Re: [PATCH] Always use TYPE_MODE instead of DECL_MODE for vector field

[Richard Biener <richard.guenther@gmail.com>]

On Fri, Oct 21, 2022 at 6:18 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Fri, Oct 21, 2022 at 2:33 AM Richard Biener
> <richard.guenther@gmail.com> wrote:
> >
> > On Thu, Oct 20, 2022 at 6:58 PM H.J. Lu via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> > >
> > > commit e034c5c895722e0092d2239cd8c2991db77d6d39
> > > Author: Jakub Jelinek <jakub@redhat.com>
> > > Date:   Sat Dec 2 08:54:47 2017 +0100
> > >
> > >         PR target/78643
> > >         PR target/80583
> > >         * expr.c (get_inner_reference): If DECL_MODE of a non-bitfield
> > >         is BLKmode for vector field with vector raw mode, use TYPE_MODE
> > >         instead of DECL_MODE.
> > >
> > > fixed the case where DECL_MODE of a vector field is BLKmode and its
> > > TYPE_MODE is a vector mode because of target attribute.  Remove the
> > > BLKmode check for the case where DECL_MODE of a vector field is a vector
> > > mode and its TYPE_MODE is BLKmode because of target attribute.
> > >
> > > gcc/
> > >
> > >         PR target/107304
> > >         * expr.c (get_inner_reference): Always use TYPE_MODE for vector
> > >         field with vector raw mode.
> > >
> > > gcc/testsuite/
> > >
> > >         PR target/107304
> > >         * gcc.target/i386/pr107304.c: New test.
> > > ---
> > >  gcc/expr.cc                              |  3 +-
> > >  gcc/testsuite/gcc.target/i386/pr107304.c | 39 ++++++++++++++++++++++++
> > >  2 files changed, 40 insertions(+), 2 deletions(-)
> > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr107304.c
> > >
> > > diff --git a/gcc/expr.cc b/gcc/expr.cc
> > > index efe387e6173..9145193c2c1 100644
> > > --- a/gcc/expr.cc
> > > +++ b/gcc/expr.cc
> > > @@ -7905,8 +7905,7 @@ get_inner_reference (tree exp, poly_int64_pod *pbitsize,
> > >           /* For vector fields re-check the target flags, as DECL_MODE
> > >              could have been set with different target flags than
> > >              the current function has.  */
> > > -         if (mode == BLKmode
> > > -             && VECTOR_TYPE_P (TREE_TYPE (field))
> > > +         if (VECTOR_TYPE_P (TREE_TYPE (field))
> > >               && VECTOR_MODE_P (TYPE_MODE_RAW (TREE_TYPE (field))))
> >
> > Isn't the check on TYPE_MODE_RAW also wrong then?  Btw, the mode could
>
> TYPE_MODE_RAW is always set to a vector mode for a vector type:
>
>        /* Find an appropriate mode for the vector type.  */
>         if (TYPE_MODE (type) == VOIDmode)
>           SET_TYPE_MODE (type,
>                          mode_for_vector (SCALAR_TYPE_MODE (innertype),
>                                           nunits).else_blk ());

But mode_for_vector can return a MODE_INT!

  /* For integers, try mapping it to a same-sized scalar mode.  */
  if (GET_MODE_CLASS (innermode) == MODE_INT)
    {
      poly_uint64 nbits = nunits * GET_MODE_BITSIZE (innermode);
      if (int_mode_for_size (nbits, 0).exists (&mode)
          && have_regs_of_mode[mode])
        return mode;

> But TYPE_MODE returns BLKmode if the vector mode is unsupported.
>
> > also be an integer mode.
>
> For a vector field, mode is either BLK mode or the vector mode.  Jakub,
> can you comment on it?

I think that for

typedef int v2si __attribute__((vector_size(8)));

struct X { int i; v2si j; };

v2si should get DImode with -mno-sse?

> >
> > >             mode = TYPE_MODE (TREE_TYPE (field));
> > >         }
> > > diff --git a/gcc/testsuite/gcc.target/i386/pr107304.c b/gcc/testsuite/gcc.target/i386/pr107304.c
> > > new file mode 100644
> > > index 00000000000..24d68795e7f
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.target/i386/pr107304.c
> > > @@ -0,0 +1,39 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O0 -march=tigerlake" } */
> > > +
> > > +#include <stdint.h>
> > > +
> > > +typedef union {
> > > +  uint8_t v __attribute__((aligned(256))) __attribute__ ((vector_size(64 * sizeof(uint8_t))));
> > > +  uint8_t i[64] __attribute__((aligned(256)));
> > > +} stress_vec_u8_64_t;
> > > +
> > > +typedef struct {
> > > + struct {
> > > +  stress_vec_u8_64_t s;
> > > +  stress_vec_u8_64_t o;
> > > +  stress_vec_u8_64_t mask1;
> > > +  stress_vec_u8_64_t mask2;
> > > + } u8_64;
> > > +} stress_vec_data_t;
> > > +
> > > +__attribute__((target_clones("arch=alderlake", "default")))
> > > +void
> > > +stress_vecshuf_u8_64(stress_vec_data_t *data)
> > > +{
> > > +  stress_vec_u8_64_t *__restrict s;
> > > +  stress_vec_u8_64_t *__restrict mask1;
> > > +  stress_vec_u8_64_t *__restrict mask2;
> > > +  register int i;
> > > +
> > > +  s = &data->u8_64.s;
> > > +  mask1 = &data->u8_64.mask1;
> > > +  mask2 = &data->u8_64.mask2;
> > > +
> > > +  for (i = 0; i < 256; i++) {  /* was i < 65536 */
> > > +      stress_vec_u8_64_t tmp;
> > > +
> > > +      tmp.v = __builtin_shuffle(s->v, mask1->v);
> > > +      s->v = __builtin_shuffle(tmp.v, mask2->v);
> > > +  }
> > > +}
> > > --
> > > 2.37.3
> > >
>
>
>
> --
> H.J.