PR91166 - Unfolded ZIPs of constants

[Richard Biener]

On Tue, 23 Jul 2019, Prathamesh Kulkarni wrote:

> On Tue, 23 Jul 2019 at 16:36, Richard Biener <rguenther@suse.de> wrote:
> >
> > On Tue, 23 Jul 2019, Prathamesh Kulkarni wrote:
> >
> > > On Fri, 19 Jul 2019 at 18:12, Richard Sandiford
> > > <richard.sandiford@arm.com> wrote:
> > > >
> > > > Not really my area, but FWIW...
> > > >
> > > > Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:
> > > > > Hi,
> > > > > The attached patch tries to fix PR91166.
> > > > > Does it look OK ?
> > > > > Bootstrap+test in progress on aarch64-linux-gnu and x86_64-unknown-linux-gnu.
> > > > >
> > > > > Thanks,
> > > > > Prathamesh
> > > > >
> > > > > 2019-07-17  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>
> > > > >
> > > > >       PR middle-end/91166
> > > > >       * match.pd (vec_perm_expr(v, v, mask) -> v): New pattern.
> > > > >       (define_predicates): Add entry for uniform_vector_p.
> > > > >
> > > > > testsuite/
> > > > >       * gcc.target/aarch64/sve/pr91166.c: New test.
> > > > >
> > > > > diff --git a/gcc/match.pd b/gcc/match.pd
> > > > > index 4a7aa0185d8..2ad98c28fd8 100644
> > > > > --- a/gcc/match.pd
> > > > > +++ b/gcc/match.pd
> > > > > @@ -36,7 +36,8 @@ along with GCC; see the file COPYING3.  If not see
> > > > >     integer_valued_real_p
> > > > >     integer_pow2p
> > > > >     uniform_integer_cst_p
> > > > > -   HONOR_NANS)
> > > > > +   HONOR_NANS
> > > > > +   uniform_vector_p)
> > > > >
> > > > >  /* Operator lists.  */
> > > > >  (define_operator_list tcc_comparison
> > > > > @@ -5568,3 +5569,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > > > >           { bitsize_int (at * tree_to_uhwi (TYPE_SIZE (TREE_TYPE (type)))); })
> > > > >         (if (changed)
> > > > >          (vec_perm { op0; } { op1; } { op2; }))))))))))
> > > > > +
> > > > > +/* VEC_PERM_EXPR (v, v, mask) -> v where v contains same element.  */
> > > > > +(simplify
> > > > > + (vec_perm (vec_duplicate@0 @1) @0 @2)
> > > > > + { @0; })
> > > > > +
> > > > > +(simplify
> > > > > + (vec_perm uniform_vector_p@0 @0 @1)
> > > > > + { @0; })
> > > >
> > > > No need for the curly braces here, can use "@0" as the target of
> > > > the simplification.
> > > >
> > > > It'd probably be worth using (match ...) to define a new predicate
> > > > that handles (vec_duplicate ...), VECTOR_CST and CONSTRUCTOR,
> > > > calling into uniform_vector_p for the latter two.
> > > Hi,
> > > Thanks for the suggestions.
> > > Does this version look OK ?
> >
> > Can you write
> >
> > +(simplify
> > + (vec_perm (vec_same_elem_p@0 @1) @0 @2)
> > + @0)
> >
> > as
> >
> >  (vec_perm vec_same_elem_p@0 @0 @1)
> >
> > ?
> (simplify
>  (vec_perm vec_same_elem_p@0 @0 @1)
>  @0)
> 
> results in:
> gimple-match.c: In function ‘bool
> gimple_simplify_VEC_PERM_EXPR(gimple_match_op*, gimple**, tree_node*
> (*)(tree), code_helper, tree, tree, tree, tree)’:
> gimple-match.c:103031:36: error: cannot convert ‘tree_node* (*)(tree)’
> {aka ‘tree_node* (*)(tree_node*)’} to ‘tree_node**’
>    if (gimple_vec_same_elem_p (op0, valueize))
>                                     ^~~~~~~~
> 
> because gimple_vec_same_elem_p has tree *res_ops as 2nd param and
> we're passing valueize as 2nd arg.

Ah, you need the

(match vec_same_elem_p
 @0
 (if (uniform_vector_p (@0)))

(match vec_same_elem_p
 (vec_duplicate @0))

form then.

> Thanks,
> Prathamesh
> >
> > Otherwise looks OK.
> >
> > Thanks,
> > Richard.
> >
> > > Thanks,
> > > Prathamesh
> > >
> > > >
> > > > Thanks,
> > > > Richard
> > > >
> > > > > diff --git a/gcc/testsuite/gcc.target/aarch64/sve/pr91166.c b/gcc/testsuite/gcc.target/aarch64/sve/pr91166.c
> > > > > new file mode 100644
> > > > > index 00000000000..42654be3b31
> > > > > --- /dev/null
> > > > > +++ b/gcc/testsuite/gcc.target/aarch64/sve/pr91166.c
> > > > > @@ -0,0 +1,20 @@
> > > > > +/* { dg-do compile } */
> > > > > +/* { dg-options "-O3 -march=armv8.2-a+sve -fdump-tree-optimized" } */
> > > > > +
> > > > > +void
> > > > > +f1 (double x[][4])
> > > > > +{
> > > > > +  for (int i = 0; i < 4; ++i)
> > > > > +    for (int j = 0; j < 4; ++j)
> > > > > +      x[i][j] = 0;
> > > > > +}
> > > > > +
> > > > > +void
> > > > > +f2 (double x[][4], double y)
> > > > > +{
> > > > > +  for (int i = 0; i < 4; ++i)
> > > > > +    for (int j = 0; j < 4; ++j)
> > > > > +      x[i][j] = y;
> > > > > +}
> > > > > +
> > > > > +/* { dg-final { scan-tree-dump-not "VEC_PERM_EXPR" "optimized"} } */
> > >
> >
> > --
> > Richard Biener <rguenther@suse.de>
> > SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany;
> > GF: Felix Imendörffer, Mary Higgins, Sri Rasiah; HRB 21284 (AG Nürnberg)
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany;
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah; HRB 21284 (AG Nürnberg)