On Fri, 5 Feb 2021, Kyrylo Tkachov wrote:
> Hi Richard,
>
> > -----Original Message-----
> > From: Gcc-patches <gcc-patches-boun...@gcc.gnu.org> On Behalf Of
> > Richard Biener
> > Sent: 01 October 2020 14:15
> > To: gcc-patches@gcc.gnu.org
> > Subject: [PATCH] tree-optimization/97236 - fix bad use of
> > VMAT_CONTIGUOUS
> >
> > This avoids using VMAT_CONTIGUOUS with single-element interleaving
> > when using V1mode vectors. Instead keep VMAT_ELEMENTWISE but
> > continue to avoid load-lanes and gathers.
> >
> > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
>
> I've checked that this fix also fixes the recently-reported PR 98949 on
> aarch64 on the GCC 9 branch.
> I've bootstrapped and tested it on the branch on aarch64-none-linux.
> Is it okay to backport to the branch?
Yes.
Thanks,
Richard.
> Thanks,
> Kyrill
>
> >
> > Richard.
> >
> > 2020-10-01 Richard Biener <rguent...@suse.de>
> >
> > PR tree-optimization/97236
> > * tree-vect-stmts.c (get_group_load_store_type): Keep
> > VMAT_ELEMENTWISE for single-element vectors.
> >
> > * gcc.dg/vect/pr97236.c: New testcase.
> > ---
> > gcc/testsuite/gcc.dg/vect/pr97236.c | 43
> > +++++++++++++++++++++++++++++
> > gcc/tree-vect-stmts.c | 20 ++++++--------
> > 2 files changed, 52 insertions(+), 11 deletions(-)
> > create mode 100644 gcc/testsuite/gcc.dg/vect/pr97236.c
> >
> > diff --git a/gcc/testsuite/gcc.dg/vect/pr97236.c
> > b/gcc/testsuite/gcc.dg/vect/pr97236.c
> > new file mode 100644
> > index 00000000000..03e0cc38984
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/vect/pr97236.c
> > @@ -0,0 +1,43 @@
> > +typedef unsigned char __uint8_t;
> > +typedef __uint8_t uint8_t;
> > +typedef struct plane_t {
> > + uint8_t *p_pixels;
> > + int i_lines;
> > + int i_pitch;
> > +} plane_t;
> > +
> > +typedef struct {
> > + plane_t p[5];
> > +} picture_t;
> > +
> > +#define N 4
> > +
> > +void __attribute__((noipa))
> > +picture_Clone(picture_t *picture, picture_t *res)
> > +{
> > + for (int i = 0; i < N; i++) {
> > + res->p[i].p_pixels = picture->p[i].p_pixels;
> > + res->p[i].i_lines = picture->p[i].i_lines;
> > + res->p[i].i_pitch = picture->p[i].i_pitch;
> > + }
> > +}
> > +
> > +int
> > +main()
> > +{
> > + picture_t aaa, bbb;
> > + uint8_t pixels[10] = {1, 1, 1, 1, 1, 1, 1, 1};
> > +
> > + for (unsigned i = 0; i < N; i++)
> > + aaa.p[i].p_pixels = pixels;
> > +
> > + picture_Clone (&aaa, &bbb);
> > +
> > + uint8_t c;
> > + for (unsigned i = 0; i < N; i++)
> > + c += bbb.p[i].p_pixels[0];
> > +
> > + if (c != N)
> > + __builtin_abort ();
> > + return 0;
> > +}
> > diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
> > index 191957c3543..3575f25241f 100644
> > --- a/gcc/tree-vect-stmts.c
> > +++ b/gcc/tree-vect-stmts.c
> > @@ -2235,25 +2235,23 @@ get_group_load_store_type (vec_info *vinfo,
> > stmt_vec_info stmt_info,
> > /* First cope with the degenerate case of a single-element
> > vector. */
> > if (known_eq (TYPE_VECTOR_SUBPARTS (vectype), 1U))
> > - *memory_access_type = VMAT_CONTIGUOUS;
> > + ;
> >
> > /* Otherwise try using LOAD/STORE_LANES. */
> > - if (*memory_access_type == VMAT_ELEMENTWISE
> > - && (vls_type == VLS_LOAD
> > - ? vect_load_lanes_supported (vectype, group_size,
> > masked_p)
> > - : vect_store_lanes_supported (vectype, group_size,
> > - masked_p)))
> > + else if (vls_type == VLS_LOAD
> > + ? vect_load_lanes_supported (vectype, group_size,
> > masked_p)
> > + : vect_store_lanes_supported (vectype, group_size,
> > + masked_p))
> > {
> > *memory_access_type = VMAT_LOAD_STORE_LANES;
> > overrun_p = would_overrun_p;
> > }
> >
> > /* If that fails, try using permuting loads. */
> > - if (*memory_access_type == VMAT_ELEMENTWISE
> > - && (vls_type == VLS_LOAD
> > - ? vect_grouped_load_supported (vectype,
> > single_element_p,
> > - group_size)
> > - : vect_grouped_store_supported (vectype, group_size)))
> > + else if (vls_type == VLS_LOAD
> > + ? vect_grouped_load_supported (vectype,
> > single_element_p,
> > + group_size)
> > + : vect_grouped_store_supported (vectype, group_size))
> > {
> > *memory_access_type = VMAT_CONTIGUOUS_PERMUTE;
> > overrun_p = would_overrun_p;
> > --
> > 2.26.2
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)
