public inbox for libc-alpha@sourceware.org
 help / color / mirror / Atom feed
From: YunQiang Su <syq@gcc.gnu.org>
To: libc-alpha@sourceware.org
Subject: [PATCH 4/6] MIPS/math: Implement optimized fmaximum/fminmum(,_mag)(,f)
Date: Mon, 13 May 2024 16:14:27 +0800	[thread overview]
Message-ID: <20240513081429.1749898-5-syq@gcc.gnu.org> (raw)
In-Reply-To: <20240513081429.1749898-1-syq@gcc.gnu.org>

MIPSr6 defines max/maxa/min/mina instructions, which have slight
different with fmaximum/fminimum: return the number instead of NAN if
NUM vs NAN.  Detecting NAN is required before these instructions.

Another problem of the generic implemention for MIPS is that, it
use copysign for +0/-0 problem.  In fact we can use GET_HIGH_WORD or
GET_FLOAT_WORD, and then determine by the sign bit:

  int32_t xi;
  GET_HIGH_WORD (xi, x);
  return (xi < 0 ? y : x);

GET_HIGH_WORD/GET_FLOAT_WORD are much more friendly to MIPS FPU.
We can use `mfhc1` to GET_HIGH_WORD, and `mfc1` to GET_FLOAT_WORD.

Since the abs.fmt instructions will signal if one operand is qNAN or
sNaN, M_FABS, aka __builtin_fabs will issue at least 4 instructions
(mfc1/ext/ins/mtc1).  So detecting NaN is also required.

	* sysdeps/mips/ieee754/s_fmaximum.c
	* sysdeps/mips/ieee754/s_fmaximum_mag.c
	* sysdeps/mips/ieee754/s_fmaximum_magf.c
	* sysdeps/mips/ieee754/s_fmaximumf.c
	* sysdeps/mips/ieee754/s_fminimum.c
	* sysdeps/mips/ieee754/s_fminimum_mag.c
	* sysdeps/mips/ieee754/s_fminimum_magf.c
	* sysdeps/mips/ieee754/s_fminimumf.c

Signed-off-by: YunQiang Su <syq@gcc.gnu.org>
---
 sysdeps/mips/ieee754/s_fmaximum.c      | 48 ++++++++++++++++++++++
 sysdeps/mips/ieee754/s_fmaximum_mag.c  | 57 ++++++++++++++++++++++++++
 sysdeps/mips/ieee754/s_fmaximum_magf.c | 55 +++++++++++++++++++++++++
 sysdeps/mips/ieee754/s_fmaximumf.c     | 46 +++++++++++++++++++++
 sysdeps/mips/ieee754/s_fminimum.c      | 48 ++++++++++++++++++++++
 sysdeps/mips/ieee754/s_fminimum_mag.c  | 57 ++++++++++++++++++++++++++
 sysdeps/mips/ieee754/s_fminimum_magf.c | 55 +++++++++++++++++++++++++
 sysdeps/mips/ieee754/s_fminimumf.c     | 46 +++++++++++++++++++++
 8 files changed, 412 insertions(+)
 create mode 100644 sysdeps/mips/ieee754/s_fmaximum.c
 create mode 100644 sysdeps/mips/ieee754/s_fmaximum_mag.c
 create mode 100644 sysdeps/mips/ieee754/s_fmaximum_magf.c
 create mode 100644 sysdeps/mips/ieee754/s_fmaximumf.c
 create mode 100644 sysdeps/mips/ieee754/s_fminimum.c
 create mode 100644 sysdeps/mips/ieee754/s_fminimum_mag.c
 create mode 100644 sysdeps/mips/ieee754/s_fminimum_magf.c
 create mode 100644 sysdeps/mips/ieee754/s_fminimumf.c

diff --git a/sysdeps/mips/ieee754/s_fmaximum.c b/sysdeps/mips/ieee754/s_fmaximum.c
new file mode 100644
index 0000000000..5a1e6a0313
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fmaximum.c
@@ -0,0 +1,48 @@
+/* fmaximum().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-double.h>
+#include <math_private.h>
+
+double
+__fmaximum (double x, double y)
+{
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)                         \
+    && !defined(__mips_single_float)
+  /* MAX.d returns NUM if NUM vs qNAN.  */
+  if (isunordered (x, y))
+    return x + y;
+  double ret;
+  asm volatile("max.d	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  if (isgreater (x, y))
+    return x;
+  else if (isless (x, y))
+    return y;
+  if (isunordered (x, y))
+    return x + y;
+
+  int32_t xi;
+  GET_HIGH_WORD (xi, x);
+  return (xi < 0 ? y : x);
+#endif
+}
+
+libm_alias_double (__fmaximum, fmaximum)
diff --git a/sysdeps/mips/ieee754/s_fmaximum_mag.c b/sysdeps/mips/ieee754/s_fmaximum_mag.c
new file mode 100644
index 0000000000..0eac275167
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fmaximum_mag.c
@@ -0,0 +1,57 @@
+/* fmaximum_mag().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-double.h>
+#include <math_private.h>
+#include <math-type-macros-double.h>
+
+double
+__fmaximum_mag (double x, double y)
+{
+  /* MAXA.d return NUM if NUM vs qNAN.  ABS.d signals both sNAN and qNAN on
+     pre-R5.  */
+  if (isunordered (x, y))
+    return x + y;
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)                         \
+    && !defined(__mips_single_float)
+  double ret;
+  asm volatile("maxa.d	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  double ax;
+  double ay;
+#  if defined(__mips_hard_float) && !defined(__mips_single_float)
+  asm volatile("abs.d	%0, %1" : "=f"(ax) : "f"(x));
+  asm volatile("abs.d	%0, %1" : "=f"(ay) : "f"(y));
+#  else
+  ax = M_FABS (x);
+  ay = M_FABS (y);
+#  endif
+  if (isgreater (ax, ay))
+    return x;
+  else if (isless (ax, ay))
+    return y;
+
+  int32_t xi;
+  GET_HIGH_WORD (xi, x);
+  return (xi < 0 ? y : x);
+#endif
+}
+
+libm_alias_double (__fmaximum_mag, fmaximum_mag)
diff --git a/sysdeps/mips/ieee754/s_fmaximum_magf.c b/sysdeps/mips/ieee754/s_fmaximum_magf.c
new file mode 100644
index 0000000000..dd871bac07
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fmaximum_magf.c
@@ -0,0 +1,55 @@
+/* fmaximum_magf().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-float.h>
+#include <math-type-macros-float.h>
+
+float
+__fmaximum_magf (float x, float y)
+{
+  /* MAXA.s return NUM if NUM vs qNAN.  ABS.s signals both sNAN and qNAN on
+     pre-R5.  */
+  if (isunordered (x, y))
+    return x + y;
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)
+  float ret;
+  asm volatile("maxa.s	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  float ax;
+  float ay;
+#  if defined(__mips_hard_float)
+  asm volatile("abs.s	%0, %1" : "=f"(ax) : "f"(x));
+  asm volatile("abs.s	%0, %1" : "=f"(ay) : "f"(y));
+#  else
+  ax = M_FABS (x);
+  ay = M_FABS (y);
+#  endif
+  if (isgreater (ax, ay))
+    return x;
+  else if (isless (ax, ay))
+    return y;
+
+  int32_t xi;
+  GET_FLOAT_WORD (xi, x);
+  return (xi < 0 ? y : x);
+#endif
+}
+
+libm_alias_float (__fmaximum_mag, fmaximum_mag)
diff --git a/sysdeps/mips/ieee754/s_fmaximumf.c b/sysdeps/mips/ieee754/s_fmaximumf.c
new file mode 100644
index 0000000000..a266ee76b6
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fmaximumf.c
@@ -0,0 +1,46 @@
+/* fmaximumf().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-float.h>
+
+float
+__fmaximumf (float x, float y)
+{
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)
+  /* MAX.s returns NUM if NUM vs qNAN.  */
+  if (isunordered (x, y))
+    return x + y;
+  float ret;
+  asm volatile("max.s	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  if (isgreater (x, y))
+    return x;
+  else if (isless (x, y))
+    return y;
+  if (isunordered (x, y))
+    return x + y;
+
+  int32_t xi;
+  GET_FLOAT_WORD (xi, x);
+  return (xi < 0 ? y : x);
+#endif
+}
+
+libm_alias_float (__fmaximum, fmaximum)
diff --git a/sysdeps/mips/ieee754/s_fminimum.c b/sysdeps/mips/ieee754/s_fminimum.c
new file mode 100644
index 0000000000..083da390ae
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fminimum.c
@@ -0,0 +1,48 @@
+/* fminimum().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-double.h>
+#include <math_private.h>
+
+double
+__fminimum (double x, double y)
+{
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)                         \
+    && !defined(__mips_single_float)
+  /* MIN.d returns NUM if NUM vs qNAN.  */
+  if (isunordered (x, y))
+    return x + y;
+  double ret;
+  asm volatile("min.d	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  if (isgreater (x, y))
+    return y;
+  else if (isless (x, y))
+    return x;
+  if (isunordered (x, y))
+    return x + y;
+
+  int32_t xi;
+  GET_HIGH_WORD (xi, x);
+  return (xi < 0 ? x : y);
+#endif
+}
+
+libm_alias_double (__fminimum, fminimum)
diff --git a/sysdeps/mips/ieee754/s_fminimum_mag.c b/sysdeps/mips/ieee754/s_fminimum_mag.c
new file mode 100644
index 0000000000..7adaa1c279
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fminimum_mag.c
@@ -0,0 +1,57 @@
+/* fminimum_mag().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-double.h>
+#include <math_private.h>
+#include <math-type-macros-double.h>
+
+double
+__fminimum_mag (double x, double y)
+{
+  /* MINA.d return NUM if NUM vs qNAN.  ABS.d signals both sNAN and qNAN on
+     pre-R5.  */
+  if (isunordered (x, y))
+    return x + y;
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)                         \
+    && !defined(__mips_single_float)
+  double ret;
+  asm volatile("mina.d	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  double ax;
+  double ay;
+#  if defined(__mips_hard_float) && !defined(__mips_single_float)
+  asm volatile("abs.d	%0, %1" : "=f"(ax) : "f"(x));
+  asm volatile("abs.d	%0, %1" : "=f"(ay) : "f"(y));
+#  else
+  ax = M_FABS (x);
+  ay = M_FABS (y);
+#  endif
+  if (isgreater (ax, ay))
+    return y;
+  else if (isless (ax, ay))
+    return x;
+
+  int32_t xi;
+  GET_HIGH_WORD (xi, x);
+  return (xi < 0 ? x : y);
+#endif
+}
+
+libm_alias_double (__fminimum_mag, fminimum_mag)
diff --git a/sysdeps/mips/ieee754/s_fminimum_magf.c b/sysdeps/mips/ieee754/s_fminimum_magf.c
new file mode 100644
index 0000000000..6839e2914d
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fminimum_magf.c
@@ -0,0 +1,55 @@
+/* fminimum_magf().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-float.h>
+#include <math-type-macros-float.h>
+
+float
+__fminimum_magf (float x, float y)
+{
+  /* MAXA.s return NUM if NUM vs qNAN.  ABS.s signals both sNAN and qNAN on
+     pre-R5.  */
+  if (isunordered (x, y))
+    return x + y;
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)
+  float ret;
+  asm volatile("mina.s	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  float ax;
+  float ay;
+#  if defined(__mips_hard_float)
+  asm volatile("abs.s	%0, %1" : "=f"(ax) : "f"(x));
+  asm volatile("abs.s	%0, %1" : "=f"(ay) : "f"(y));
+#  else
+  ax = M_FABS (x);
+  ay = M_FABS (y);
+#  endif
+  if (isgreater (ax, ay))
+    return y;
+  else if (isless (ax, ay))
+    return x;
+
+  int32_t xi;
+  GET_FLOAT_WORD (xi, x);
+  return (xi < 0 ? x : y);
+#endif
+}
+
+libm_alias_float (__fminimum_mag, fminimum_mag)
diff --git a/sysdeps/mips/ieee754/s_fminimumf.c b/sysdeps/mips/ieee754/s_fminimumf.c
new file mode 100644
index 0000000000..f37ca1c23b
--- /dev/null
+++ b/sysdeps/mips/ieee754/s_fminimumf.c
@@ -0,0 +1,46 @@
+/* fminimumf().  MIPS version.
+   Copyright (C) 2024 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <libm-alias-float.h>
+
+float
+__fminimumf (float x, float y)
+{
+#if __mips_isa_rev >= 6 && defined(__mips_hard_float)
+  /* MIN.s returns NUM if NUM vs qNAN.  */
+  if (isunordered (x, y))
+    return x + y;
+  float ret;
+  asm volatile("min.s	%0, %1, %2" : "=f"(ret) : "f"(x), "f"(y));
+  return ret;
+#else
+  if (isgreater (x, y))
+    return y;
+  else if (isless (x, y))
+    return x;
+  if (isunordered (x, y))
+    return x + y;
+
+  int xi;
+  GET_FLOAT_WORD (xi, x);
+  return (xi < 0 ? x : y);
+#endif
+}
+
+libm_alias_float (__fminimum, fminimum)
-- 
2.39.2


  parent reply	other threads:[~2024-05-13  8:14 UTC|newest]

Thread overview: 13+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2024-05-13  8:14 [PATCH 0/6] MIPS: Improve math YunQiang Su
2024-05-13  8:14 ` [PATCH 1/6] MIPSr6/math: Use builtin fma and fmaf YunQiang Su
2024-05-22 13:42   ` Adhemerval Zanella Netto
2024-05-25 11:55     ` YunQiang Su
2024-05-31 15:53       ` YunQiang Su
2024-05-13  8:14 ` [PATCH 2/6] MIPS/math: Define port-specific GET_HIGH_WORD YunQiang Su
2024-05-22 18:36   ` Adhemerval Zanella Netto
2024-05-13  8:14 ` [PATCH 3/6] MIPS/math: Implement optimized issignaling(f) YunQiang Su
2024-05-23 14:37   ` Adhemerval Zanella Netto
2024-05-13  8:14 ` YunQiang Su [this message]
2024-05-13  8:14 ` [PATCH 5/6] MIPS/math: Implement optimized fmax(mag)(f)/fmin(mag)(f) YunQiang Su
2024-05-13  8:14 ` [PATCH 6/6] MIPS/math: Implement optimized f(max,min)imum(_mag)_num(f) YunQiang Su
2024-05-22 13:21 ` [PATCH 0/6] MIPS: Improve math YunQiang Su

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=20240513081429.1749898-5-syq@gcc.gnu.org \
    --to=syq@gcc.gnu.org \
    --cc=libc-alpha@sourceware.org \
    /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).