueize returns match/no-match/dunno.
> >>
> >>>
> >>> and why's this done with virtual functions rather than a callback that we can
> >>> cheaply check for NULLness in the default implementation?
> >>
> >> I can transform it into a hook. But as mentioned I'll need two hooks.
> >>
> >>>
> >>> So - what does ICF want to make "equal" that isn't equal normally and how's
> >>> that "valueization"?
> >>
> >> E.g. for a FUNCTION_DECL, ICF always return true because it can only calls
> >> the operand_equal_p after callgraph is compared. Similarly for LABEL_DECLs,
> >> we have a map (m_label_bb_map). Please take a look at patch 6/9 in this
> >> series.
> >
> > Hmm, ok, so you basically replace recursive calls to operand_equal_p with

_recursive calls_

> >
> >   operand_equal_valueize (t1, t2, 0)
> >   || operand_equal_p (t1, t2, 0)
> >
> > no?
>
> This is not going to work ..

I wonder if

class base
{
  virtual operand_equal_p (tree a, tree b, int f);
};

base::operand_equal_p (tree a, tree b, int f)
{
  as-is now, recursing to virtual operand_equal_p
}

class deriv : public base
{
  vritual operand_equal_p (tree a, tree b, int f);
};

deriv::operand_equal_p (tree a, tree b, int f)
{
  // just example
  if (TREE_CODE (a) == TREE_CODE (b)
     && TREE_CODE (a) == FUNCTION_DECL)
    return true;

  return base::operand_equal_p (tree a, tree b, int f);
}

would work?  ICF would call deriv::operand_equal_p and
base::operand_equal_p would recurse via the derived implementation.

That at least is cleaner from the "looks".

> >  But the same could be achieved by actually making t1 and t2 equal
> > according to operand_equal_p rules via the valueization hook?  So replace
> > FUNCTION_DECLs with their prevailing ones, LABEL_DECLs with theirs, etc.
> >
> > As given your abstraction is quite awkward to use, say, from value-numbering
> > which knows how to "valueize" a single tree but doesn't compare things.
> >
> > To make it work for your case you'd valueize not only SSA names but also
> > all DECL_P I guess.  After all your operand_equal_valueize only does
> > something for "leafs" but is called for all intermediate expressions as well.
>
> ... because I need to be called for all intermediate expression. One simple
> example can be a ADDR_EXPR of a DECL. The first call will recursively call
> operand_equal_p for the DECL and the DECL can be compared with operand_equal_valueize
> in ICF.
>
> Note that current ICF code is more complex than only selection of a canonical
> form of a tree.
>
> I'm not saying the suggested API change is beautiful. But having a more specific
> equal hook seams to me a reasonable extension to current operand_equal_p.
> Moreover, we'll be able to kill all the ICF duplicate comparison machinery.

I wonder if all FUNCTION_DECL are really equal.  If you just compare
the callgraph
you don't notice differences in the following (with disabled DSE/FRE
to retain both
stores to *dest)

void fna();
void fnb();

void foo (void *dest)
{
  *dest = (void *)fna;
  *dest = (void *)fnb;
}

void bar (void *dest)
{
  *dest = (void *)fnb;
  *dest = (void *)fna;
}

and if you compare IPA refs you'd need to identify the ref stmts as the same?


> Martin
>
> >
> > Richard.
> >
> >> Thanks,
> >> Martin
> >>
> >>>
> >>> Thanks,
> >>> Richard.
> >>>
> >>>> Martin
> >>>>
> >>>>>
> >>>>> Richard.
> >>>>>
> >>>>>> Thanks,
> >>>>>> Martin
> >>>>>>
> >>>>>>> IMHO that would also make the logic within
> >>>>>>> operand_equal_p clearer, because you don't have to think about all the
> >>>>>>> potential callback functions that might be called.
> >>>>>>>
> >>>>>>>
> >>>>>>> Ciao,
> >>>>>>> Michael.
> >>>>>>>
> >>>>>>
> >>>>
> >>
>