[PATCH 2/2] Revert "x86: Do not use __builtin_fpclassify for _Float64x/long double"
Carlos O'Donell
carlos@redhat.com
Tue Jan 13 15:07:55 GMT 2026
On 1/13/26 9:19 AM, Adhemerval Zanella wrote:
> This reverts commit 50112103993b042f52fb6afb0e4eee24fe4cb9af. It
> breaks libstdc++ and other languages bootstrap.
Earlier comments apply, I'd like to see us continue to fix this for
the clang/gcc scenarios.
Revert matches.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
> ---
> bits/fp-builtin-denormal.h | 28 --------
> math/Makefile | 1 -
> math/math.h | 89 +------------------------
> sysdeps/x86/bits/fp-builtin-denormal.h | 25 -------
> sysdeps/x86/fpu/Makefile | 1 -
> sysdeps/x86/fpu/test-builtin-denormal.c | 53 ---------------
> 6 files changed, 2 insertions(+), 195 deletions(-)
> delete mode 100644 bits/fp-builtin-denormal.h
> delete mode 100644 sysdeps/x86/bits/fp-builtin-denormal.h
> delete mode 100644 sysdeps/x86/fpu/test-builtin-denormal.c
>
> diff --git a/bits/fp-builtin-denormal.h b/bits/fp-builtin-denormal.h
> deleted file mode 100644
> index 457ab0c97a..0000000000
> --- a/bits/fp-builtin-denormal.h
> +++ /dev/null
> @@ -1,28 +0,0 @@
> -/* Denormal number definitions.
> - Copyright (C) 2025-2026 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 _MATH_H
> -# error "Never use <bits/fp-builtin-denormal.h> directly; include <math.h> instead."
> -#endif
> -
> -/* __FP_BUILTIN_FPCLASSIFY_DENORMAL is defined to 1 if compiler supports
> - handling pseudo-denormal numbers with fpclassify builtin. Pseudo-denormal
> - is a non-standard denormalized floating-point number only supported by
> - Intel double extended-precision (long double). By default assume 1 to
> - enable the usage of compiler builtin on math.h. */
> -#define __FP_BUILTIN_FPCLASSIFY_DENORMAL 1
> diff --git a/math/Makefile b/math/Makefile
> index 69f80e143f..b546626f15 100644
> --- a/math/Makefile
> +++ b/math/Makefile
> @@ -28,7 +28,6 @@ headers := \
> bits/floatn-common.h \
> bits/floatn.h \
> bits/flt-eval-method.h \
> - bits/fp-builtin-denormal.h \
> bits/fp-fast.h \
> bits/fp-logb.h \
> bits/iscanonical.h \
> diff --git a/math/math.h b/math/math.h
> index 77dfcdc36b..75830d7696 100644
> --- a/math/math.h
> +++ b/math/math.h
> @@ -1064,86 +1064,6 @@ extern int signgam;
> : FUNC ## l ARGS)
> #endif
>
> -
> -/* Depending on the type of TG_ARG and extra DEFINE to check, either call the
> - BUILTIN with ARGS_B or an appropriately suffixed version of FUNC with
> - arguments (including parentheses) ARGS_B. The function call is used for
> - long double and/or _Float64x is the builtin can not be safely used on all
> - arguments (defined by DEFINE). */
> -
> -#include <bits/fp-builtin-denormal.h>
> -
> -#ifdef __NO_LONG_DOUBLE_MATH
> -# define __MATH_TG_BUILTIN_CLASSIFY(TG_ARG, BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - BUILTIN ARGS_B
> -#elif __HAVE_DISTINCT_FLOAT128
> -# if __HAVE_GENERIC_SELECTION
> -# if __HAVE_FLOATN_NOT_TYPEDEF && __HAVE_FLOAT32
> -# define __MATH_TG_BUILTIN_CLASSIFY_F32(BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - _Float32: BUILTIN ARGS_B,
> -# else
> -# define __MATH_TG_BUILTIN_CLASSIFY_F32(BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE)
> -# endif
> -# define __MATH_TG_BUILTIN_CLASSIFY_LDOUBLE(BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - long double: DEFINE ? BUILTIN ARGS_B : __ ## FUNC ## l ARGS_F,
> -# if __HAVE_FLOATN_NOT_TYPEDEF && __HAVE_FLOAT64X
> -# if __HAVE_FLOAT64X_LONG_DOUBLE
> -# define __MATH_TG_BUILTIN_CLASSIFY_F64X(BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - _Float64x: DEFINE ? BUILTIN ARGS_B : __ ## FUNC ## l ARGS_F,
> -# else
> -# define __MATH_TG_BUILTIN_CLASSIFY_F64X(BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - _Float64x: DEFINE ? BUILTIN ARGS_B : __ ## FUNC ## f128 ARGS_F,
> -# endif
> -# else
> -# define __MATH_TG_BUILTIN_CLASSIFY_F64X(BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE)
> -# endif
> -# define __MATH_TG_BUILTIN_CLASSIFY_F128(BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - _Float128: BUILTIN ARGS_B
> -# define __MATH_TG_BUILTIN_CLASSIFY(TG_ARG, BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - _Generic ((TG_ARG), \
> - float: BUILTIN ARGS_B, \
> - __MATH_TG_BUILTIN_CLASSIFY_F32 (BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - default: BUILTIN ARGS_B, \
> - __MATH_TG_BUILTIN_CLASSIFY_LDOUBLE (BUILTIN, ARGS_B, FUNC, \
> - ARGS_F, DEFINE) \
> - __MATH_TG_BUILTIN_CLASSIFY_F64X (BUILTIN, ARGS_B, FUNC, ARGS_F,\
> - DEFINE) \
> - __MATH_TG_BUILTIN_CLASSIFY_F128 (BUILTIN, ARGS_B, FUNC, ARGS_F,\
> - DEFINE))
> -# else
> -# define __MATH_TG_BUILTIN_CLASSIFY(TG_ARG, BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - __builtin_choose_expr \
> - (__builtin_types_compatible_p (__typeof (TG_ARG), float), \
> - __builtin ## BUILTIN ARGS_B, \
> - __builtin_choose_expr \
> - (__builtin_types_compatible_p (__typeof (TG_ARG), double), \
> - __builtin ## BUILTIN ARGS_B, \
> - __builtin_choose_expr \
> - (__builtin_types_compatible_p (__typeof (TG_ARG), long double), \
> - DEFINE ? BUILTIN ARGS_B : __ ## FUNC ## l ARGS_F, \
> - BUILTIN ARGS_B)))
> -# endif
> -#else
> -# define __MATH_TG_BUILTIN_CLASSIFY(TG_ARG, BUILTIN, ARGS_B, FUNC, ARGS_F, \
> - DEFINE) \
> - (sizeof (TG_ARG) == sizeof (float) \
> - ? BUILTIN ARGS_B \
> - : sizeof (TG_ARG) == sizeof (double) \
> - ? BUILTIN ARGS_B \
> - : DEFINE ? BUILTIN ARGS_B : __ ## FUNC ## l ARGS_F)
> -#endif
> -
> /* ISO C99 defines some generic macros which work on any data type. */
> #ifdef __USE_ISOC99
>
> @@ -1181,13 +1101,8 @@ enum
> with -Os. No further use of this definition of fpclassify is
> expected in C++ mode, since libstdc++ provides its own version
> of fpclassify in cmath (which undefines fpclassify). */
> -# define fpclassify(x) \
> - __MATH_TG_BUILTIN_CLASSIFY ((x), \
> - __builtin_fpclassify, (FP_NAN, FP_INFINITE, \
> - FP_NORMAL, FP_SUBNORMAL, \
> - FP_ZERO, x), \
> - fpclassify, (x), \
> - __FP_BUILTIN_FPCLASSIFY_DENORMAL)
> +# define fpclassify(x) __builtin_fpclassify (FP_NAN, FP_INFINITE, \
> + FP_NORMAL, FP_SUBNORMAL, FP_ZERO, x)
> # else
> # define fpclassify(x) __MATH_TG ((x), __fpclassify, (x))
> # endif
> diff --git a/sysdeps/x86/bits/fp-builtin-denormal.h b/sysdeps/x86/bits/fp-builtin-denormal.h
> deleted file mode 100644
> index 33549c876a..0000000000
> --- a/sysdeps/x86/bits/fp-builtin-denormal.h
> +++ /dev/null
> @@ -1,25 +0,0 @@
> -/* Define __FP_BUILTIN_DENORMAL.
> - Copyright (C) 2025-2026 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 _MATH_H
> -# error "Never use <bits/fp-builtin-denormal.h> directly; include <math.h> instead."
> -#endif
> -
> -/* Neither GCC (bug 123161) nor clang (issue 172533) handles pseudo-normal
> - numbers correctly with fpclassify builtin. */
> -#define __FP_BUILTIN_FPCLASSIFY_DENORMAL 0
> diff --git a/sysdeps/x86/fpu/Makefile b/sysdeps/x86/fpu/Makefile
> index 8566d3c80f..91b2444f3a 100644
> --- a/sysdeps/x86/fpu/Makefile
> +++ b/sysdeps/x86/fpu/Makefile
> @@ -4,7 +4,6 @@ CPPFLAGS += -I../soft-fp
>
> libm-support += powl_helper
> tests += \
> - test-builtin-denormal \
> test-fenv-clear-sse \
> test-fenv-sse \
> test-fenv-sse-2 \
> diff --git a/sysdeps/x86/fpu/test-builtin-denormal.c b/sysdeps/x86/fpu/test-builtin-denormal.c
> deleted file mode 100644
> index 6584ec339f..0000000000
> --- a/sysdeps/x86/fpu/test-builtin-denormal.c
> +++ /dev/null
> @@ -1,53 +0,0 @@
> -/* Ccheck if math.h optimizations to call compiler builtin
> - does not trigger FE_INVALID on x86 denormal numbers.
> - Copyright (C) 2025-2026 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 <array_length.h>
> -#include <fenv.h>
> -#include <math.h>
> -#include <math_ldbl.h>
> -#include <support/check.h>
> -
> -#define pseudo_inf { .parts = { 0x00000000, 0x00000000, 0x7fff }}
> -#define pseudo_zero { .parts = { 0x00000000, 0x00000000, 0x0100 }}
> -#define pseudo_qnan { .parts = { 0x00000001, 0x00000000, 0x7fff }}
> -#define pseudo_snan { .parts = { 0x00000001, 0x40000000, 0x7fff }}
> -#define pseudo_unnormal { .parts = { 0x00000001, 0x40000000, 0x0100 }}
> -
> -static const ieee_long_double_shape_type inputs[] = {
> - pseudo_inf,
> - pseudo_zero,
> - pseudo_qnan,
> - pseudo_snan,
> - pseudo_unnormal
> -};
> -
> -static int
> -do_test (void)
> -{
> - for (int i = 0; i < array_length (inputs); i++)
> - {
> - TEST_COMPARE (feclearexcept (FE_INVALID), 0);
> - TEST_COMPARE (fpclassify (inputs[i].value), FP_NAN);
> - TEST_COMPARE (fetestexcept (FE_INVALID), 0);
> - }
> -
> - return 0;
> -}
> -
> -#include <support/test-driver.c>
--
Cheers,
Carlos.
More information about the Libc-alpha
mailing list