public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Richard Biener <richard.guenther@gmail.com>
To: Richard Biener <richard.guenther@gmail.com>,
	GCC Patches <gcc-patches@gcc.gnu.org>,
	 Richard Sandiford <richard.sandiford@arm.com>
Subject: Re: [PATCH 1/4] Canonicalize argument order for commutative functions
Date: Tue, 30 Nov 2021 08:40:04 +0100	[thread overview]
Message-ID: <CAFiYyc0C7z0LxKN3eZ1gzyP6Dj4x-igTrcX4LyEurV31RuQrAg@mail.gmail.com> (raw)
In-Reply-To: <mpt8rx73r8o.fsf@arm.com>

On Mon, Nov 29, 2021 at 4:40 PM Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> Sorry for the slow response, was away last week.
>
> Richard Biener <richard.guenther@gmail.com> writes:
> > On Wed, Nov 10, 2021 at 1:50 PM Richard Sandiford via Gcc-patches
> > <gcc-patches@gcc.gnu.org> wrote:
> >>
> >> This patch uses information about internal functions to canonicalize
> >> the argument order of calls.
> >>
> >> Tested on aarch64-linux-gnu and x86_64-linux-gnu.  OK to install?
> >
> > OK.  Note the gimple_resimplifyN functions also canonicalize operand
> > order, currently for is_tree_code only:
> >
> >   /* Canonicalize operand order.  */
> >   bool canonicalized = false;
> >   if (res_op->code.is_tree_code ()
> >       && (TREE_CODE_CLASS ((enum tree_code) res_op->code) == tcc_comparison
> >           || commutative_tree_code (res_op->code))
> >       && tree_swap_operands_p (res_op->ops[0], res_op->ops[1]))
> >     {
> >       std::swap (res_op->ops[0], res_op->ops[1]);
> >       if (TREE_CODE_CLASS ((enum tree_code) res_op->code) == tcc_comparison)
> >         res_op->code = swap_tree_comparison (res_op->code);
> >       canonicalized = true;
> >     }
> >
> > that's maybe not the best place.  The function assumes the operands
> > are already valueized,
> > so it maybe should be valueization that does the canonicalization -
> > but I think doing it
> > elsewhere made operand order unreliable (we do end up with
> > non-canonical order in
> > the IL sometimes).
> >
> > So maybe you should amend the code in resimplifyN as well.
>
> Hmm, yeah, thanks for the heads up.  Does this updated version look OK?
> Tested as before.

Yes - OK.

Thanks,
Richard.

> Thanks,
> Richard
>
>
> gcc/
>         * gimple-fold.c: Include internal-fn.h.
>         (fold_stmt_1): If a function maps to an internal one, use
>         first_commutative_argument to canonicalize the order of
>         commutative arguments.
>         * gimple-match-head.c (gimple_resimplify2, gimple_resimplify3)
>         (gimple_resimplify4, gimple_resimplify5): Extend commutativity
>         checks to functions.
>
> gcc/testsuite/
>         * gcc.dg/fmax-fmin-1.c: New test.
> ---
>  gcc/gimple-fold.c                  | 25 ++++++++++++--
>  gcc/gimple-match-head.c            | 52 ++++++++++++++++++++----------
>  gcc/testsuite/gcc.dg/fmax-fmin-1.c | 18 +++++++++++
>  3 files changed, 75 insertions(+), 20 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.dg/fmax-fmin-1.c
>
> diff --git a/gcc/gimple-fold.c b/gcc/gimple-fold.c
> index 44fba12e150..1d8fd74f72c 100644
> --- a/gcc/gimple-fold.c
> +++ b/gcc/gimple-fold.c
> @@ -69,6 +69,7 @@ along with GCC; see the file COPYING3.  If not see
>  #include "varasm.h"
>  #include "memmodel.h"
>  #include "optabs.h"
> +#include "internal-fn.h"
>
>  enum strlen_range_kind {
>    /* Compute the exact constant string length.  */
> @@ -6109,18 +6110,36 @@ fold_stmt_1 (gimple_stmt_iterator *gsi, bool inplace, tree (*valueize) (tree))
>        break;
>      case GIMPLE_CALL:
>        {
> -       for (i = 0; i < gimple_call_num_args (stmt); ++i)
> +       gcall *call = as_a<gcall *> (stmt);
> +       for (i = 0; i < gimple_call_num_args (call); ++i)
>           {
> -           tree *arg = gimple_call_arg_ptr (stmt, i);
> +           tree *arg = gimple_call_arg_ptr (call, i);
>             if (REFERENCE_CLASS_P (*arg)
>                 && maybe_canonicalize_mem_ref_addr (arg))
>               changed = true;
>           }
> -       tree *lhs = gimple_call_lhs_ptr (stmt);
> +       tree *lhs = gimple_call_lhs_ptr (call);
>         if (*lhs
>             && REFERENCE_CLASS_P (*lhs)
>             && maybe_canonicalize_mem_ref_addr (lhs))
>           changed = true;
> +       if (*lhs)
> +         {
> +           combined_fn cfn = gimple_call_combined_fn (call);
> +           internal_fn ifn = associated_internal_fn (cfn, TREE_TYPE (*lhs));
> +           int opno = first_commutative_argument (ifn);
> +           if (opno >= 0)
> +             {
> +               tree arg1 = gimple_call_arg (call, opno);
> +               tree arg2 = gimple_call_arg (call, opno + 1);
> +               if (tree_swap_operands_p (arg1, arg2))
> +                 {
> +                   gimple_call_set_arg (call, opno, arg2);
> +                   gimple_call_set_arg (call, opno + 1, arg1);
> +                   changed = true;
> +                 }
> +             }
> +         }
>         break;
>        }
>      case GIMPLE_ASM:
> diff --git a/gcc/gimple-match-head.c b/gcc/gimple-match-head.c
> index c481a625581..2d9364ca5de 100644
> --- a/gcc/gimple-match-head.c
> +++ b/gcc/gimple-match-head.c
> @@ -294,18 +294,16 @@ gimple_resimplify2 (gimple_seq *seq, gimple_match_op *res_op,
>
>    /* Canonicalize operand order.  */
>    bool canonicalized = false;
> -  if (res_op->code.is_tree_code ())
> +  bool is_comparison
> +    = (res_op->code.is_tree_code ()
> +       && TREE_CODE_CLASS (tree_code (res_op->code)) == tcc_comparison);
> +  if ((is_comparison || commutative_binary_op_p (res_op->code, res_op->type))
> +      && tree_swap_operands_p (res_op->ops[0], res_op->ops[1]))
>      {
> -      auto code = tree_code (res_op->code);
> -      if ((TREE_CODE_CLASS (code) == tcc_comparison
> -          || commutative_tree_code (code))
> -         && tree_swap_operands_p (res_op->ops[0], res_op->ops[1]))
> -       {
> -         std::swap (res_op->ops[0], res_op->ops[1]);
> -         if (TREE_CODE_CLASS (code) == tcc_comparison)
> -           res_op->code = swap_tree_comparison (code);
> -         canonicalized = true;
> -       }
> +      std::swap (res_op->ops[0], res_op->ops[1]);
> +      if (is_comparison)
> +       res_op->code = swap_tree_comparison (tree_code (res_op->code));
> +      canonicalized = true;
>      }
>
>    /* Limit recursion, see gimple_resimplify1.  */
> @@ -376,11 +374,11 @@ gimple_resimplify3 (gimple_seq *seq, gimple_match_op *res_op,
>
>    /* Canonicalize operand order.  */
>    bool canonicalized = false;
> -  if (res_op->code.is_tree_code ()
> -      && commutative_ternary_tree_code (tree_code (res_op->code))
> -      && tree_swap_operands_p (res_op->ops[0], res_op->ops[1]))
> +  int argno = first_commutative_argument (res_op->code, res_op->type);
> +  if (argno >= 0
> +      && tree_swap_operands_p (res_op->ops[argno], res_op->ops[argno + 1]))
>      {
> -      std::swap (res_op->ops[0], res_op->ops[1]);
> +      std::swap (res_op->ops[argno], res_op->ops[argno + 1]);
>        canonicalized = true;
>      }
>
> @@ -424,6 +422,16 @@ gimple_resimplify4 (gimple_seq *seq, gimple_match_op *res_op,
>  {
>    /* No constant folding is defined for four-operand functions.  */
>
> +  /* Canonicalize operand order.  */
> +  bool canonicalized = false;
> +  int argno = first_commutative_argument (res_op->code, res_op->type);
> +  if (argno >= 0
> +      && tree_swap_operands_p (res_op->ops[argno], res_op->ops[argno + 1]))
> +    {
> +      std::swap (res_op->ops[argno], res_op->ops[argno + 1]);
> +      canonicalized = true;
> +    }
> +
>    /* Limit recursion, see gimple_resimplify1.  */
>    static unsigned depth;
>    if (depth > 10)
> @@ -450,7 +458,7 @@ gimple_resimplify4 (gimple_seq *seq, gimple_match_op *res_op,
>    if (maybe_resimplify_conditional_op (seq, res_op, valueize))
>      return true;
>
> -  return false;
> +  return canonicalized;
>  }
>
>  /* Helper that matches and simplifies the toplevel result from
> @@ -465,6 +473,16 @@ gimple_resimplify5 (gimple_seq *seq, gimple_match_op *res_op,
>  {
>    /* No constant folding is defined for five-operand functions.  */
>
> +  /* Canonicalize operand order.  */
> +  bool canonicalized = false;
> +  int argno = first_commutative_argument (res_op->code, res_op->type);
> +  if (argno >= 0
> +      && tree_swap_operands_p (res_op->ops[argno], res_op->ops[argno + 1]))
> +    {
> +      std::swap (res_op->ops[argno], res_op->ops[argno + 1]);
> +      canonicalized = true;
> +    }
> +
>    gimple_match_op res_op2 (*res_op);
>    if (gimple_simplify (&res_op2, seq, valueize,
>                        res_op->code, res_op->type,
> @@ -478,7 +496,7 @@ gimple_resimplify5 (gimple_seq *seq, gimple_match_op *res_op,
>    if (maybe_resimplify_conditional_op (seq, res_op, valueize))
>      return true;
>
> -  return false;
> +  return canonicalized;
>  }
>
>  /* Match and simplify the toplevel valueized operation THIS.
> diff --git a/gcc/testsuite/gcc.dg/fmax-fmin-1.c b/gcc/testsuite/gcc.dg/fmax-fmin-1.c
> new file mode 100644
> index 00000000000..e7e0518d8bb
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/fmax-fmin-1.c
> @@ -0,0 +1,18 @@
> +/* { dg-options "-O -fdump-tree-optimized" } */
> +
> +void
> +f1 (double *res, double x, double y)
> +{
> +  res[0] = __builtin_fmax (x, y);
> +  res[1] = __builtin_fmax (y, x);
> +}
> +
> +void
> +f2 (double *res, double x, double y)
> +{
> +  res[0] = __builtin_fmin (x, y);
> +  res[1] = __builtin_fmin (y, x);
> +}
> +
> +/* { dg-final { scan-tree-dump-times {__builtin_fmax} 1 "optimized" } } */
> +/* { dg-final { scan-tree-dump-times {__builtin_fmin} 1 "optimized" } } */
> --
> 2.25.1
>

      reply	other threads:[~2021-11-30  7:40 UTC|newest]

Thread overview: 4+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-11-10 12:48 Richard Sandiford
2021-11-11 10:57 ` Richard Biener
2021-11-29 15:40   ` Richard Sandiford
2021-11-30  7:40     ` Richard Biener [this message]

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=CAFiYyc0C7z0LxKN3eZ1gzyP6Dj4x-igTrcX4LyEurV31RuQrAg@mail.gmail.com \
    --to=richard.guenther@gmail.com \
    --cc=gcc-patches@gcc.gnu.org \
    --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).