* [PATCH 0/3] Improve x86 rounding implementation when FE_INEXACT trap is enabled
@ 2024-04-03 19:39 Adhemerval Zanella
2024-04-03 19:39 ` [PATCH 1/3] math: math: x86 ceill traps when FE_INEXACT is enabled (BZ 31600) Adhemerval Zanella
` (2 more replies)
0 siblings, 3 replies; 9+ messages in thread
From: Adhemerval Zanella @ 2024-04-03 19:39 UTC (permalink / raw)
To: libc-alpha; +Cc: H . J . Lu, Joseph Myers
Some x86 rounding implementation that uses 387 raises invalid
floating point exceptions when traps are enabled. This is a GNU
extension outside the scope of the C standard.
Adhemerval Zanella (3):
math: math: x86 ceill traps when FE_INEXACT is enabled (BZ 31600)
math: math: x86 floor traps when FE_INEXACT is enabled (BZ 31601)
math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603)
math/Makefile | 7 ++
math/test-ceil-except-2.c | 67 +++++++++++++++++++
math/test-floor-except-2.c | 67 +++++++++++++++++++
math/test-trunc-except-2.c | 67 +++++++++++++++++++
sysdeps/i386/fpu/s_ceil.S | 34 ----------
sysdeps/i386/fpu/s_ceil.c | 25 +++++++
sysdeps/i386/fpu/s_ceilf.S | 34 ----------
sysdeps/i386/fpu/s_ceilf.c | 25 +++++++
sysdeps/i386/fpu/s_ceill.S | 39 -----------
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/i386/fpu/{s_trunc.S => s_trunc.c} | 24 ++-----
sysdeps/i386/fpu/{s_truncf.S => s_truncf.c} | 24 ++-----
.../fpu/s_truncl.S => x86/fpu/s_ceill.c} | 29 ++------
sysdeps/x86/fpu/s_floorl.c | 25 +++++++
sysdeps/x86/fpu/s_nearestint_387_template.c | 36 ++++++++++
.../fpu/s_truncl.S => x86/fpu/s_truncl.c} | 23 ++-----
sysdeps/x86_64/fpu/s_ceill.S | 34 ----------
sysdeps/x86_64/fpu/s_floorl.S | 33 ---------
22 files changed, 394 insertions(+), 356 deletions(-)
create mode 100644 math/test-ceil-except-2.c
create mode 100644 math/test-floor-except-2.c
create mode 100644 math/test-trunc-except-2.c
delete mode 100644 sysdeps/i386/fpu/s_ceil.S
create mode 100644 sysdeps/i386/fpu/s_ceil.c
delete mode 100644 sysdeps/i386/fpu/s_ceilf.S
create mode 100644 sysdeps/i386/fpu/s_ceilf.c
delete mode 100644 sysdeps/i386/fpu/s_ceill.S
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
rename sysdeps/i386/fpu/{s_trunc.S => s_trunc.c} (69%)
rename sysdeps/i386/fpu/{s_truncf.S => s_truncf.c} (68%)
rename sysdeps/{i386/fpu/s_truncl.S => x86/fpu/s_ceill.c} (63%)
create mode 100644 sysdeps/x86/fpu/s_floorl.c
create mode 100644 sysdeps/x86/fpu/s_nearestint_387_template.c
rename sysdeps/{x86_64/fpu/s_truncl.S => x86/fpu/s_truncl.c} (70%)
delete mode 100644 sysdeps/x86_64/fpu/s_ceill.S
delete mode 100644 sysdeps/x86_64/fpu/s_floorl.S
--
2.34.1
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 1/3] math: math: x86 ceill traps when FE_INEXACT is enabled (BZ 31600)
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 ` 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 19:39 ` [PATCH 3/3] math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603) Adhemerval Zanella
2 siblings, 1 reply; 9+ messages in thread
From: Adhemerval Zanella @ 2024-04-03 19:39 UTC (permalink / raw)
To: libc-alpha; +Cc: H . J . Lu, Joseph Myers
The implementations of ceil 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 | 3 +
math/test-ceil-except-2.c | 67 +++++++++++++++++++++
sysdeps/i386/fpu/s_ceil.S | 34 -----------
sysdeps/i386/fpu/s_ceil.c | 25 ++++++++
sysdeps/i386/fpu/s_ceilf.S | 34 -----------
sysdeps/i386/fpu/s_ceilf.c | 25 ++++++++
sysdeps/i386/fpu/s_ceill.S | 39 ------------
sysdeps/x86/fpu/s_ceill.c | 25 ++++++++
sysdeps/x86/fpu/s_nearestint_387_template.c | 36 +++++++++++
sysdeps/x86_64/fpu/s_ceill.S | 34 -----------
10 files changed, 181 insertions(+), 141 deletions(-)
create mode 100644 math/test-ceil-except-2.c
delete mode 100644 sysdeps/i386/fpu/s_ceil.S
create mode 100644 sysdeps/i386/fpu/s_ceil.c
delete mode 100644 sysdeps/i386/fpu/s_ceilf.S
create mode 100644 sysdeps/i386/fpu/s_ceilf.c
delete mode 100644 sysdeps/i386/fpu/s_ceill.S
create mode 100644 sysdeps/x86/fpu/s_ceill.c
create mode 100644 sysdeps/x86/fpu/s_nearestint_387_template.c
delete mode 100644 sysdeps/x86_64/fpu/s_ceill.S
diff --git a/math/Makefile b/math/Makefile
index 121a709121..d2a740eebe 100644
--- a/math/Makefile
+++ b/math/Makefile
@@ -498,6 +498,7 @@ tests = \
bug-nextafter \
bug-nexttoward \
bug-tgmath1 \
+ test-ceil-except-2 \
test-femode \
test-femode-traps \
test-fenv basic-test \
@@ -989,6 +990,8 @@ 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
+
include ../Rules
gen-all-calls = $(gen-libm-calls) $(gen-calls)
diff --git a/math/test-ceil-except-2.c b/math/test-ceil-except-2.c
new file mode 100644
index 0000000000..394a272d89
--- /dev/null
+++ b/math/test-ceil-except-2.c
@@ -0,0 +1,67 @@
+/* Test ceil 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; \
+ /* ceil must work when traps on "inexact" are enabled. */ \
+ b = ceil ## 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_ceil.S b/sysdeps/i386/fpu/s_ceil.S
deleted file mode 100644
index 99984f9b8d..0000000000
--- a/sysdeps/i386/fpu/s_ceil.S
+++ /dev/null
@@ -1,34 +0,0 @@
-/*
- * Public domain.
- */
-
-#include <machine/asm.h>
-#include <libm-alias-double.h>
-
-RCSID("$NetBSD: s_ceil.S,v 1.4 1995/05/08 23:52:13 jtc Exp $")
-
-ENTRY(__ceil)
- 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 $0x0800,%edx /* round towards +oo */
- orl 4(%esp),%edx
- andl $0xfbff,%edx
- movl %edx,(%esp)
- fldcw (%esp) /* load modified control word */
-
- frndint /* round */
-
- fldenv 4(%esp) /* restore original environment */
-
- addl $32,%esp
- cfi_adjust_cfa_offset (-32)
- ret
-END (__ceil)
-libm_alias_double (__ceil, ceil)
diff --git a/sysdeps/i386/fpu/s_ceil.c b/sysdeps/i386/fpu/s_ceil.c
new file mode 100644
index 0000000000..349135c5d3
--- /dev/null
+++ b/sysdeps/i386/fpu/s_ceil.c
@@ -0,0 +1,25 @@
+/* Return smallest integral value not less than argument. i386 version.
+ Copyright (C) 2024 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <libm-alias-double.h>
+
+#define FUNC __ceil
+#define TYPE double
+#define FE_OPTION FE_UPWARD
+#include "s_nearestint_387_template.c"
+libm_alias_double (__ceil, ceil)
diff --git a/sysdeps/i386/fpu/s_ceilf.S b/sysdeps/i386/fpu/s_ceilf.S
deleted file mode 100644
index 03e8e22609..0000000000
--- a/sysdeps/i386/fpu/s_ceilf.S
+++ /dev/null
@@ -1,34 +0,0 @@
-/*
- * Public domain.
- */
-
-#include <machine/asm.h>
-#include <libm-alias-float.h>
-
-RCSID("$NetBSD: s_ceilf.S,v 1.3 1995/05/08 23:52:44 jtc Exp $")
-
-ENTRY(__ceilf)
- 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 $0x0800,%edx /* round towards +oo */
- orl 4(%esp),%edx
- andl $0xfbff,%edx
- movl %edx,(%esp)
- fldcw (%esp) /* load modified control word */
-
- frndint /* round */
-
- fldenv 4(%esp) /* restore original environment */
-
- addl $32,%esp
- cfi_adjust_cfa_offset (-32)
- ret
-END (__ceilf)
-libm_alias_float (__ceil, ceil)
diff --git a/sysdeps/i386/fpu/s_ceilf.c b/sysdeps/i386/fpu/s_ceilf.c
new file mode 100644
index 0000000000..e73a20fd71
--- /dev/null
+++ b/sysdeps/i386/fpu/s_ceilf.c
@@ -0,0 +1,25 @@
+/* Return largest integral value not less than argument. i386 version.
+ Copyright (C) 2024 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <libm-alias-float.h>
+
+#define FUNC __ceilf
+#define TYPE float
+#define FE_OPTION FE_UPWARD
+#include "s_nearestint_387_template.c"
+libm_alias_float (__ceil, ceil)
diff --git a/sysdeps/i386/fpu/s_ceill.S b/sysdeps/i386/fpu/s_ceill.S
deleted file mode 100644
index a551fce7f9..0000000000
--- a/sysdeps/i386/fpu/s_ceill.S
+++ /dev/null
@@ -1,39 +0,0 @@
-/*
- * Public domain.
- */
-
-#include <libm-alias-ldouble.h>
-#include <machine/asm.h>
-
-RCSID("$NetBSD: $")
-
-ENTRY(__ceill)
- 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 $0x0800,%edx /* round towards +oo */
- orl 4(%esp),%edx
- andl $0xfbff,%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 (__ceill)
-libm_alias_ldouble (__ceil, ceil)
diff --git a/sysdeps/x86/fpu/s_ceill.c b/sysdeps/x86/fpu/s_ceill.c
new file mode 100644
index 0000000000..860dd2c960
--- /dev/null
+++ b/sysdeps/x86/fpu/s_ceill.c
@@ -0,0 +1,25 @@
+/* Return smallest 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 __ceill
+#define TYPE long double
+#define FE_OPTION FE_UPWARD
+#include "s_nearestint_387_template.c"
+libm_alias_ldouble (__ceil, ceil)
diff --git a/sysdeps/x86/fpu/s_nearestint_387_template.c b/sysdeps/x86/fpu/s_nearestint_387_template.c
new file mode 100644
index 0000000000..95fca93f87
--- /dev/null
+++ b/sysdeps/x86/fpu/s_nearestint_387_template.c
@@ -0,0 +1,36 @@
+/* Nearest integet template for x86.
+ 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/>. */
+
+#define NO_MATH_REDIRECT
+#include <math.h>
+#include <fenv_private.h>
+
+TYPE
+FUNC (TYPE x)
+{
+ fenv_t fenv;
+ TYPE r;
+
+ libc_feholdexcept_setround_387 (&fenv, FE_OPTION);
+ asm volatile ("frndint" : "=t" (r) : "0" (x));
+ /* Preserve "invalid" exceptions from sNaN input. */
+ fenv.__status_word |= libc_fetestexcept_387 (FE_INVALID);
+ libc_fesetenv_387 (&fenv);
+
+ return r;
+}
diff --git a/sysdeps/x86_64/fpu/s_ceill.S b/sysdeps/x86_64/fpu/s_ceill.S
deleted file mode 100644
index 16dbecd56d..0000000000
--- a/sysdeps/x86_64/fpu/s_ceill.S
+++ /dev/null
@@ -1,34 +0,0 @@
-/*
- * Public domain.
- */
-
-#include <libm-alias-ldouble.h>
-#include <machine/asm.h>
-
-
-ENTRY(__ceill)
- 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 $0x0800,%edx /* round towards +oo */
- orl -28(%rsp),%edx
- andl $0xfbff,%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 (__ceill)
-libm_alias_ldouble (__ceil, ceil)
--
2.34.1
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 2/3] math: math: x86 floor traps when FE_INEXACT is enabled (BZ 31601)
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 19:39 ` Adhemerval Zanella
2024-04-03 20:03 ` H.J. Lu
2024-04-03 19:39 ` [PATCH 3/3] math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603) Adhemerval Zanella
2 siblings, 1 reply; 9+ messages in thread
From: Adhemerval Zanella @ 2024-04-03 19:39 UTC (permalink / raw)
To: libc-alpha; +Cc: H . J . Lu, Joseph Myers
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
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 3/3] math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603)
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 19:39 ` [PATCH 2/3] math: math: x86 floor traps when FE_INEXACT is enabled (BZ 31601) Adhemerval Zanella
@ 2024-04-03 19:39 ` Adhemerval Zanella
2024-04-03 20:04 ` H.J. Lu
2 siblings, 1 reply; 9+ messages in thread
From: Adhemerval Zanella @ 2024-04-03 19:39 UTC (permalink / raw)
To: libc-alpha; +Cc: H . J . Lu, Joseph Myers
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
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 1/3] math: math: x86 ceill traps when FE_INEXACT is enabled (BZ 31600)
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
0 siblings, 0 replies; 9+ messages in thread
From: H.J. Lu @ 2024-04-03 20:03 UTC (permalink / raw)
To: Adhemerval Zanella; +Cc: libc-alpha, Joseph Myers
On Wed, Apr 3, 2024 at 12:39 PM Adhemerval Zanella
<adhemerval.zanella@linaro.org> wrote:
>
> The implementations of ceil 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 | 3 +
> math/test-ceil-except-2.c | 67 +++++++++++++++++++++
> sysdeps/i386/fpu/s_ceil.S | 34 -----------
> sysdeps/i386/fpu/s_ceil.c | 25 ++++++++
> sysdeps/i386/fpu/s_ceilf.S | 34 -----------
> sysdeps/i386/fpu/s_ceilf.c | 25 ++++++++
> sysdeps/i386/fpu/s_ceill.S | 39 ------------
> sysdeps/x86/fpu/s_ceill.c | 25 ++++++++
> sysdeps/x86/fpu/s_nearestint_387_template.c | 36 +++++++++++
> sysdeps/x86_64/fpu/s_ceill.S | 34 -----------
> 10 files changed, 181 insertions(+), 141 deletions(-)
> create mode 100644 math/test-ceil-except-2.c
> delete mode 100644 sysdeps/i386/fpu/s_ceil.S
> create mode 100644 sysdeps/i386/fpu/s_ceil.c
> delete mode 100644 sysdeps/i386/fpu/s_ceilf.S
> create mode 100644 sysdeps/i386/fpu/s_ceilf.c
> delete mode 100644 sysdeps/i386/fpu/s_ceill.S
> create mode 100644 sysdeps/x86/fpu/s_ceill.c
> create mode 100644 sysdeps/x86/fpu/s_nearestint_387_template.c
> delete mode 100644 sysdeps/x86_64/fpu/s_ceill.S
>
> diff --git a/math/Makefile b/math/Makefile
> index 121a709121..d2a740eebe 100644
> --- a/math/Makefile
> +++ b/math/Makefile
> @@ -498,6 +498,7 @@ tests = \
> bug-nextafter \
> bug-nexttoward \
> bug-tgmath1 \
> + test-ceil-except-2 \
> test-femode \
> test-femode-traps \
> test-fenv basic-test \
> @@ -989,6 +990,8 @@ 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
> +
> include ../Rules
>
> gen-all-calls = $(gen-libm-calls) $(gen-calls)
> diff --git a/math/test-ceil-except-2.c b/math/test-ceil-except-2.c
> new file mode 100644
> index 0000000000..394a272d89
> --- /dev/null
> +++ b/math/test-ceil-except-2.c
> @@ -0,0 +1,67 @@
> +/* Test ceil 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; \
> + /* ceil must work when traps on "inexact" are enabled. */ \
> + b = ceil ## 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_ceil.S b/sysdeps/i386/fpu/s_ceil.S
> deleted file mode 100644
> index 99984f9b8d..0000000000
> --- a/sysdeps/i386/fpu/s_ceil.S
> +++ /dev/null
> @@ -1,34 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <machine/asm.h>
> -#include <libm-alias-double.h>
> -
> -RCSID("$NetBSD: s_ceil.S,v 1.4 1995/05/08 23:52:13 jtc Exp $")
> -
> -ENTRY(__ceil)
> - 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 $0x0800,%edx /* round towards +oo */
> - orl 4(%esp),%edx
> - andl $0xfbff,%edx
> - movl %edx,(%esp)
> - fldcw (%esp) /* load modified control word */
> -
> - frndint /* round */
> -
> - fldenv 4(%esp) /* restore original environment */
> -
> - addl $32,%esp
> - cfi_adjust_cfa_offset (-32)
> - ret
> -END (__ceil)
> -libm_alias_double (__ceil, ceil)
> diff --git a/sysdeps/i386/fpu/s_ceil.c b/sysdeps/i386/fpu/s_ceil.c
> new file mode 100644
> index 0000000000..349135c5d3
> --- /dev/null
> +++ b/sysdeps/i386/fpu/s_ceil.c
> @@ -0,0 +1,25 @@
> +/* Return smallest integral value not less than argument. i386 version.
> + Copyright (C) 2024 Free Software Foundation, Inc.
> + This file is part of the GNU C Library.
> +
> + The GNU C Library is free software; you can redistribute it and/or
> + modify it under the terms of the GNU Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 of the License, or (at your option) any later version.
> +
> + The GNU C Library is distributed in the hope that it will be useful,
> + but WITHOUT ANY WARRANTY; without even the implied warranty of
> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <https://www.gnu.org/licenses/>. */
> +
> +#include <libm-alias-double.h>
> +
> +#define FUNC __ceil
> +#define TYPE double
> +#define FE_OPTION FE_UPWARD
> +#include "s_nearestint_387_template.c"
> +libm_alias_double (__ceil, ceil)
> diff --git a/sysdeps/i386/fpu/s_ceilf.S b/sysdeps/i386/fpu/s_ceilf.S
> deleted file mode 100644
> index 03e8e22609..0000000000
> --- a/sysdeps/i386/fpu/s_ceilf.S
> +++ /dev/null
> @@ -1,34 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <machine/asm.h>
> -#include <libm-alias-float.h>
> -
> -RCSID("$NetBSD: s_ceilf.S,v 1.3 1995/05/08 23:52:44 jtc Exp $")
> -
> -ENTRY(__ceilf)
> - 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 $0x0800,%edx /* round towards +oo */
> - orl 4(%esp),%edx
> - andl $0xfbff,%edx
> - movl %edx,(%esp)
> - fldcw (%esp) /* load modified control word */
> -
> - frndint /* round */
> -
> - fldenv 4(%esp) /* restore original environment */
> -
> - addl $32,%esp
> - cfi_adjust_cfa_offset (-32)
> - ret
> -END (__ceilf)
> -libm_alias_float (__ceil, ceil)
> diff --git a/sysdeps/i386/fpu/s_ceilf.c b/sysdeps/i386/fpu/s_ceilf.c
> new file mode 100644
> index 0000000000..e73a20fd71
> --- /dev/null
> +++ b/sysdeps/i386/fpu/s_ceilf.c
> @@ -0,0 +1,25 @@
> +/* Return largest integral value not less than argument. i386 version.
> + Copyright (C) 2024 Free Software Foundation, Inc.
> + This file is part of the GNU C Library.
> +
> + The GNU C Library is free software; you can redistribute it and/or
> + modify it under the terms of the GNU Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 of the License, or (at your option) any later version.
> +
> + The GNU C Library is distributed in the hope that it will be useful,
> + but WITHOUT ANY WARRANTY; without even the implied warranty of
> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <https://www.gnu.org/licenses/>. */
> +
> +#include <libm-alias-float.h>
> +
> +#define FUNC __ceilf
> +#define TYPE float
> +#define FE_OPTION FE_UPWARD
> +#include "s_nearestint_387_template.c"
> +libm_alias_float (__ceil, ceil)
> diff --git a/sysdeps/i386/fpu/s_ceill.S b/sysdeps/i386/fpu/s_ceill.S
> deleted file mode 100644
> index a551fce7f9..0000000000
> --- a/sysdeps/i386/fpu/s_ceill.S
> +++ /dev/null
> @@ -1,39 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <libm-alias-ldouble.h>
> -#include <machine/asm.h>
> -
> -RCSID("$NetBSD: $")
> -
> -ENTRY(__ceill)
> - 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 $0x0800,%edx /* round towards +oo */
> - orl 4(%esp),%edx
> - andl $0xfbff,%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 (__ceill)
> -libm_alias_ldouble (__ceil, ceil)
> diff --git a/sysdeps/x86/fpu/s_ceill.c b/sysdeps/x86/fpu/s_ceill.c
> new file mode 100644
> index 0000000000..860dd2c960
> --- /dev/null
> +++ b/sysdeps/x86/fpu/s_ceill.c
> @@ -0,0 +1,25 @@
> +/* Return smallest 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 __ceill
> +#define TYPE long double
> +#define FE_OPTION FE_UPWARD
> +#include "s_nearestint_387_template.c"
> +libm_alias_ldouble (__ceil, ceil)
> diff --git a/sysdeps/x86/fpu/s_nearestint_387_template.c b/sysdeps/x86/fpu/s_nearestint_387_template.c
> new file mode 100644
> index 0000000000..95fca93f87
> --- /dev/null
> +++ b/sysdeps/x86/fpu/s_nearestint_387_template.c
> @@ -0,0 +1,36 @@
> +/* Nearest integet template for x86.
> + 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/>. */
> +
> +#define NO_MATH_REDIRECT
> +#include <math.h>
> +#include <fenv_private.h>
> +
> +TYPE
> +FUNC (TYPE x)
> +{
> + fenv_t fenv;
> + TYPE r;
> +
> + libc_feholdexcept_setround_387 (&fenv, FE_OPTION);
> + asm volatile ("frndint" : "=t" (r) : "0" (x));
> + /* Preserve "invalid" exceptions from sNaN input. */
> + fenv.__status_word |= libc_fetestexcept_387 (FE_INVALID);
> + libc_fesetenv_387 (&fenv);
> +
> + return r;
> +}
> diff --git a/sysdeps/x86_64/fpu/s_ceill.S b/sysdeps/x86_64/fpu/s_ceill.S
> deleted file mode 100644
> index 16dbecd56d..0000000000
> --- a/sysdeps/x86_64/fpu/s_ceill.S
> +++ /dev/null
> @@ -1,34 +0,0 @@
> -/*
> - * Public domain.
> - */
> -
> -#include <libm-alias-ldouble.h>
> -#include <machine/asm.h>
> -
> -
> -ENTRY(__ceill)
> - 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 $0x0800,%edx /* round towards +oo */
> - orl -28(%rsp),%edx
> - andl $0xfbff,%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 (__ceill)
> -libm_alias_ldouble (__ceil, ceil)
> --
> 2.34.1
>
LGTM.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Thanks.
--
H.J.
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 2/3] math: math: x86 floor traps when FE_INEXACT is enabled (BZ 31601)
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
0 siblings, 0 replies; 9+ messages in thread
From: H.J. Lu @ 2024-04-03 20:03 UTC (permalink / raw)
To: Adhemerval Zanella; +Cc: libc-alpha, Joseph Myers
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.
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603)
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
0 siblings, 1 reply; 9+ messages in thread
From: H.J. Lu @ 2024-04-03 20:04 UTC (permalink / raw)
To: Adhemerval Zanella; +Cc: libc-alpha, Joseph Myers
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.
--
H.J.
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603)
2024-04-03 20:04 ` H.J. Lu
@ 2024-04-04 5:27 ` Paul Zimmermann
2024-04-04 11:52 ` Adhemerval Zanella Netto
0 siblings, 1 reply; 9+ messages in thread
From: Paul Zimmermann @ 2024-04-04 5:27 UTC (permalink / raw)
To: H.J. Lu; +Cc: adhemerval.zanella, libc-alpha, josmyers
Hi HJ,
there is a typo in the subject: "math: math:" should be "math:", no?
Paul
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] math: math: x86 trunc traps when FE_INEXACT is enabled (BZ 31603)
2024-04-04 5:27 ` Paul Zimmermann
@ 2024-04-04 11:52 ` Adhemerval Zanella Netto
0 siblings, 0 replies; 9+ messages in thread
From: Adhemerval Zanella Netto @ 2024-04-04 11:52 UTC (permalink / raw)
To: Paul Zimmermann, H.J. Lu; +Cc: libc-alpha, josmyers
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.
^ permalink raw reply [flat|nested] 9+ messages in thread
end of thread, other threads:[~2024-04-04 11:53 UTC | newest]
Thread overview: 9+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
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
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
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).