[1/4] Canonicalize argument order for commutative functions

Message ID mptmtmcuqgb.fsf@arm.com
State Committed
Headers
Series [1/4] Canonicalize argument order for commutative functions |

Commit Message

Richard Sandiford Nov. 10, 2021, 12:48 p.m. UTC
  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?

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.

gcc/testsuite/
	* gcc.dg/fmax-fmin-1.c: New test.
---
 gcc/gimple-fold.c                  | 25 ++++++++++++++++++++++---
 gcc/testsuite/gcc.dg/fmax-fmin-1.c | 18 ++++++++++++++++++
 2 files changed, 40 insertions(+), 3 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/fmax-fmin-1.c
  

Comments

Richard Biener Nov. 11, 2021, 10:57 a.m. UTC | #1
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.

Richard.


> 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.
>
> gcc/testsuite/
>         * gcc.dg/fmax-fmin-1.c: New test.
> ---
>  gcc/gimple-fold.c                  | 25 ++++++++++++++++++++++---
>  gcc/testsuite/gcc.dg/fmax-fmin-1.c | 18 ++++++++++++++++++
>  2 files changed, 40 insertions(+), 3 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 a937f130815..6a7d4507c89 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.  */
> @@ -6140,18 +6141,36 @@ 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/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
>
  
Richard Sandiford Nov. 29, 2021, 3:40 p.m. UTC | #2
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.

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" } } */
  
Richard Biener Nov. 30, 2021, 7:40 a.m. UTC | #3
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
>
  

Patch

diff --git a/gcc/gimple-fold.c b/gcc/gimple-fold.c
index a937f130815..6a7d4507c89 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.  */
@@ -6140,18 +6141,36 @@  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/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" } } */