From: "H.J. Lu" <hjl.tools@gmail.com>
To: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Cc: libc-alpha@sourceware.org, Joseph Myers <josmyers@redhat.com>
Subject: Re: [PATCH 2/3] math: math: x86 floor traps when FE_INEXACT is enabled (BZ 31601)
Date: Wed, 3 Apr 2024 13:03:38 -0700 [thread overview]
Message-ID: <CAMe9rOoC1G_W30Hs1Bj=-VHJRV_NCO+sNNi9+SUFUHs1Zs8edw@mail.gmail.com> (raw)
In-Reply-To: <20240403193919.1533786-3-adhemerval.zanella@linaro.org>
On Wed, Apr 3, 2024 at 12:39 PM Adhemerval Zanella
<adhemerval.zanella@linaro.org> wrote:
>
> The implementations of floor functions using x87 floating point (i386 and
> 86_64 long double only) traps when FE_INEXACT is enabled. Although
> this is a GNU extension outside the scope of the C standard, other
> architectures that also support traps do not show this behavior.
>
> The fix moves the implementation to a common one that holds any
> exceptions with a 'fnclex' (libc_feholdexcept_setround_387).
>
> Checked on x86_64-linux-gnu and i686-linux-gnu.
> ---
> math/Makefile | 2 ++
> math/test-floor-except-2.c | 67 +++++++++++++++++++++++++++++++++++
> sysdeps/i386/fpu/s_floor.S | 34 ------------------
> sysdeps/i386/fpu/s_floor.c | 25 +++++++++++++
> sysdeps/i386/fpu/s_floorf.S | 34 ------------------
> sysdeps/i386/fpu/s_floorf.c | 25 +++++++++++++
> sysdeps/i386/fpu/s_floorl.S | 39 --------------------
> sysdeps/x86/fpu/s_floorl.c | 25 +++++++++++++
> sysdeps/x86_64/fpu/s_floorl.S | 33 -----------------
> 9 files changed, 144 insertions(+), 140 deletions(-)
> create mode 100644 math/test-floor-except-2.c
> delete mode 100644 sysdeps/i386/fpu/s_floor.S
> create mode 100644 sysdeps/i386/fpu/s_floor.c
> delete mode 100644 sysdeps/i386/fpu/s_floorf.S
> create mode 100644 sysdeps/i386/fpu/s_floorf.c
> delete mode 100644 sysdeps/i386/fpu/s_floorl.S
> create mode 100644 sysdeps/x86/fpu/s_floorl.c
> delete mode 100644 sysdeps/x86_64/fpu/s_floorl.S
>
> diff --git a/math/Makefile b/math/Makefile
> index d2a740eebe..121fe2881a 100644
> --- a/math/Makefile
> +++ b/math/Makefile
> @@ -511,6 +511,7 @@ tests = \
> test-fetestexceptflag \
> test-fexcept \
> test-fexcept-traps \
> + test-floor-except-2 \
> test-flt-eval-method \
> test-fp-ilogb-constants \
> test-fp-llogb-constants \
> @@ -991,6 +992,7 @@ CFLAGS-test-fe-snans-always-signal.c += $(config-cflags-signaling-nans)
> CFLAGS-test-nan-const.c += -fno-builtin
>
> CFLAGS-test-ceil-except-2.c += -fno-builtin
> +CFLAGS-test-floor-except-2.c += -fno-builtin
>
> include ../Rules
>
> diff --git a/math/test-floor-except-2.c b/math/test-floor-except-2.c
> new file mode 100644
> index 0000000000..d99e835909
> --- /dev/null
> +++ b/math/test-floor-except-2.c
> @@ -0,0 +1,67 @@
> +/* Test floor functions do not disable exception traps.
> + 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 <fenv.h>
> +#include <math.h>
> +#include <stdio.h>
> +
> +#ifndef FE_INEXACT
> +# define FE_INEXACT 0
> +#endif
> +
> +#define TEST_FUNC(NAME, FLOAT, SUFFIX) \
> +static int \
> +NAME (void) \
> +{ \
> + int result = 0; \
> + volatile FLOAT a, b __attribute__ ((unused)); \
> + a = 1.5; \
> + /* floor must work when traps on "inexact" are enabled. */ \
> + b = floor ## SUFFIX (a); \
> + /* And it must have left those traps enabled. */ \
> + if (fegetexcept () == FE_INEXACT) \
> + puts ("PASS: " #FLOAT); \
> + else \
> + { \
> + puts ("FAIL: " #FLOAT); \
> + result = 1; \
> + } \
> + return result; \
> +}
> +
> +TEST_FUNC (float_test, float, f)
> +TEST_FUNC (double_test, double, )
> +TEST_FUNC (ldouble_test, long double, l)
> +
> +static int
> +do_test (void)
> +{
> + if (feenableexcept (FE_INEXACT) == -1)
> + {
> + puts ("enabling FE_INEXACT traps failed, cannot test");
> + return 77;
> + }
> + int result = float_test ();
> + feenableexcept (FE_INEXACT);
> + result |= double_test ();
> + feenableexcept (FE_INEXACT);
> + result |= ldouble_test ();
> + return result;
> +}
> +
> +#include <support/test-driver.c>
> diff --git a/sysdeps/i386/fpu/s_floor.S b/sysdeps/i386/fpu/s_floor.S
> deleted file mode 100644
> index 7143fdcc9a..0000000000
> --- a/sysdeps/i386/fpu/s_floor.S
> +++ /dev/null
> @@ -1,34 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <machine/asm.h>
> -#include <libm-alias-double.h>
> -
> -RCSID("$NetBSD: s_floor.S,v 1.4 1995/05/09 00:01:59 jtc Exp $")
> -
> -ENTRY(__floor)
> - fldl 4(%esp)
> - subl $32,%esp
> - cfi_adjust_cfa_offset (32)
> -
> - fnstenv 4(%esp) /* store fpu environment */
> -
> - /* We use here %edx although only the low 1 bits are defined.
> - But none of the operations should care and they are faster
> - than the 16 bit operations. */
> - movl $0x400,%edx /* round towards -oo */
> - orl 4(%esp),%edx
> - andl $0xf7ff,%edx
> - movl %edx,(%esp)
> - fldcw (%esp) /* load modified control word */
> -
> - frndint /* round */
> -
> - fldenv 4(%esp) /* restore original environment */
> -
> - addl $32,%esp
> - cfi_adjust_cfa_offset (-32)
> - ret
> -END (__floor)
> -libm_alias_double (__floor, floor)
> diff --git a/sysdeps/i386/fpu/s_floor.c b/sysdeps/i386/fpu/s_floor.c
> new file mode 100644
> index 0000000000..cc50e33b59
> --- /dev/null
> +++ b/sysdeps/i386/fpu/s_floor.c
> @@ -0,0 +1,25 @@
> +/* Return smallest integral value not less than argument. i386 version.
> + Copyright (C) 2024 Free Software Foundation, Inc.
> + This file is part of the GNU C Library.
> +
> + The GNU C Library is free software; you can redistribute it and/or
> + modify it under the terms of the GNU Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 of the License, or (at your option) any later version.
> +
> + The GNU C Library is distributed in the hope that it will be useful,
> + but WITHOUT ANY WARRANTY; without even the implied warranty of
> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <https://www.gnu.org/licenses/>. */
> +
> +#include <libm-alias-double.h>
> +
> +#define FUNC __floor
> +#define TYPE double
> +#define FE_OPTION FE_DOWNWARD
> +#include "s_nearestint_387_template.c"
> +libm_alias_double (__floor, floor)
> diff --git a/sysdeps/i386/fpu/s_floorf.S b/sysdeps/i386/fpu/s_floorf.S
> deleted file mode 100644
> index 8fad9c0698..0000000000
> --- a/sysdeps/i386/fpu/s_floorf.S
> +++ /dev/null
> @@ -1,34 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <machine/asm.h>
> -#include <libm-alias-float.h>
> -
> -RCSID("$NetBSD: s_floorf.S,v 1.3 1995/05/09 00:04:32 jtc Exp $")
> -
> -ENTRY(__floorf)
> - flds 4(%esp)
> - subl $32,%esp
> - cfi_adjust_cfa_offset (32)
> -
> - fnstenv 4(%esp) /* store fpu environment */
> -
> - /* We use here %edx although only the low 1 bits are defined.
> - But none of the operations should care and they are faster
> - than the 16 bit operations. */
> - movl $0x400,%edx /* round towards -oo */
> - orl 4(%esp),%edx
> - andl $0xf7ff,%edx
> - movl %edx,(%esp)
> - fldcw (%esp) /* load modified control word */
> -
> - frndint /* round */
> -
> - fldenv 4(%esp) /* restore original environment */
> -
> - addl $32,%esp
> - cfi_adjust_cfa_offset (-32)
> - ret
> -END (__floorf)
> -libm_alias_float (__floor, floor)
> diff --git a/sysdeps/i386/fpu/s_floorf.c b/sysdeps/i386/fpu/s_floorf.c
> new file mode 100644
> index 0000000000..fa9454e56b
> --- /dev/null
> +++ b/sysdeps/i386/fpu/s_floorf.c
> @@ -0,0 +1,25 @@
> +/* Largest integral value not greater than argument i386 version.
> + Copyright (C) 2024 Free Software Foundation, Inc.
> + This file is part of the GNU C Library.
> +
> + The GNU C Library is free software; you can redistribute it and/or
> + modify it under the terms of the GNU Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 of the License, or (at your option) any later version.
> +
> + The GNU C Library is distributed in the hope that it will be useful,
> + but WITHOUT ANY WARRANTY; without even the implied warranty of
> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <https://www.gnu.org/licenses/>. */
> +
> +#include <libm-alias-float.h>
> +
> +#define FUNC __floorf
> +#define TYPE float
> +#define FE_OPTION FE_DOWNWARD
> +#include "s_nearestint_387_template.c"
> +libm_alias_float (__floor, floor)
> diff --git a/sysdeps/i386/fpu/s_floorl.S b/sysdeps/i386/fpu/s_floorl.S
> deleted file mode 100644
> index 3ec28b477b..0000000000
> --- a/sysdeps/i386/fpu/s_floorl.S
> +++ /dev/null
> @@ -1,39 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <libm-alias-ldouble.h>
> -#include <machine/asm.h>
> -
> -RCSID("$NetBSD: $")
> -
> -ENTRY(__floorl)
> - fldt 4(%esp)
> - subl $32,%esp
> - cfi_adjust_cfa_offset (32)
> -
> - fnstenv 4(%esp) /* store fpu environment */
> -
> - /* We use here %edx although only the low 1 bits are defined.
> - But none of the operations should care and they are faster
> - than the 16 bit operations. */
> - movl $0x400,%edx /* round towards -oo */
> - orl 4(%esp),%edx
> - andl $0xf7ff,%edx
> - movl %edx,(%esp)
> - fldcw (%esp) /* load modified control word */
> -
> - frndint /* round */
> -
> - /* Preserve "invalid" exceptions from sNaN input. */
> - fnstsw
> - andl $0x1, %eax
> - orl %eax, 8(%esp)
> -
> - fldenv 4(%esp) /* restore original environment */
> -
> - addl $32,%esp
> - cfi_adjust_cfa_offset (-32)
> - ret
> -END (__floorl)
> -libm_alias_ldouble (__floor, floor)
> diff --git a/sysdeps/x86/fpu/s_floorl.c b/sysdeps/x86/fpu/s_floorl.c
> new file mode 100644
> index 0000000000..9c92d33fbe
> --- /dev/null
> +++ b/sysdeps/x86/fpu/s_floorl.c
> @@ -0,0 +1,25 @@
> +/* Return largest integral value not less than argument. x86 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 <libm-alias-ldouble.h>
> +
> +#define FUNC __floorl
> +#define TYPE long double
> +#define FE_OPTION FE_DOWNWARD
> +#include "s_nearestint_387_template.c"
> +libm_alias_ldouble (__floor, floor)
> diff --git a/sysdeps/x86_64/fpu/s_floorl.S b/sysdeps/x86_64/fpu/s_floorl.S
> deleted file mode 100644
> index b74d1a4d6b..0000000000
> --- a/sysdeps/x86_64/fpu/s_floorl.S
> +++ /dev/null
> @@ -1,33 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <libm-alias-ldouble.h>
> -#include <machine/asm.h>
> -
> -ENTRY(__floorl)
> - fldt 8(%rsp)
> -
> - fnstenv -28(%rsp) /* store fpu environment */
> -
> - /* We use here %edx although only the low 1 bits are defined.
> - But none of the operations should care and they are faster
> - than the 16 bit operations. */
> - movl $0x400,%edx /* round towards -oo */
> - orl -28(%rsp),%edx
> - andl $0xf7ff,%edx
> - movl %edx,-32(%rsp)
> - fldcw -32(%rsp) /* load modified control word */
> -
> - frndint /* round */
> -
> - /* Preserve "invalid" exceptions from sNaN input. */
> - fnstsw
> - andl $0x1, %eax
> - orl %eax, -24(%rsp)
> -
> - fldenv -28(%rsp) /* restore original environment */
> -
> - ret
> -END (__floorl)
> -libm_alias_ldouble (__floor, floor)
> --
> 2.34.1
>
LGTM.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Thanks.
--
H.J.
next prev parent reply other threads:[~2024-04-03 20:04 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-04-03 19:39 [PATCH 0/3] Improve x86 rounding implementation when FE_INEXACT trap is enabled Adhemerval Zanella
2024-04-03 19:39 ` [PATCH 1/3] math: math: x86 ceill traps when FE_INEXACT is enabled (BZ 31600) Adhemerval Zanella
2024-04-03 20:03 ` H.J. Lu
2024-04-03 19:39 ` [PATCH 2/3] math: math: x86 floor traps when FE_INEXACT is enabled (BZ 31601) Adhemerval Zanella
2024-04-03 20:03 ` H.J. Lu [this message]
2024-04-03 19:39 ` [PATCH 3/3] math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603) Adhemerval Zanella
2024-04-03 20:04 ` H.J. Lu
2024-04-04 5:27 ` Paul Zimmermann
2024-04-04 11:52 ` Adhemerval Zanella Netto
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
List information: https://www.gnu.org/software/libc/involved.html
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CAMe9rOoC1G_W30Hs1Bj=-VHJRV_NCO+sNNi9+SUFUHs1Zs8edw@mail.gmail.com' \
--to=hjl.tools@gmail.com \
--cc=adhemerval.zanella@linaro.org \
--cc=josmyers@redhat.com \
--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).