| Message ID | 20240403193919.1533786-4-adhemerval.zanella@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | Improve x86 rounding implementation when FE_INEXACT trap is enabled | expand |
On Wed, Apr 3, 2024 at 12:39 PM Adhemerval Zanella <adhemerval.zanella@linaro.org> wrote: > > The implementations of trunc functions using x87 floating point (i386 and > x86_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-trunc-except-2.c | 67 +++++++++++++++++++ > sysdeps/i386/fpu/{s_trunc.S => s_trunc.c} | 24 ++----- > sysdeps/i386/fpu/{s_truncf.S => s_truncf.c} | 24 ++----- > sysdeps/i386/fpu/s_truncl.S | 40 ----------- > .../fpu/s_truncl.S => x86/fpu/s_truncl.c} | 23 ++----- > 6 files changed, 87 insertions(+), 93 deletions(-) > create mode 100644 math/test-trunc-except-2.c > rename sysdeps/i386/fpu/{s_trunc.S => s_trunc.c} (69%) > rename sysdeps/i386/fpu/{s_truncf.S => s_truncf.c} (68%) > delete mode 100644 sysdeps/i386/fpu/s_truncl.S > rename sysdeps/{x86_64/fpu/s_truncl.S => x86/fpu/s_truncl.c} (70%) > > diff --git a/math/Makefile b/math/Makefile > index 121fe2881a..a9fef9e2db 100644 > --- a/math/Makefile > +++ b/math/Makefile > @@ -539,6 +539,7 @@ tests = \ > test-tgmath-int \ > test-tgmath-ret \ > test-tgmath2 \ > + test-trunc-except-2 \ > tst-CMPLX \ > tst-CMPLX2 \ > tst-definitions \ > @@ -993,6 +994,7 @@ CFLAGS-test-nan-const.c += -fno-builtin > > CFLAGS-test-ceil-except-2.c += -fno-builtin > CFLAGS-test-floor-except-2.c += -fno-builtin > +CFLAGS-test-trunc-except-2.c += -fno-builtin > > include ../Rules > > diff --git a/math/test-trunc-except-2.c b/math/test-trunc-except-2.c > new file mode 100644 > index 0000000000..8933c6ab41 > --- /dev/null > +++ b/math/test-trunc-except-2.c > @@ -0,0 +1,67 @@ > +/* Test trunc 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; \ > + /* trunc must work when traps on "inexact" are enabled. */ \ > + b = trunc ## 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_trunc.S b/sysdeps/i386/fpu/s_trunc.c > similarity index 69% > rename from sysdeps/i386/fpu/s_trunc.S > rename to sysdeps/i386/fpu/s_trunc.c > index 40e45c9f9c..ac16f4967c 100644 > --- a/sysdeps/i386/fpu/s_trunc.S > +++ b/sysdeps/i386/fpu/s_trunc.c > @@ -1,5 +1,5 @@ > -/* Truncate double value. > - Copyright (C) 1997-2024 Free Software Foundation, Inc. > +/* Round to integer, toward zero. 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 > @@ -16,22 +16,10 @@ > License along with the GNU C Library; if not, see > <https://www.gnu.org/licenses/>. */ > > -#include <machine/asm.h> > #include <libm-alias-double.h> > > -ENTRY(__trunc) > - fldl 4(%esp) > - subl $32, %esp > - cfi_adjust_cfa_offset (32) > - fnstenv 4(%esp) > - movl $0xc00, %edx > - orl 4(%esp), %edx > - movl %edx, (%esp) > - fldcw (%esp) > - frndint > - fldenv 4(%esp) > - addl $32, %esp > - cfi_adjust_cfa_offset (-32) > - ret > -END(__trunc) > +#define FUNC __trunc > +#define TYPE double > +#define FE_OPTION FE_TOWARDZERO > +#include "s_nearestint_387_template.c" > libm_alias_double (__trunc, trunc) > diff --git a/sysdeps/i386/fpu/s_truncf.S b/sysdeps/i386/fpu/s_truncf.c > similarity index 68% > rename from sysdeps/i386/fpu/s_truncf.S > rename to sysdeps/i386/fpu/s_truncf.c > index 0b26e09d61..240d3507ef 100644 > --- a/sysdeps/i386/fpu/s_truncf.S > +++ b/sysdeps/i386/fpu/s_truncf.c > @@ -1,5 +1,5 @@ > -/* Truncate float value. > - Copyright (C) 1997-2024 Free Software Foundation, Inc. > +/* Round to integer, toward zero. 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 > @@ -16,22 +16,10 @@ > License along with the GNU C Library; if not, see > <https://www.gnu.org/licenses/>. */ > > -#include <machine/asm.h> > #include <libm-alias-float.h> > > -ENTRY(__truncf) > - flds 4(%esp) > - subl $32, %esp > - cfi_adjust_cfa_offset (32) > - fnstenv 4(%esp) > - movl $0xc00, %edx > - orl 4(%esp), %edx > - movl %edx, (%esp) > - fldcw (%esp) > - frndint > - fldenv 4(%esp) > - addl $32, %esp > - cfi_adjust_cfa_offset (-32) > - ret > -END(__truncf) > +#define FUNC __truncf > +#define TYPE float > +#define FE_OPTION FE_TOWARDZERO > +#include "s_nearestint_387_template.c" > libm_alias_float (__trunc, trunc) > diff --git a/sysdeps/i386/fpu/s_truncl.S b/sysdeps/i386/fpu/s_truncl.S > deleted file mode 100644 > index dfd0ca4a57..0000000000 > --- a/sysdeps/i386/fpu/s_truncl.S > +++ /dev/null > @@ -1,40 +0,0 @@ > -/* Truncate long double value. > - Copyright (C) 1997-2024 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> > -#include <machine/asm.h> > - > -ENTRY(__truncl) > - fldt 4(%esp) > - subl $32, %esp > - cfi_adjust_cfa_offset (32) > - fnstenv 4(%esp) > - movl $0xc00, %edx > - orl 4(%esp), %edx > - movl %edx, (%esp) > - fldcw (%esp) > - frndint > - fnstsw > - andl $0x1, %eax > - orl %eax, 8(%esp) > - fldenv 4(%esp) > - addl $32, %esp > - cfi_adjust_cfa_offset (-32) > - ret > -END(__truncl) > -libm_alias_ldouble (__trunc, trunc) > diff --git a/sysdeps/x86_64/fpu/s_truncl.S b/sysdeps/x86/fpu/s_truncl.c > similarity index 70% > rename from sysdeps/x86_64/fpu/s_truncl.S > rename to sysdeps/x86/fpu/s_truncl.c > index e3d64a84e8..e2bac7fa38 100644 > --- a/sysdeps/x86_64/fpu/s_truncl.S > +++ b/sysdeps/x86/fpu/s_truncl.c > @@ -1,5 +1,5 @@ > -/* Truncate long double value. > - Copyright (C) 1997-2024 Free Software Foundation, Inc. > +/* Round to integer, toward zero. 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 > @@ -17,20 +17,9 @@ > <https://www.gnu.org/licenses/>. */ > > #include <libm-alias-ldouble.h> > -#include <machine/asm.h> > > -ENTRY(__truncl) > - fldt 8(%rsp) > - fnstenv -28(%rsp) > - movl $0xc00, %edx > - orl -28(%rsp), %edx > - movl %edx, -32(%rsp) > - fldcw -32(%rsp) > - frndint > - fnstsw > - andl $0x1, %eax > - orl %eax, -24(%rsp) > - fldenv -28(%rsp) > - ret > -END(__truncl) > +#define FUNC __truncl > +#define TYPE long double > +#define FE_OPTION FE_TOWARDZERO > +#include "s_nearestint_387_template.c" > libm_alias_ldouble (__trunc, trunc) > -- > 2.34.1 > LGTM. Reviewed-by: H.J. Lu <hjl.tools@gmail.com> Thanks.
Hi HJ, there is a typo in the subject: "math: math:" should be "math:", no? Paul
On 04/04/24 02:27, Paul Zimmermann wrote: > Hi HJ, > > there is a typo in the subject: "math: math:" should be "math:", no? > Yeah, I will fix it.
diff --git a/math/Makefile b/math/Makefile index 121fe2881a..a9fef9e2db 100644 --- a/math/Makefile +++ b/math/Makefile @@ -539,6 +539,7 @@ tests = \ test-tgmath-int \ test-tgmath-ret \ test-tgmath2 \ + test-trunc-except-2 \ tst-CMPLX \ tst-CMPLX2 \ tst-definitions \ @@ -993,6 +994,7 @@ CFLAGS-test-nan-const.c += -fno-builtin CFLAGS-test-ceil-except-2.c += -fno-builtin CFLAGS-test-floor-except-2.c += -fno-builtin +CFLAGS-test-trunc-except-2.c += -fno-builtin include ../Rules diff --git a/math/test-trunc-except-2.c b/math/test-trunc-except-2.c new file mode 100644 index 0000000000..8933c6ab41 --- /dev/null +++ b/math/test-trunc-except-2.c @@ -0,0 +1,67 @@ +/* Test trunc 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; \ + /* trunc must work when traps on "inexact" are enabled. */ \ + b = trunc ## 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_trunc.S b/sysdeps/i386/fpu/s_trunc.c similarity index 69% rename from sysdeps/i386/fpu/s_trunc.S rename to sysdeps/i386/fpu/s_trunc.c index 40e45c9f9c..ac16f4967c 100644 --- a/sysdeps/i386/fpu/s_trunc.S +++ b/sysdeps/i386/fpu/s_trunc.c @@ -1,5 +1,5 @@ -/* Truncate double value. - Copyright (C) 1997-2024 Free Software Foundation, Inc. +/* Round to integer, toward zero. 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 @@ -16,22 +16,10 @@ License along with the GNU C Library; if not, see <https://www.gnu.org/licenses/>. */ -#include <machine/asm.h> #include <libm-alias-double.h> -ENTRY(__trunc) - fldl 4(%esp) - subl $32, %esp - cfi_adjust_cfa_offset (32) - fnstenv 4(%esp) - movl $0xc00, %edx - orl 4(%esp), %edx - movl %edx, (%esp) - fldcw (%esp) - frndint - fldenv 4(%esp) - addl $32, %esp - cfi_adjust_cfa_offset (-32) - ret -END(__trunc) +#define FUNC __trunc +#define TYPE double +#define FE_OPTION FE_TOWARDZERO +#include "s_nearestint_387_template.c" libm_alias_double (__trunc, trunc) diff --git a/sysdeps/i386/fpu/s_truncf.S b/sysdeps/i386/fpu/s_truncf.c similarity index 68% rename from sysdeps/i386/fpu/s_truncf.S rename to sysdeps/i386/fpu/s_truncf.c index 0b26e09d61..240d3507ef 100644 --- a/sysdeps/i386/fpu/s_truncf.S +++ b/sysdeps/i386/fpu/s_truncf.c @@ -1,5 +1,5 @@ -/* Truncate float value. - Copyright (C) 1997-2024 Free Software Foundation, Inc. +/* Round to integer, toward zero. 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 @@ -16,22 +16,10 @@ License along with the GNU C Library; if not, see <https://www.gnu.org/licenses/>. */ -#include <machine/asm.h> #include <libm-alias-float.h> -ENTRY(__truncf) - flds 4(%esp) - subl $32, %esp - cfi_adjust_cfa_offset (32) - fnstenv 4(%esp) - movl $0xc00, %edx - orl 4(%esp), %edx - movl %edx, (%esp) - fldcw (%esp) - frndint - fldenv 4(%esp) - addl $32, %esp - cfi_adjust_cfa_offset (-32) - ret -END(__truncf) +#define FUNC __truncf +#define TYPE float +#define FE_OPTION FE_TOWARDZERO +#include "s_nearestint_387_template.c" libm_alias_float (__trunc, trunc) diff --git a/sysdeps/i386/fpu/s_truncl.S b/sysdeps/i386/fpu/s_truncl.S deleted file mode 100644 index dfd0ca4a57..0000000000 --- a/sysdeps/i386/fpu/s_truncl.S +++ /dev/null @@ -1,40 +0,0 @@ -/* Truncate long double value. - Copyright (C) 1997-2024 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> -#include <machine/asm.h> - -ENTRY(__truncl) - fldt 4(%esp) - subl $32, %esp - cfi_adjust_cfa_offset (32) - fnstenv 4(%esp) - movl $0xc00, %edx - orl 4(%esp), %edx - movl %edx, (%esp) - fldcw (%esp) - frndint - fnstsw - andl $0x1, %eax - orl %eax, 8(%esp) - fldenv 4(%esp) - addl $32, %esp - cfi_adjust_cfa_offset (-32) - ret -END(__truncl) -libm_alias_ldouble (__trunc, trunc) diff --git a/sysdeps/x86_64/fpu/s_truncl.S b/sysdeps/x86/fpu/s_truncl.c similarity index 70% rename from sysdeps/x86_64/fpu/s_truncl.S rename to sysdeps/x86/fpu/s_truncl.c index e3d64a84e8..e2bac7fa38 100644 --- a/sysdeps/x86_64/fpu/s_truncl.S +++ b/sysdeps/x86/fpu/s_truncl.c @@ -1,5 +1,5 @@ -/* Truncate long double value. - Copyright (C) 1997-2024 Free Software Foundation, Inc. +/* Round to integer, toward zero. 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 @@ -17,20 +17,9 @@ <https://www.gnu.org/licenses/>. */ #include <libm-alias-ldouble.h> -#include <machine/asm.h> -ENTRY(__truncl) - fldt 8(%rsp) - fnstenv -28(%rsp) - movl $0xc00, %edx - orl -28(%rsp), %edx - movl %edx, -32(%rsp) - fldcw -32(%rsp) - frndint - fnstsw - andl $0x1, %eax - orl %eax, -24(%rsp) - fldenv -28(%rsp) - ret -END(__truncl) +#define FUNC __truncl +#define TYPE long double +#define FE_OPTION FE_TOWARDZERO +#include "s_nearestint_387_template.c" libm_alias_ldouble (__trunc, trunc)
The implementations of trunc functions using x87 floating point (i386 and x86_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-trunc-except-2.c | 67 +++++++++++++++++++ sysdeps/i386/fpu/{s_trunc.S => s_trunc.c} | 24 ++----- sysdeps/i386/fpu/{s_truncf.S => s_truncf.c} | 24 ++----- sysdeps/i386/fpu/s_truncl.S | 40 ----------- .../fpu/s_truncl.S => x86/fpu/s_truncl.c} | 23 ++----- 6 files changed, 87 insertions(+), 93 deletions(-) create mode 100644 math/test-trunc-except-2.c rename sysdeps/i386/fpu/{s_trunc.S => s_trunc.c} (69%) rename sysdeps/i386/fpu/{s_truncf.S => s_truncf.c} (68%) delete mode 100644 sysdeps/i386/fpu/s_truncl.S rename sysdeps/{x86_64/fpu/s_truncl.S => x86/fpu/s_truncl.c} (70%)