LoongArch: Optimize f{max,min}imum{,f}

The code now looks like:

	fclass.s        $fa2, $fa0
	movfr2gr.s      $t0, $fa2
	slli.w          $t0, $t0, 0x0
	fclass.s        $fa2, $fa1
	movfr2gr.s      $t1, $fa2
	or              $t0, $t0, $t1
	andi            $t0, $t0, 0x3
	bnez            $t0, 1f
	fmin.s          $fa0, $fa0, $fa1
	ret
	1:
	fmul.s		$fa0, $fa0, $fa1
	ret

This looks really bad, with expensive movfr2gr instructions, redundant
sign-extensions and masking (arguably it's a compiler
missed-optimzation), and a branch.  Rewrite it with inline assembly:

	fcmp.cor.s      $fcc0, $fa0, $fa0
	fcmp.cor.s      $fcc1, $fa1, $fa1
	fsel            $fa2, $fa0, $fa1, $fcc0
	fsel            $fa0, $fa1, $fa0, $fcc1
	fmax.s          $fa0, $fa2, $fa0
	ret

Note that we cannot make it more readable with
"double a = __builtin_isnanf (x) ? y : x" because this C statement only
happens to produce what we want with https://gcc.gnu.org/PR66462, if
this bug is fixed in the future the generated code may change.

Signed-off-by: Xi Ruoyao <xry111@xry111.site>

diff --git a/sysdeps/loongarch/fpu/s_fmaximum.c b/sysdeps/loongarch/fpu/math-type-macros-double.h
similarity index 54%
rename from sysdeps/loongarch/fpu/s_fmaximum.c
rename to sysdeps/loongarch/fpu/math-type-macros-double.h
index 07fc7214a1af234b9305892fd65f377771bc8686..dd53885144b9eeec22649bdd3c1911eb9b18636b 100644 (file)
--- a/sysdeps/loongarch/fpu/s_fmaximum.c
+++ b/sysdeps/loongarch/fpu/math-type-macros-double.h
@@ -1,5 +1,6 @@
-/* fmaximum().  LoongArch version.
-   Copyright (C) 2022-2025 Free Software Foundation, Inc.
+/* Helper macros for double variants of type generic functions of libm,
+   LoongArch version.
+   Copyright (C) 2025 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
@@ -16,25 +17,5 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
-#define NO_MATH_REDIRECT
-#include <math.h>
-#include <libm-alias-double.h>
-#include <fpu_control.h>
-
-double
-__fmaximum (double x, double y)
-{
-  int x_cond;
-  int y_cond;
-  asm volatile ("fclass.d \t%0, %1" : "=f" (x_cond) : "f" (x));
-  asm volatile ("fclass.d \t%0, %1" : "=f" (y_cond) : "f" (y));
-
-  if (__glibc_unlikely((x_cond | y_cond) & _FCLASS_NAN))
-      return x * y;
-  else
-    {
-      asm volatile ("fmax.d \t%0, %1, %2" : "=f" (x) : "f" (x), "f" (y));
-      return x;
-    }
-}
-libm_alias_double (__fmaximum, fmaximum)
+#include_next <math-type-macros-double.h>
+#define INSN_FMT "d"

diff --git a/sysdeps/loongarch/fpu/s_fminimum.c b/sysdeps/loongarch/fpu/math-type-macros-float.h
similarity index 54%
rename from sysdeps/loongarch/fpu/s_fminimum.c
rename to sysdeps/loongarch/fpu/math-type-macros-float.h
index a63c357957be3052f631743910951d40e37879e2..ee5f7847c8240ba0cd9169ab411d174119bd47dd 100644 (file)
--- a/sysdeps/loongarch/fpu/s_fminimum.c
+++ b/sysdeps/loongarch/fpu/math-type-macros-float.h
@@ -1,5 +1,6 @@
-/* fminimum().  LoongArch version.
-   Copyright (C) 2022-2025 Free Software Foundation, Inc.
+/* Helper macros for float variants of type generic functions of libm,
+   LoongArch version.
+   Copyright (C) 2025 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
@@ -16,25 +17,5 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
-#define NO_MATH_REDIRECT
-#include <math.h>
-#include <libm-alias-double.h>
-#include <fpu_control.h>
-
-double
-__fminimum (double x, double y)
-{
-  int x_cond;
-  int y_cond;
-  asm volatile ("fclass.d \t%0, %1" : "=f" (x_cond) : "f" (x));
-  asm volatile ("fclass.d \t%0, %1" : "=f" (y_cond) : "f" (y));
-
-  if (__glibc_unlikely((x_cond | y_cond) & _FCLASS_NAN))
-      return x * y;
-  else
-    {
-      asm volatile ("fmin.d \t%0, %1, %2" : "=f" (x) : "f" (x), "f" (y));
-      return x;
-    }
-}
-libm_alias_double (__fminimum, fminimum)
+#include_next <math-type-macros-float.h>
+#define INSN_FMT "s"

diff --git a/sysdeps/loongarch/fpu/s_fmaximumf.c b/sysdeps/loongarch/fpu/s_fmaximum_template.c
similarity index 56%
rename from sysdeps/loongarch/fpu/s_fmaximumf.c
rename to sysdeps/loongarch/fpu/s_fmaximum_template.c
index a518ccf3481d9185dfb9cd108cdaa1401f947ae3..701850ea73d9f3ad3d02490a38fb67f33291158d 100644 (file)
--- a/sysdeps/loongarch/fpu/s_fmaximumf.c
+++ b/sysdeps/loongarch/fpu/s_fmaximum_template.c
@@ -1,5 +1,5 @@
-/* fmaximumf().  LoongArch version.
-   Copyright (C) 2022-2025 Free Software Foundation, Inc.
+/* Return maximum of X and Y.  LoongArch version.
+   Copyright (C) 2025 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
@@ -16,25 +16,24 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
-#define NO_MATH_REDIRECT
+#ifndef INSN_FMT
+#include_next <s_fmaximum_template.c>
+#else
+
 #include <math.h>
-#include <libm-alias-float.h>
-#include <fpu_control.h>
 
-float
-__fmaximumf (float x, float y)
+FLOAT
+M_DECL_FUNC (__fmaximum) (FLOAT x, FLOAT y)
 {
-  int x_cond;
-  int y_cond;
-  asm volatile ("fclass.s \t%0, %1" : "=f" (x_cond) : "f" (x));
-  asm volatile ("fclass.s \t%0, %1" : "=f" (y_cond) : "f" (y));
-
-  if (__glibc_unlikely((x_cond | y_cond) & _FCLASS_NAN))
-      return x * y;
-  else
-    {
-      asm volatile ("fmax.s \t%0, %1, %2" : "=f" (x) : "f" (x), "f" (y));
-      return x;
-    }
+  FLOAT a, b;
+  asm("fcmp.cor." INSN_FMT "\t$fcc0, %2, %2\n\t"
+      "fcmp.cor." INSN_FMT "\t$fcc1, %3, %3\n\t"
+      "fsel"              "\t%0, %2, %3, $fcc0\n\t"
+      "fsel"              "\t%1, %3, %2, $fcc1\n\t"
+      "fmax."    INSN_FMT "\t%1, %0, %1"
+      : "=&f" (a), "=f" (b) : "f" (x), "f" (y) : "fcc0", "fcc1");
+  return b;
 }
-libm_alias_float (__fmaximum, fmaximum)
+declare_mgen_alias (__fmaximum, fmaximum);
+
+#endif

diff --git a/sysdeps/loongarch/fpu/s_fminimumf.c b/sysdeps/loongarch/fpu/s_fminimum_template.c
similarity index 56%
rename from sysdeps/loongarch/fpu/s_fminimumf.c
rename to sysdeps/loongarch/fpu/s_fminimum_template.c
index 973a9f75d16ae381f28a38be271869be3190da77..80db72ee3c0ae14f95621eae4acd2395ddc703a6 100644 (file)
--- a/sysdeps/loongarch/fpu/s_fminimumf.c
+++ b/sysdeps/loongarch/fpu/s_fminimum_template.c
@@ -1,5 +1,5 @@
-/* fminimumf().  LoongArch version.
-   Copyright (C) 2022-2025 Free Software Foundation, Inc.
+/* Return minimum of X and Y.  LoongArch version.
+   Copyright (C) 2025 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
@@ -16,25 +16,24 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
-#define NO_MATH_REDIRECT
+#ifndef INSN_FMT
+#include_next <s_fminimum_template.c>
+#else
+
 #include <math.h>
-#include <libm-alias-float.h>
-#include <fpu_control.h>
 
-float
-__fminimumf (float x, float y)
+FLOAT
+M_DECL_FUNC (__fminimum) (FLOAT x, FLOAT y)
 {
-  int x_cond;
-  int y_cond;
-  asm volatile ("fclass.s \t%0, %1" : "=f" (x_cond) : "f" (x));
-  asm volatile ("fclass.s \t%0, %1" : "=f" (y_cond) : "f" (y));
-
-  if (__glibc_unlikely((x_cond | y_cond) & _FCLASS_NAN))
-      return x * y;
-  else
-    {
-      asm volatile ("fmin.s \t%0, %1, %2" : "=f" (x) : "f" (x), "f" (y));
-      return x;
-    }
+  FLOAT a, b;
+  asm("fcmp.cor." INSN_FMT "\t$fcc0, %2, %2\n\t"
+      "fcmp.cor." INSN_FMT "\t$fcc1, %3, %3\n\t"
+      "fsel"              "\t%0, %2, %3, $fcc0\n\t"
+      "fsel"              "\t%1, %3, %2, $fcc1\n\t"
+      "fmin."    INSN_FMT "\t%1, %0, %1"
+      : "=&f" (a), "=f" (b) : "f" (x), "f" (y) : "fcc0", "fcc1");
+  return b;
 }
-libm_alias_float (__fminimum, fminimum)
+declare_mgen_alias (__fminimum, fminimum);
+
+#endif