The match.pd patterns to merge two vector permutes into one fail when a
potentially no-op view convert expression 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.
---
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..1f91b9857c8 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)
+ && TYPE_SIZE (TREE_TYPE (TREE_TYPE (@0)))
+ == TYPE_SIZE (TREE_TYPE (TREE_TYPE (@1))))
(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" } } */
--
2.44.0
On Fri, May 24, 2024 at 11:30 AM Richard Biener <rguent...@suse.de> wrote:
>
> On Fri, 24 May 2024, Manolis Tsamis wrote:
>
> > On Fri, May 24, 2024 at 10:46 AM Richard Biener <rguent...@suse.de> wrote:
> > >
> > > On Fri, 24 May 2024, Manolis Tsamis wrote:
> > >
> > > > 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.
> > >
> > > Err, yes - you want to check the element sizes of course.
> > >
> > From what I understand, checking the element size should be enough.
> > Otherwise we can check both TYPE_SIZE and element_precision to be
> > equal.
> > So OK to commit with just element_precision?
>
> Please just check the element size. I'm always worried when
> using TYPE_PRECISION on FP types and for shuffles it's really
> only about size.
>
> > BTW I also noticed from these testcases that there is a gcc 13 -> 14
> > regression with weird XORs being introduced:
> >
> > typedef int veci __attribute__ ((vector_size (4 * sizeof (int))));
> > void fun (veci *a, veci *b, veci *c) {
> > *c = __builtin_shufflevector (*a, *b, 0, 5, 2, 7);
> > }
> >
> > gcc 13.3:
> > adrp x3, .LC0
> > ldr q0, [x0]
> > ldr q1, [x1]
> > ldr q2, [x3, #:lo12:.LC0]
> > tbl v0.16b, {v0.16b - v1.16b}, v2.16b
> > str q0, [x2]
> >
> > gcc 14.1:
> > ldr q30, [x1]
> > adrp x3, .LC0
> > ldr q31, [x0]
> > ldr q29, [x3, #:lo12:.LC0]
> > eor v31.16b, v31.16b, v30.16b
> > eor v30.16b, v31.16b, v30.16b
> > eor v31.16b, v31.16b, v30.16b
> > tbl v30.16b, {v30.16b - v31.16b}, v29.16b
> > str q30, [x2]
>
> You'd need to bisect that but I'd guess we got some extra
> match patterns triggering?
>
> > Manolis
> >
> > > > 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)
> > > >
> > >
> > > --
> > > 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)
> >
>
> --
> 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)
