From: Xi Ruoyao Date: Mon, 24 Feb 2025 12:49:15 +0000 (+0800) Subject: LoongArch: Optimize f{max,min}imum{,f} X-Git-Tag: glibc-2.42~439 X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=0195552e158e007a456e568018c25280a96c3b99;p=thirdparty%2Fglibc.git 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 --- 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 07fc7214a1..dd53885144 100644 --- 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 . */ -#define NO_MATH_REDIRECT -#include -#include -#include - -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 +#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 a63c357957..ee5f7847c8 100644 --- 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 . */ -#define NO_MATH_REDIRECT -#include -#include -#include - -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 +#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 a518ccf348..701850ea73 100644 --- 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 . */ -#define NO_MATH_REDIRECT +#ifndef INSN_FMT +#include_next +#else + #include -#include -#include -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, $fcc1\n\t" + "fmax." INSN_FMT "\t%1, %0, %1" + : "=&f" (a), "=f" (b) : "f" (x), "f" (y) : "fcc0", "fcc1"); + return b; } -libm_alias_float (__fmaximum, fmaximum) +declare_mgen_alias (__fmaximum, fmaximum); + +#endif diff --git a/sysdeps/loongarch/fpu/s_fminimumf.c b/sysdeps/loongarch/fpu/s_fminimum_template.c similarity index 56% rename from sysdeps/loongarch/fpu/s_fminimumf.c rename to sysdeps/loongarch/fpu/s_fminimum_template.c index 973a9f75d1..80db72ee3c 100644 --- a/sysdeps/loongarch/fpu/s_fminimumf.c +++ b/sysdeps/loongarch/fpu/s_fminimum_template.c @@ -1,5 +1,5 @@ -/* fminimumf(). LoongArch version. - Copyright (C) 2022-2025 Free Software Foundation, Inc. +/* Return minimum of X and Y. LoongArch version. + Copyright (C) 2025 Free Software Foundation, Inc. This file is part of the GNU C Library. The GNU C Library is free software; you can redistribute it and/or @@ -16,25 +16,24 @@ License along with the GNU C Library; if not, see . */ -#define NO_MATH_REDIRECT +#ifndef INSN_FMT +#include_next +#else + #include -#include -#include -float -__fminimumf (float x, float y) +FLOAT +M_DECL_FUNC (__fminimum) (FLOAT x, FLOAT y) { - int x_cond; - int y_cond; - asm volatile ("fclass.s \t%0, %1" : "=f" (x_cond) : "f" (x)); - asm volatile ("fclass.s \t%0, %1" : "=f" (y_cond) : "f" (y)); - - if (__glibc_unlikely((x_cond | y_cond) & _FCLASS_NAN)) - return x * y; - else - { - asm volatile ("fmin.s \t%0, %1, %2" : "=f" (x) : "f" (x), "f" (y)); - return x; - } + FLOAT a, b; + asm("fcmp.cor." INSN_FMT "\t$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