Re: [PATCH] fold-const: Handle AND, IOR, XOR with stepped vectors [PR112971].

[Richard Biener <rguenther@suse.de>]

On Wed, 20 Dec 2023, Richard Sandiford wrote:

> Richard Biener <rguenther@suse.de> writes:
> > On Tue, 19 Dec 2023, Andrew Pinski wrote:
> >
> >> On Tue, Dec 19, 2023 at 2:40?AM Richard Sandiford
> >> <richard.sandiford@arm.com> wrote:
> >> >
> >> > Richard Biener <rguenther@suse.de> writes:
> >> > > On Tue, 19 Dec 2023, juzhe.zhong@rivai.ai wrote:
> >> > >
> >> > >> Hi, Richard.
> >> > >>
> >> > >> After investigating the codes:
> >> > >> /* Return true if EXPR is the integer constant zero or a complex constant
> >> > >>    of zero, or a location wrapper for such a constant.  */
> >> > >>
> >> > >> bool
> >> > >> integer_zerop (const_tree expr)
> >> > >> {
> >> > >>   STRIP_ANY_LOCATION_WRAPPER (expr);
> >> > >>
> >> > >>   switch (TREE_CODE (expr))
> >> > >>     {
> >> > >>     case INTEGER_CST:
> >> > >>       return wi::to_wide (expr) == 0;
> >> > >>     case COMPLEX_CST:
> >> > >>       return (integer_zerop (TREE_REALPART (expr))
> >> > >>               && integer_zerop (TREE_IMAGPART (expr)));
> >> > >>     case VECTOR_CST:
> >> > >>       return (VECTOR_CST_NPATTERNS (expr) == 1
> >> > >>               && VECTOR_CST_DUPLICATE_P (expr)
> >> > >>               && integer_zerop (VECTOR_CST_ENCODED_ELT (expr, 0)));
> >> > >>     default:
> >> > >>       return false;
> >> > >>     }
> >> > >> }
> >> > >>
> >> > >> I wonder whether we can simplify the codes as follows :?
> >> > >>       if (integer_zerop (arg1) || integer_zerop (arg2))
> >> > >>         step_ok_p = (code == BIT_AND_EXPR || code == BIT_IOR_EXPR
> >> > >>                      || code == BIT_XOR_EXPR);
> >> > >
> >> > > Possibly.  I'll let Richard S. comment on the whole structure.
> >> >
> >> > The current code is handling cases that require elementwise arithmetic.
> >> > ISTM that what we're really doing here is identifying cases where
> >> > whole-vector arithmetic is possible instead.  I think that should be
> >> > a separate pre-step, rather than integrated into the current code.
> >> >
> >> > Largely this would consist of writing out match.pd-style folds in
> >> > C++ code, so Andrew's fix in comment 7 seems neater to me.
> >> 
> >> I didn't like the change to match.pd (even with a comment on why)
> >> because it violates the whole idea behind canonicalization of
> >> constants being 2nd operand of commutative and comparison expressions.
> >> Maybe there are only a few limited match/simplify patterns which need
> >> to add the :c for constants not being the 2nd operand but there needs
> >> to be a comment on why :c is needed for this.
> >
> > Agreed.  Note that in theory we of course could define extra
> > canonicalization rules for CST op CST in tree_swap_operands_p,
> > there are some cases those surivive, mostly around FP and
> > dynamic rounding state.  I'd rather go that route if we decide
> > that match.pd should catch this.
> 
> I don't think that'll help in all cases though.  E.g. consider an
> unfoldable:
> 
>   (or 1 CST)
> 
> where CST is "complicated".  We'd probably canonicalise that to:
> 
>   (or CST 1)
> 
> And that's good if we have:
> 
>   (or (or CST 1) 2)
> 
> since it could be folded to:
> 
>   (or CST 3)
> 
> But there are other constants CST2 for which (or CST CST2) is foldable
> and (op 1 CST2) isn't.  So:
> 
>   (or (or 1 CST) CST2)
> 
> would sometimes be foldable in cases where:
> 
>   (or (or CST 1) CST2)
> 
> isn't.

OK, I was thinking of only handling VECTOR_CST_NELTS_PER_PATTERN
for example as additional (cheap) heuristic so we put
VECTOR_CST_DUPLICATE_P second (this would cover the particular
cases in this thread).

> >> >
> >> > But if this must happen in const_binop instead, then we could have
> >> > a function like:
> >> 
> >> The reasoning of why it should be in const_binop rather than in match
> >> is because both operands are constants.
> >
> > +1
> 
> I can see that argument for the traditional case where all constants
> can be folded at compile time.  But that isn't the case for VLA constants.
> (op CST1 CST2) is sometimes not knowable at compile time.  And I think
> match.pd should apply to those cases just as much as to (op VAR CST).
> 
> VLA vector "constants" are really in intermediate category between
> variable and constant.
> 
> The approach that the patch takes is to add a "rule of thumb" that
> applies to all (op X CST), regardless of whether X is constant.
> It doesn't work by doing constant arithmetic per se.  If we add
> those rules of thumb here, I think it'll keep growing and growing.

But doesn't this mean we can't rely on folding (match.pd) not seeing
un-constant-folded operations and thus proper canonicalization?
Which means we'd possibly have to alter _all_ (op X CST) cases to
use :c?

> Thanks,
> Richard
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)