On Fri, May 24, 2024 at 9:31 AM Richard Biener <rguent...@suse.de> wrote:
>
> On Wed, 22 May 2024, Manolis Tsamis wrote:
>
> > The match.pd patterns to merge two vector permutes into one fail when a
> > potentially no-op view convert expressions is between the two permutes.
> > This change lifts this restriction.
> >
> > gcc/ChangeLog:
> >
> > * match.pd: Allow no-op view_convert between permutes.
> >
> > gcc/testsuite/ChangeLog:
> >
> > * gcc.dg/fold-perm-2.c: New test.
> >
> > Signed-off-by: Manolis Tsamis <manolis.tsa...@vrull.eu>
> > ---
> >
> > gcc/match.pd | 14 ++++++++------
> > gcc/testsuite/gcc.dg/fold-perm-2.c | 16 ++++++++++++++++
> > 2 files changed, 24 insertions(+), 6 deletions(-)
> > create mode 100644 gcc/testsuite/gcc.dg/fold-perm-2.c
> >
> > diff --git a/gcc/match.pd b/gcc/match.pd
> > index 07e743ae464..cbb3c5d86e0 100644
> > --- a/gcc/match.pd
> > +++ b/gcc/match.pd
> > @@ -10039,19 +10039,21 @@ and,
> > d = VEC_PERM_EXPR <a, b, NEW_VCST>; */
> >
> > (simplify
> > - (vec_perm (vec_perm@0 @1 @2 VECTOR_CST@3) @0 VECTOR_CST@4)
> > + (vec_perm (view_convert?@0 (vec_perm@1 @2 @3 VECTOR_CST@4)) @0
> > VECTOR_CST@5)
> > (if (TYPE_VECTOR_SUBPARTS (type).is_constant ())
> > (with
> > {
> > machine_mode result_mode = TYPE_MODE (type);
> > - machine_mode op_mode = TYPE_MODE (TREE_TYPE (@1));
> > + machine_mode op_mode = TYPE_MODE (TREE_TYPE (@2));
> > int nelts = TYPE_VECTOR_SUBPARTS (type).to_constant ();
> > vec_perm_builder builder0;
> > vec_perm_builder builder1;
> > vec_perm_builder builder2 (nelts, nelts, 1);
> > }
> > - (if (tree_to_vec_perm_builder (&builder0, @3)
> > - && tree_to_vec_perm_builder (&builder1, @4))
> > + (if (tree_to_vec_perm_builder (&builder0, @4)
> > + && tree_to_vec_perm_builder (&builder1, @5)
> > + && element_precision (TREE_TYPE (@0))
> > + == element_precision (TREE_TYPE (@1)))
>
> I think you want to check TYPE_SIZE (TREE_TYPE (@0/@1)) for equality
> instead.
>
I think TYPE_SIZE is not enough as we need the vector elements to have
the same size, not just the vector as a whole.
For example, when using the TYPE_SIZE check instead the following
testcase miscompiles
typedef int veci __attribute__ ((vector_size (4 * sizeof (int))));
typedef double vecd __attribute__ ((vector_size (2 * sizeof (double))));
void fun (veci *a, veci *b, veci *c)
{
char data[16];
veci r1 = __builtin_shufflevector (*a, *b, 0, 5, 2, 7);
vecd r2;
__builtin_memcpy(data, &r1, sizeof(veci));
__builtin_memcpy(&r2, data, sizeof(vecd));
vecd r3 = __builtin_shufflevector (r2, r2, 1, 0);
__builtin_memcpy(data, &r3, sizeof(vecd));
__builtin_memcpy(c, data, sizeof(veci));
}
To:
ldr q31, [x0]
rev64 v31.4s, v31.4s
str q31, [x2]
ret
> Otherwise OK.
>
> Thanks,
> Richard.
>
> > (with
> > {
> > vec_perm_indices sel0 (builder0, 2, nelts);
> > @@ -10073,10 +10075,10 @@ and,
> > ? (!can_vec_perm_const_p (result_mode, op_mode, sel0, false)
> > || !can_vec_perm_const_p (result_mode, op_mode, sel1,
> > false))
> > : !can_vec_perm_const_p (result_mode, op_mode, sel1, false)))
> > - op0 = vec_perm_indices_to_tree (TREE_TYPE (@4), sel2);
> > + op0 = vec_perm_indices_to_tree (TREE_TYPE (@5), sel2);
> > }
> > (if (op0)
> > - (vec_perm @1 @2 { op0; })))))))
> > + (view_convert (vec_perm @2 @3 { op0; }))))))))
> >
> > /* Merge
> > c = VEC_PERM_EXPR <a, b, VCST0>;
> > diff --git a/gcc/testsuite/gcc.dg/fold-perm-2.c
> > b/gcc/testsuite/gcc.dg/fold-perm-2.c
> > new file mode 100644
> > index 00000000000..1a4ab4065de
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/fold-perm-2.c
> > @@ -0,0 +1,16 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O -fdump-tree-fre1" } */
> > +
> > +typedef int veci __attribute__ ((vector_size (4 * sizeof (int))));
> > +typedef unsigned int vecu __attribute__ ((vector_size (4 * sizeof
> > (unsigned int))));
> > +
> > +void fun (veci *a, veci *b, veci *c)
> > +{
> > + veci r1 = __builtin_shufflevector (*a, *b, 0, 5, 2, 7);
> > + vecu r2 = __builtin_convertvector (r1, vecu);
> > + vecu r3 = __builtin_shufflevector (r2, r2, 2, 3, 1, 0);
> > + *c = __builtin_convertvector (r3, veci);
> > +}
> > +
> > +/* { dg-final { scan-tree-dump "VEC_PERM_EXPR.*{ 2, 7, 5, 0 }" "fre1" } }
> > */
> > +/* { dg-final { scan-tree-dump-times "VEC_PERM_EXPR" 1 "fre1" } } */
> >
>
> --
> Richard Biener <rguent...@suse.de>
> SUSE Software Solutions Germany GmbH,
> Frankenstrasse 146, 90461 Nuernberg, Germany;
> GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
