Re: [PATCH v2] x86: Update STORE_MAX_PIECES

[Uros Bizjak <ubizjak@gmail.com>]

On Wed, Aug 4, 2021 at 3:34 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Tue, Aug 3, 2021 at 6:56 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > 1. Update x86 STORE_MAX_PIECES to use OImode and XImode only if inter-unit
> > move is enabled since x86 uses vec_duplicate, which is enabled only when
> > inter-unit move is enabled, to implement store_by_pieces.
> > 2. Update op_by_pieces_d::op_by_pieces_d to set m_max_size to
> > STORE_MAX_PIECES for store_by_pieces and to COMPARE_MAX_PIECES for
> > compare_by_pieces.
> >
> > gcc/
> >
> >         PR target/101742
> >         * expr.c (op_by_pieces_d::op_by_pieces_d): Set m_max_size to
> >         STORE_MAX_PIECES for store_by_pieces and to COMPARE_MAX_PIECES
> >         for compare_by_pieces.
> >         * config/i386/i386.h (STORE_MAX_PIECES): Use OImode and XImode
> >         only if TARGET_INTER_UNIT_MOVES_TO_VEC is true.
> >
> > gcc/testsuite/
> >
> >         PR target/101742
> >         * gcc.target/i386/pr101742a.c: New test.
> >         * gcc.target/i386/pr101742b.c: Likewise.
> > ---
> >  gcc/config/i386/i386.h                    | 20 +++++++++++---------
> >  gcc/expr.c                                |  6 +++++-
> >  gcc/testsuite/gcc.target/i386/pr101742a.c | 16 ++++++++++++++++
> >  gcc/testsuite/gcc.target/i386/pr101742b.c |  4 ++++
> >  4 files changed, 36 insertions(+), 10 deletions(-)
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr101742a.c
> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr101742b.c
> >
> > diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h
> > index bed9cd9da18..9b416abd5f4 100644
> > --- a/gcc/config/i386/i386.h
> > +++ b/gcc/config/i386/i386.h
> > @@ -1783,15 +1783,17 @@ typedef struct ix86_args {
> >  /* STORE_MAX_PIECES is the number of bytes at a time that we can
> >     store efficiently.  */
> >  #define STORE_MAX_PIECES \
> > -  ((TARGET_AVX512F && !TARGET_PREFER_AVX256) \
> > -   ? 64 \
> > -   : ((TARGET_AVX \
> > -       && !TARGET_PREFER_AVX128 \
> > -       && !TARGET_AVX256_SPLIT_UNALIGNED_STORE) \
> > -      ? 32 \
> > -      : ((TARGET_SSE2 \
> > -         && TARGET_SSE_UNALIGNED_STORE_OPTIMAL) \
> > -        ? 16 : UNITS_PER_WORD)))
> > +  (TARGET_INTER_UNIT_MOVES_TO_VEC \
> > +   ? ((TARGET_AVX512F && !TARGET_PREFER_AVX256) \
> > +      ? 64 \
> > +      : ((TARGET_AVX \
> > +         && !TARGET_PREFER_AVX128 \
> > +         && !TARGET_AVX256_SPLIT_UNALIGNED_STORE) \
> > +         ? 32 \
> > +         : ((TARGET_SSE2 \
> > +             && TARGET_SSE_UNALIGNED_STORE_OPTIMAL) \
> > +             ? 16 : UNITS_PER_WORD))) \
> > +   : UNITS_PER_WORD)
> >
> >  /* If a memory-to-memory move would take MOVE_RATIO or more simple
> >     move-instruction pairs, we will do a cpymem or libcall instead.
>
> expr.c has been fixed.   Here is the v2 patch for x86 backend.
> OK for master?

OK, but please add the comment about vec_duplicate before the define
to explain the situation with TARGET_INTER_UNIT_MOVES_TO_VEC.

Thanks,
Uros.