public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Richard Biener <richard.guenther@gmail.com>
To: Jakub Jelinek <jakub@redhat.com>,
	Eric Botcazou <ebotcazou@adacore.com>,
		GCC Patches <gcc-patches@gcc.gnu.org>,
	Richard Sandiford <richard.sandiford@arm.com>
Subject: Re: [1/2] PR88598: Optimise x * { 0 or 1, 0 or 1, ... }
Date: Mon, 07 Jan 2019 11:35:00 -0000	[thread overview]
Message-ID: <CAFiYyc3zZqPiZi-0=asiczTc+v=hBq=A13mdv_0=AbVv=WmPdA@mail.gmail.com> (raw)
In-Reply-To: <871s5spp6z.fsf@arm.com>

On Fri, Jan 4, 2019 at 1:44 PM Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> Jakub Jelinek <jakub@redhat.com> writes:
> > On Fri, Jan 04, 2019 at 12:13:13PM +0000, Richard Sandiford wrote:
> >> > Can we avoid the gratuitous use of template here?  We were told that C++ would
> >> > be used only when it makes things more straightforward and it's the contrary
> >> > in this case, to wit the need for the ugly RECURSE macro in the middle.
> >>
> >> I did it that way so that it would be easy to add things like
> >> zero_or_minus_onep without cut-&-pasting the whole structure.
> >
> > IMHO we can make such a change only when it is needed.
>
> The other predicates in tree.c suggest that we won't though.
> E.g. there was never any attempt to unify integer_zerop vs. integer_onep
> and real_zerop vs. real_onep.
>
> >> The way to do that in C would be to use a macro for the full
> >> function, but that's even uglier due to the extra backslashes.
> >
> > Or just make the function static inline and pass the function pointers
> > to it as arguments?  If it is inlined, it will be the same, it could be
> > even always_inline if that is really needed.
>
> For that to work for recursive functions I think we'd need to pass the
> caller predicate in too, which means one more function pointer overall.
>
> Anyway, here's the patch without the template.

OK.

Thanks,
Richard.

> Thanks,
> Richard
>
>
> 2019-01-04  Richard Sandiford  <richard.sandiford@arm.com>
>
> gcc/
>         PR tree-optimization/88598
>         * tree.h (initializer_each_zero_or_onep): Declare.
>         * tree.c (initializer_each_zero_or_onep): New function.
>         (signed_or_unsigned_type_for): Handle float types too.
>         (unsigned_type_for, signed_type_for): Update comments accordingly.
>         * match.pd: Fold x * { 0 or 1, 0 or 1, ...} to
>         x & { 0 or -1, 0 or -1, ... }.
>
> gcc/testsuite/
>         PR tree-optimization/88598
>         * gcc.dg/pr88598-1.c: New test.
>         * gcc.dg/pr88598-2.c: Likewise.
>         * gcc.dg/pr88598-3.c: Likewise.
>         * gcc.dg/pr88598-4.c: Likewise.
>         * gcc.dg/pr88598-5.c: Likewise.
>
> Index: gcc/tree.h
> ===================================================================
> --- gcc/tree.h  2019-01-04 12:40:51.000000000 +0000
> +++ gcc/tree.h  2019-01-04 12:40:51.990582844 +0000
> @@ -4506,6 +4506,7 @@ extern tree first_field (const_tree);
>     combinations indicate definitive answers.  */
>
>  extern bool initializer_zerop (const_tree, bool * = NULL);
> +extern bool initializer_each_zero_or_onep (const_tree);
>
>  extern wide_int vector_cst_int_elt (const_tree, unsigned int);
>  extern tree vector_cst_elt (const_tree, unsigned int);
> Index: gcc/tree.c
> ===================================================================
> --- gcc/tree.c  2019-01-04 12:40:51.000000000 +0000
> +++ gcc/tree.c  2019-01-04 12:40:51.990582844 +0000
> @@ -11229,6 +11229,45 @@ initializer_zerop (const_tree init, bool
>      }
>  }
>
> +/* Return true if EXPR is an initializer expression in which every element
> +   is a constant that is numerically equal to 0 or 1.  The elements do not
> +   need to be equal to each other.  */
> +
> +bool
> +initializer_each_zero_or_onep (const_tree expr)
> +{
> +  STRIP_ANY_LOCATION_WRAPPER (expr);
> +
> +  switch (TREE_CODE (expr))
> +    {
> +    case INTEGER_CST:
> +      return integer_zerop (expr) || integer_onep (expr);
> +
> +    case REAL_CST:
> +      return real_zerop (expr) || real_onep (expr);
> +
> +    case VECTOR_CST:
> +      {
> +       unsigned HOST_WIDE_INT nelts = vector_cst_encoded_nelts (expr);
> +       if (VECTOR_CST_STEPPED_P (expr)
> +           && !TYPE_VECTOR_SUBPARTS (TREE_TYPE (expr)).is_constant (&nelts))
> +         return false;
> +
> +       for (unsigned int i = 0; i < nelts; ++i)
> +         {
> +           tree elt = VECTOR_CST_ENCODED_ELT (expr, i);
> +           if (!initializer_each_zero_or_onep (elt))
> +             return false;
> +         }
> +
> +       return true;
> +      }
> +
> +    default:
> +      return false;
> +    }
> +}
> +
>  /* Check if vector VEC consists of all the equal elements and
>     that the number of elements corresponds to the type of VEC.
>     The function returns first element of the vector
> @@ -11672,7 +11711,10 @@ int_cst_value (const_tree x)
>
>  /* If TYPE is an integral or pointer type, return an integer type with
>     the same precision which is unsigned iff UNSIGNEDP is true, or itself
> -   if TYPE is already an integer type of signedness UNSIGNEDP.  */
> +   if TYPE is already an integer type of signedness UNSIGNEDP.
> +   If TYPE is a floating-point type, return an integer type with the same
> +   bitsize and with the signedness given by UNSIGNEDP; this is useful
> +   when doing bit-level operations on a floating-point value.  */
>
>  tree
>  signed_or_unsigned_type_for (int unsignedp, tree type)
> @@ -11702,17 +11744,23 @@ signed_or_unsigned_type_for (int unsigne
>        return build_complex_type (inner2);
>      }
>
> -  if (!INTEGRAL_TYPE_P (type)
> -      && !POINTER_TYPE_P (type)
> -      && TREE_CODE (type) != OFFSET_TYPE)
> +  unsigned int bits;
> +  if (INTEGRAL_TYPE_P (type)
> +      || POINTER_TYPE_P (type)
> +      || TREE_CODE (type) == OFFSET_TYPE)
> +    bits = TYPE_PRECISION (type);
> +  else if (TREE_CODE (type) == REAL_TYPE)
> +    bits = GET_MODE_BITSIZE (SCALAR_TYPE_MODE (type));
> +  else
>      return NULL_TREE;
>
> -  return build_nonstandard_integer_type (TYPE_PRECISION (type), unsignedp);
> +  return build_nonstandard_integer_type (bits, unsignedp);
>  }
>
>  /* If TYPE is an integral or pointer type, return an integer type with
>     the same precision which is unsigned, or itself if TYPE is already an
> -   unsigned integer type.  */
> +   unsigned integer type.  If TYPE is a floating-point type, return an
> +   unsigned integer type with the same bitsize as TYPE.  */
>
>  tree
>  unsigned_type_for (tree type)
> @@ -11722,7 +11770,8 @@ unsigned_type_for (tree type)
>
>  /* If TYPE is an integral or pointer type, return an integer type with
>     the same precision which is signed, or itself if TYPE is already a
> -   signed integer type.  */
> +   signed integer type.  If TYPE is a floating-point type, return a
> +   signed integer type with the same bitsize as TYPE.  */
>
>  tree
>  signed_type_for (tree type)
> Index: gcc/match.pd
> ===================================================================
> --- gcc/match.pd        2019-01-04 12:40:51.000000000 +0000
> +++ gcc/match.pd        2019-01-04 12:40:51.982582910 +0000
> @@ -29,6 +29,7 @@ along with GCC; see the file COPYING3.
>     integer_each_onep integer_truep integer_nonzerop
>     real_zerop real_onep real_minus_onep
>     zerop
> +   initializer_each_zero_or_onep
>     CONSTANT_CLASS_P
>     tree_expr_nonnegative_p
>     tree_expr_nonzero_p
> @@ -194,6 +195,22 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>             || !COMPLEX_FLOAT_TYPE_P (type)))
>     (negate @0)))
>
> +/* Transform x * { 0 or 1, 0 or 1, ... } into x & { 0 or -1, 0 or -1, ...},
> +   unless the target has native support for the former but not the latter.  */
> +(simplify
> + (mult @0 VECTOR_CST@1)
> + (if (initializer_each_zero_or_onep (@1)
> +      && !HONOR_SNANS (type)
> +      && !HONOR_SIGNED_ZEROS (type))
> +  (with { tree itype = FLOAT_TYPE_P (type) ? unsigned_type_for (type) : type; }
> +   (if (itype
> +       && (!VECTOR_MODE_P (TYPE_MODE (type))
> +           || (VECTOR_MODE_P (TYPE_MODE (itype))
> +               && optab_handler (and_optab,
> +                                 TYPE_MODE (itype)) != CODE_FOR_nothing)))
> +    (view_convert (bit_and:itype (view_convert @0)
> +                                (ne @1 { build_zero_cst (type); })))))))
> +
>  (for cmp (gt ge lt le)
>       outp (convert convert negate negate)
>       outn (negate negate convert convert)
> Index: gcc/testsuite/gcc.dg/pr88598-1.c
> ===================================================================
> --- /dev/null   2018-12-31 11:20:29.178325188 +0000
> +++ gcc/testsuite/gcc.dg/pr88598-1.c    2019-01-04 12:40:51.982582910 +0000
> @@ -0,0 +1,27 @@
> +/* { dg-do run } */
> +/* { dg-options "-O -fdump-tree-ccp1" } */
> +
> +typedef int v4si __attribute__ ((vector_size (16)));
> +
> +int
> +main ()
> +{
> +  volatile v4si x1 = { 4, 5, 6, 7 };
> +  volatile v4si x2 = { 10, 11, 12, 13 };
> +  volatile v4si x3 = { 20, 21, 22, 23 };
> +
> +  x1 *= (v4si) { 0, 1, 1, 0 };
> +  x2 *= (v4si) { 1, 0, 0, 1 };
> +  x3 *= (v4si) { 0, 0, 1, 0 };
> +
> +  if (__builtin_memcmp ((void *) &x1, &(v4si) { 0, 5, 6, 0 }, sizeof (v4si))
> +      || __builtin_memcmp ((void *) &x2, &(v4si) { 10, 0, 0, 13 },
> +                          sizeof (v4si))
> +      || __builtin_memcmp ((void *) &x3, &(v4si) { 0, 0, 22, 0 },
> +                          sizeof (v4si)))
> +    __builtin_abort ();
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump-not { \* } "ccp1" } } */
> Index: gcc/testsuite/gcc.dg/pr88598-2.c
> ===================================================================
> --- /dev/null   2018-12-31 11:20:29.178325188 +0000
> +++ gcc/testsuite/gcc.dg/pr88598-2.c    2019-01-04 12:40:51.986582877 +0000
> @@ -0,0 +1,30 @@
> +/* { dg-do run { target double64 } } */
> +/* { dg-options "-O -fdump-tree-ccp1" } */
> +/* { dg-add-options ieee } */
> +
> +typedef double v4df __attribute__ ((vector_size (32)));
> +
> +int
> +main ()
> +{
> +  volatile v4df x1 = { 4, 5, 6, -7 };
> +  volatile v4df x2 = { 10, -11, 12, 13 };
> +  volatile v4df x3 = { 20, 21, 22, 23 };
> +
> +  x1 *= (v4df) { 0, 1, 1, 0 };
> +  x2 *= (v4df) { 1, 0, 0, 1 };
> +  x3 *= (v4df) { 0.0, -0.0, 1.0, -0.0 };
> +
> +  if (__builtin_memcmp ((void *) &x1, &(v4df) { 0, 5, 6, -0.0 },
> +                       sizeof (v4df))
> +      || __builtin_memcmp ((void *) &x2, &(v4df) { 10, -0.0, 0, 13 },
> +                          sizeof (v4df))
> +      || __builtin_memcmp ((void *) &x3, &(v4df) { 0, -0.0, 22, -0.0 },
> +                          sizeof (v4df)))
> +    __builtin_abort ();
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump { \* } "ccp1" } } */
> +/* { dg-final { scan-tree-dump-not { \& } "ccp1" } } */
> Index: gcc/testsuite/gcc.dg/pr88598-3.c
> ===================================================================
> --- /dev/null   2018-12-31 11:20:29.178325188 +0000
> +++ gcc/testsuite/gcc.dg/pr88598-3.c    2019-01-04 12:40:51.986582877 +0000
> @@ -0,0 +1,29 @@
> +/* { dg-do run { target double64 } } */
> +/* { dg-options "-O -fno-signed-zeros -fdump-tree-ccp1" } */
> +/* { dg-add-options ieee } */
> +
> +typedef double v4df __attribute__ ((vector_size (32)));
> +
> +int
> +main ()
> +{
> +  volatile v4df x1 = { 4, 5, 6, -7 };
> +  volatile v4df x2 = { 10, -11, 12, 13 };
> +  volatile v4df x3 = { 20, 21, 22, 23 };
> +
> +  x1 *= (v4df) { 0, 1, 1, 0 };
> +  x2 *= (v4df) { 1, 0, 0, 1 };
> +  x3 *= (v4df) { 0.0, -0.0, 1.0, -0.0 };
> +
> +  if (__builtin_memcmp ((void *) &x1, &(v4df) { 0, 5, 6, 0 },
> +                       sizeof (v4df))
> +      || __builtin_memcmp ((void *) &x2, &(v4df) { 10, 0, 0, 13 },
> +                          sizeof (v4df))
> +      || __builtin_memcmp ((void *) &x3, &(v4df) { 0, 0, 22, 0 },
> +                          sizeof (v4df)))
> +    __builtin_abort ();
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump-not { \* } "ccp1" } } */
> Index: gcc/testsuite/gcc.dg/pr88598-4.c
> ===================================================================
> --- /dev/null   2018-12-31 11:20:29.178325188 +0000
> +++ gcc/testsuite/gcc.dg/pr88598-4.c    2019-01-04 12:40:51.986582877 +0000
> @@ -0,0 +1,28 @@
> +/* { dg-do run } */
> +/* { dg-options "-O -fdump-tree-ccp1" } */
> +
> +typedef int v4si __attribute__ ((vector_size (16)));
> +
> +int
> +main ()
> +{
> +  volatile v4si x1 = { 4, 5, 6, 7 };
> +  volatile v4si x2 = { 10, 11, 12, 13 };
> +  volatile v4si x3 = { 20, 21, 22, 23 };
> +
> +  x1 *= (v4si) { 0, 1, 2, 3 };
> +  x2 *= (v4si) { 1, 0, 2, 0 };
> +  x3 *= (v4si) { 0, 0, -1, 0 };
> +
> +  if (__builtin_memcmp ((void *) &x1, &(v4si) { 0, 5, 12, 21 }, sizeof (v4si))
> +      || __builtin_memcmp ((void *) &x2, &(v4si) { 10, 0, 24, 0 },
> +                          sizeof (v4si))
> +      || __builtin_memcmp ((void *) &x3, &(v4si) { 0, 0, -22, 0 },
> +                          sizeof (v4si)))
> +    __builtin_abort ();
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump { \* } "ccp1" } } */
> +/* { dg-final { scan-tree-dump-not { \& } "ccp1" } } */
> Index: gcc/testsuite/gcc.dg/pr88598-5.c
> ===================================================================
> --- /dev/null   2018-12-31 11:20:29.178325188 +0000
> +++ gcc/testsuite/gcc.dg/pr88598-5.c    2019-01-04 12:40:51.986582877 +0000
> @@ -0,0 +1,29 @@
> +/* { dg-do run { target double64 } } */
> +/* { dg-options "-O -fno-signed-zeros -fdump-tree-ccp1" } */
> +/* { dg-add-options ieee } */
> +
> +typedef double v4df __attribute__ ((vector_size (32)));
> +
> +int
> +main ()
> +{
> +  volatile v4df x1 = { 4, 5, 6, 7 };
> +  volatile v4df x2 = { 10, 11, 12, 13 };
> +  volatile v4df x3 = { 20, 21, 22, 23 };
> +
> +  x1 *= (v4df) { 0, 1, 2, 3 };
> +  x2 *= (v4df) { 1, 0, 2, 0 };
> +  x3 *= (v4df) { 0, 0, -1, 0 };
> +
> +  if (__builtin_memcmp ((void *) &x1, &(v4df) { 0, 5, 12, 21 }, sizeof (v4df))
> +      || __builtin_memcmp ((void *) &x2, &(v4df) { 10, 0, 24, 0 },
> +                          sizeof (v4df))
> +      || __builtin_memcmp ((void *) &x3, &(v4df) { 0, 0, -22, 0 },
> +                          sizeof (v4df)))
> +    __builtin_abort ();
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump { \* } "ccp1" } } */
> +/* { dg-final { scan-tree-dump-not { \& } "ccp1" } } */

  reply	other threads:[~2019-01-07 11:35 UTC|newest]

Thread overview: 7+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-01-04 11:43 Richard Sandiford
2019-01-04 12:04 ` Eric Botcazou
2019-01-04 12:13   ` Richard Sandiford
2019-01-04 12:19     ` Jakub Jelinek
2019-01-04 12:44       ` Richard Sandiford
2019-01-07 11:35         ` Richard Biener [this message]
2019-01-04 12:25     ` Eric Botcazou

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to='CAFiYyc3zZqPiZi-0=asiczTc+v=hBq=A13mdv_0=AbVv=WmPdA@mail.gmail.com' \
    --to=richard.guenther@gmail.com \
    --cc=ebotcazou@adacore.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=jakub@redhat.com \
    --cc=richard.sandiford@arm.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).