[PATCH v2 1/4] LoongArch: Optimize f{max,min}imum{,f}
caiyinyu
caiyinyu@loongson.cn
Thu Feb 27 10:59:41 GMT 2025
LGTM
在 2025/2/24 下午8:49, Xi Ruoyao 写道:
> 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>
> ---
> .../loongarch/fpu/math-type-macros-double.h | 21 ++++++++++
> .../loongarch/fpu/math-type-macros-float.h | 21 ++++++++++
> sysdeps/loongarch/fpu/s_fmaximum.c | 40 -------------------
> sysdeps/loongarch/fpu/s_fmaximum_template.c | 39 ++++++++++++++++++
> sysdeps/loongarch/fpu/s_fmaximumf.c | 40 -------------------
> sysdeps/loongarch/fpu/s_fminimum.c | 40 -------------------
> sysdeps/loongarch/fpu/s_fminimum_template.c | 39 ++++++++++++++++++
> sysdeps/loongarch/fpu/s_fminimumf.c | 40 -------------------
> 8 files changed, 120 insertions(+), 160 deletions(-)
> create mode 100644 sysdeps/loongarch/fpu/math-type-macros-double.h
> create mode 100644 sysdeps/loongarch/fpu/math-type-macros-float.h
> delete mode 100644 sysdeps/loongarch/fpu/s_fmaximum.c
> create mode 100644 sysdeps/loongarch/fpu/s_fmaximum_template.c
> delete mode 100644 sysdeps/loongarch/fpu/s_fmaximumf.c
> delete mode 100644 sysdeps/loongarch/fpu/s_fminimum.c
> create mode 100644 sysdeps/loongarch/fpu/s_fminimum_template.c
> delete mode 100644 sysdeps/loongarch/fpu/s_fminimumf.c
>
> diff --git a/sysdeps/loongarch/fpu/math-type-macros-double.h b/sysdeps/loongarch/fpu/math-type-macros-double.h
> new file mode 100644
> index 0000000000..dd53885144
> --- /dev/null
> +++ b/sysdeps/loongarch/fpu/math-type-macros-double.h
> @@ -0,0 +1,21 @@
> +/* 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
> + 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_next <math-type-macros-double.h>
> +#define INSN_FMT "d"
> diff --git a/sysdeps/loongarch/fpu/math-type-macros-float.h b/sysdeps/loongarch/fpu/math-type-macros-float.h
> new file mode 100644
> index 0000000000..ee5f7847c8
> --- /dev/null
> +++ b/sysdeps/loongarch/fpu/math-type-macros-float.h
> @@ -0,0 +1,21 @@
> +/* 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
> + 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_next <math-type-macros-float.h>
> +#define INSN_FMT "s"
> diff --git a/sysdeps/loongarch/fpu/s_fmaximum.c b/sysdeps/loongarch/fpu/s_fmaximum.c
> deleted file mode 100644
> index 07fc7214a1..0000000000
> --- a/sysdeps/loongarch/fpu/s_fmaximum.c
> +++ /dev/null
> @@ -1,40 +0,0 @@
> -/* fmaximum(). LoongArch version.
> - Copyright (C) 2022-2025 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
> - 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/>. */
> -
> -#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)
> diff --git a/sysdeps/loongarch/fpu/s_fmaximum_template.c b/sysdeps/loongarch/fpu/s_fmaximum_template.c
> new file mode 100644
> index 0000000000..701850ea73
> --- /dev/null
> +++ b/sysdeps/loongarch/fpu/s_fmaximum_template.c
> @@ -0,0 +1,39 @@
> +/* 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
> + 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/>. */
> +
> +#ifndef INSN_FMT
> +#include_next <s_fmaximum_template.c>
> +#else
> +
> +#include <math.h>
> +
> +FLOAT
> +M_DECL_FUNC (__fmaximum) (FLOAT x, FLOAT y)
> +{
> + 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;
> +}
> +declare_mgen_alias (__fmaximum, fmaximum);
> +
> +#endif
> diff --git a/sysdeps/loongarch/fpu/s_fmaximumf.c b/sysdeps/loongarch/fpu/s_fmaximumf.c
> deleted file mode 100644
> index a518ccf348..0000000000
> --- a/sysdeps/loongarch/fpu/s_fmaximumf.c
> +++ /dev/null
> @@ -1,40 +0,0 @@
> -/* fmaximumf(). LoongArch version.
> - Copyright (C) 2022-2025 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
> - 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/>. */
> -
> -#define NO_MATH_REDIRECT
> -#include <math.h>
> -#include <libm-alias-float.h>
> -#include <fpu_control.h>
> -
> -float
> -__fmaximumf (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;
> - }
> -}
> -libm_alias_float (__fmaximum, fmaximum)
> diff --git a/sysdeps/loongarch/fpu/s_fminimum.c b/sysdeps/loongarch/fpu/s_fminimum.c
> deleted file mode 100644
> index a63c357957..0000000000
> --- a/sysdeps/loongarch/fpu/s_fminimum.c
> +++ /dev/null
> @@ -1,40 +0,0 @@
> -/* fminimum(). LoongArch version.
> - Copyright (C) 2022-2025 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
> - 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/>. */
> -
> -#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)
> diff --git a/sysdeps/loongarch/fpu/s_fminimum_template.c b/sysdeps/loongarch/fpu/s_fminimum_template.c
> new file mode 100644
> index 0000000000..80db72ee3c
> --- /dev/null
> +++ b/sysdeps/loongarch/fpu/s_fminimum_template.c
> @@ -0,0 +1,39 @@
> +/* 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
> + 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/>. */
> +
> +#ifndef INSN_FMT
> +#include_next <s_fminimum_template.c>
> +#else
> +
> +#include <math.h>
> +
> +FLOAT
> +M_DECL_FUNC (__fminimum) (FLOAT x, FLOAT y)
> +{
> + 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;
> +}
> +declare_mgen_alias (__fminimum, fminimum);
> +
> +#endif
> diff --git a/sysdeps/loongarch/fpu/s_fminimumf.c b/sysdeps/loongarch/fpu/s_fminimumf.c
> deleted file mode 100644
> index 973a9f75d1..0000000000
> --- a/sysdeps/loongarch/fpu/s_fminimumf.c
> +++ /dev/null
> @@ -1,40 +0,0 @@
> -/* fminimumf(). LoongArch version.
> - Copyright (C) 2022-2025 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
> - 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/>. */
> -
> -#define NO_MATH_REDIRECT
> -#include <math.h>
> -#include <libm-alias-float.h>
> -#include <fpu_control.h>
> -
> -float
> -__fminimumf (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;
> - }
> -}
> -libm_alias_float (__fminimum, fminimum)
More information about the Libc-alpha
mailing list