Re: [PATCH] match.pd: Implement missed optimization (x << c) >> c -> -(x & 1) [PR101955]

[Drew Ross <drross@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 4012 bytes --]

> With that fixed I think for non-vector integrals the above is the most
suitable
> canonical form of a sign-extension.  Note it should also work for any
other
> constant shift amount - just use the appropriate intermediate precision
for
> the truncating type.
> We _might_ want
> to consider to only use the converts when the intermediate type has
> mode precision (and as a special case allow one bit as in your above case)
> so it can expand to (sign_extend:<outer> (subreg:<inner> reg)).

Here is a pattern that that only matches to truncations that result in mode
precision (or precision of 1):

(simplify
 (rshift (nop_convert? (lshift @0 INTEGER_CST@1)) @@1)
 (if (INTEGRAL_TYPE_P (type)
      && !TYPE_UNSIGNED (type)
      && wi::gt_p (element_precision (type), wi::to_wide (@1), TYPE_SIGN
(TREE_TYPE (@1))))
  (with {
    int width = element_precision (type) - tree_to_uhwi (@1);
    tree stype = build_nonstandard_integer_type (width, 0);
   }
   (if (TYPE_PRECISION (stype) == 1 || type_has_mode_precision_p (stype))
    (convert (convert:stype @0))))))

Look ok?

> You might also want to verify what RTL expansion
> produces before/after - it at least shouldn't be worse.

The RTL is slightly better for the mode precision cases and slightly worse
for the precision 1 case.

> That said - do you have any testcase where the canonicalization is an
enabler
> for further transforms or was this requested stand-alone?

No, I don't have any specific test cases. This patch is just in response to
pr101955 <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101955>.

On Tue, Jul 25, 2023 at 2:55 AM Richard Biener <richard.guenther@gmail.com>
wrote:

> On Mon, Jul 24, 2023 at 9:42 PM Jakub Jelinek <jakub@redhat.com> wrote:
> >
> > On Mon, Jul 24, 2023 at 03:29:54PM -0400, Drew Ross via Gcc-patches
> wrote:
> > > So would something like
> > >
> > > (simplify
> > >  (rshift (nop_convert? (lshift @0 INTEGER_CST@1)) @@1)
> > >  (with { tree stype = build_nonstandard_integer_type (1, 0); }
> > >  (if (INTEGRAL_TYPE_P (type)
> > >       && !TYPE_UNSIGNED (type)
> > >       && wi::eq_p (wi::to_wide (@1), element_precision (type) - 1))
> > >   (convert (convert:stype @0)))))
> > >
> > > work?
> >
> > Certainly swap the if and with and the (with then should be indented by 1
> > column to the right of (if and (convert one further (the reason for the
> > swapping is not to call build_nonstandard_integer_type when it will not
> be
> > needed, which will be probably far more often then an actual match).
>
> With that fixed I think for non-vector integrals the above is the most
> suitable
> canonical form of a sign-extension.  Note it should also work for any other
> constant shift amount - just use the appropriate intermediate precision for
> the truncating type.  You might also want to verify what RTL expansion
> produces before/after - it at least shouldn't be worse.  We _might_ want
> to consider to only use the converts when the intermediate type has
> mode precision (and as a special case allow one bit as in your above case)
> so it can expand to (sign_extend:<outer> (subreg:<inner> reg)).
>
> > As discussed privately, the above isn't what we want for vectors and the
> 2
> > shifts are probably best on most arches because even when using -(x & 1)
> the
> > { 1, 1, 1, ... } vector would often needed to be loaded from memory.
>
> I think for vectors a vpcmpgt {0,0,0,..}, %xmm is the cheapest way of
> producing the result.  Note that to reflect this on GIMPLE you'd need
>
>   _2 = _1 < { 0,0...};
>   res = _2 ? { -1, -1, ...} : { 0, 0,...};
>
> because whether the ISA has a way to produce all-ones masks isn't known.
>
> For scalars using -(T)(_1 < 0) would also be possible.
>
> That said - do you have any testcase where the canonicalization is an
> enabler
> for further transforms or was this requested stand-alone?
>
> Thanks,
> Richard.
>
> >         Jakub
> >
>
>