Hi!
The following patch implements nextafter for std::{,b}float16_t,
though right now only without constexpr support, and adds a testcase for it.
The testcase unfortunately relevealed I've screwed up testing of my last
patch. I've tested earlier version of the patch with
--target_board=unix/-std=c++23
but didn't test the final version with that RUNTESTFLAGS, so missed
an invalid call to std::sph_neumann (too many arguments) in the test.
And, I've made a typo in the guard for numeric_limits (the reason
for the guard is I wanted to avoid defining a large macro that nothing
will then use because the std::{,b}float*_t types are C++23 only) and
so numeric_limits wasn't specialized for the types at all but
testsuite/18_support/headers/limits/synopsis_cxx23.cc test didn't
detect that.
In the nextafter implementation I'm calling __builtin_nextafterf
to get various required side-effects for nextafter from 0/-0, or from max
to inf or from min to largest subnormal to avoid needing to set errno
inline, or use inline asm specific for each processor to force math
evaluation barriers. Dunno if
#ifdef __INT16_TYPE__
using __float16_int_type = __INT16_TYPE__;
#else
using __float16_int_type = short int;
#endif
isn't too ugly, perhaps we could just blindly use short int and hope
or even assert it has the same size as _Float16 or __gnu_cxx::__bfloat16_t?
Only aarch64, arm, csky, gcn, x86, nvptx and riscv support these types
and all of them have 16-bit short (I think the only target with some
other short size is avr with certain command line switches where both
short and int are 8-bit, but such mode isn't compatible with C and C++
requirements).
Bootstrapped/regtested on x86_64-linux and i686-linux, additionally
tested with --target_board=unix/-std=c++23, ok for trunk?
2022-10-19 Jakub Jelinek <jakub@redhat.com>
PR c++/106652
* include/std/limits: Fix a typo, 202202L -> 202002L.
(numeric_limits::<_Float16>::radix, numeric_limits::<_Float32>::radix,
numeric_limits::<_Float64>::radix, numeric_limits::<_Float128>::radix,
numeric_limits::<__gnu_cxx::__bfloat16_t>::radix: Use __FLT_RADIX__
macro instead of type specific macros.
* include/c_global/cmath (nextafter(_Float16, _Float16)): New
overload.
(nextafter(__gnu_cxx::__bfloat16_t, __gnu_cxx::__bfloat16_t)):
Likewise.
* testsuite/26_numerics/headers/cmath/functions_std_c++23.cc
(test_functions): Uncomment nextafter test. Fix up sph_neumann call.
* testsuite/26_numerics/headers/cmath/nextafter_c++23.cc: New test.
Jakub
On Wed, 19 Oct 2022 at 09:23, Jakub Jelinek wrote:
>
> Hi!
>
> The following patch implements nextafter for std::{,b}float16_t,
> though right now only without constexpr support, and adds a testcase for it.
> The testcase unfortunately relevealed I've screwed up testing of my last
> patch. I've tested earlier version of the patch with
> --target_board=unix/-std=c++23
> but didn't test the final version with that RUNTESTFLAGS, so missed
> an invalid call to std::sph_neumann (too many arguments) in the test.
> And, I've made a typo in the guard for numeric_limits (the reason
> for the guard is I wanted to avoid defining a large macro that nothing
> will then use because the std::{,b}float*_t types are C++23 only) and
> so numeric_limits wasn't specialized for the types at all but
> testsuite/18_support/headers/limits/synopsis_cxx23.cc test didn't
> detect that.
> In the nextafter implementation I'm calling __builtin_nextafterf
> to get various required side-effects for nextafter from 0/-0, or from max
> to inf or from min to largest subnormal to avoid needing to set errno
> inline, or use inline asm specific for each processor to force math
> evaluation barriers. Dunno if
> #ifdef __INT16_TYPE__
> using __float16_int_type = __INT16_TYPE__;
> #else
> using __float16_int_type = short int;
> #endif
> isn't too ugly, perhaps we could just blindly use short int and hope
> or even assert it has the same size as _Float16 or __gnu_cxx::__bfloat16_t?
I'm happy with the preprocessor conditions above, but using short int
and a static assert would be OK too.
> Only aarch64, arm, csky, gcn, x86, nvptx and riscv support these types
> and all of them have 16-bit short (I think the only target with some
> other short size is avr with certain command line switches where both
> short and int are 8-bit, but such mode isn't compatible with C and C++
> requirements).
It's not, but we do try to support avr, especially for something like
<limits> which requires no OS support and has always been required for
freestanding. But will it have float16 anyway? I suppose it might do,
so maybe the preprocessor dance for __INT16_TYPE__ is better. It's
only needed in two places, so isn't too bad.
> --- libstdc++-v3/testsuite/26_numerics/headers/cmath/nextafter_c++23.cc.jj 2022-10-18 17:18:32.237061706 +0200
> +++ libstdc++-v3/testsuite/26_numerics/headers/cmath/nextafter_c++23.cc 2022-10-18 18:51:21.465258133 +0200
> @@ -0,0 +1,124 @@
> +// Copyright (C) 2022 Free Software Foundation, Inc.
> +//
> +// This file is part of the GNU ISO C++ Library. This library is free
> +// software; you can redistribute it and/or modify it under the
> +// terms of the GNU General Public License as published by the
> +// Free Software Foundation; either version 3, or (at your option)
> +// any later version.
> +
> +// This 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 General Public License for more details.
> +
> +// You should have received a copy of the GNU General Public License along
> +// with this library; see the file COPYING3. If not see
> +// <http://www.gnu.org/licenses/>.
> +
> +// { dg-do run { target c++23 } }
This should have { dg-options "-std=gnu++2b" } before the dg-do.
OK with that change.
@@ -1890,7 +1890,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
#undef __glibcxx_long_double_traps
#undef __glibcxx_long_double_tinyness_before
-#if __cplusplus > 202202L
+#if __cplusplus > 202002L
#define __glibcxx_concat3_(P,M,S) P ## M ## S
#define __glibcxx_concat3(P,M,S) __glibcxx_concat3_ (P,M,S)
@@ -1924,8 +1924,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
static constexpr bool is_signed = true; \
static constexpr bool is_integer = false; \
static constexpr bool is_exact = false; \
- static constexpr int radix \
- = __glibcxx_concat3 (__FLT, BITSIZE, _RADIX__); \
+ static constexpr int radix = __FLT_RADIX__; \
\
static constexpr _Float##BITSIZE \
epsilon() noexcept \
@@ -2023,7 +2022,7 @@ __glibcxx_float_n(128)
static constexpr bool is_signed = true;
static constexpr bool is_integer = false;
static constexpr bool is_exact = false;
- static constexpr int radix = __BFLT16_RADIX__;
+ static constexpr int radix = __FLT_RADIX__;
static constexpr __gnu_cxx::__bfloat16_t
epsilon() noexcept
@@ -2755,7 +2755,52 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
nearbyint(_Float16 __x)
{ return _Float16(__builtin_nearbyintf(__x)); }
- // nextafter not implemented so far.
+ inline _Float16
+ nextafter(_Float16 __x, _Float16 __y)
+ {
+#ifdef __INT16_TYPE__
+ using __float16_int_type = __INT16_TYPE__;
+#else
+ using __float16_int_type = short int;
+#endif
+ __float16_int_type __hx, __hy, __ix, __iy;
+ __builtin_memcpy(&__hx, &__x, sizeof(__x));
+ __builtin_memcpy(&__hy, &__y, sizeof(__x));
+ __ix = __hx & 0x7fff; // |x|
+ __iy = __hy & 0x7fff; // |y|
+ if (__ix > 0x7c00 || __iy > 0x7c00) // x or y is NaN
+ return __x + __y;
+ if (__x == __y)
+ return __y; // x == y, return y
+ if (__ix == 0) // x == 0
+ {
+ __hy = (__hy & 0x8000) | 1; // return +-__FLT16_DENORM_MIN__
+ __builtin_memcpy(&__x, &__hy, sizeof(__x));
+ __builtin_nextafterf(0.0f, 1.0f); // raise underflow
+ return __x;
+ }
+ if (__hx >= 0) // x > 0
+ {
+ if (__hx > __hy) // x > y, x -= ulp
+ --__hx;
+ else // x < y, x += ulp
+ ++__hx;
+ }
+ else // x < 0
+ {
+ if (__hy >= 0 || __hx > __hy) // x < y, x -= ulp
+ --__hx;
+ else // x > y, x += ulp
+ ++__hx;
+ }
+ __hy = __hx & 0x7c00;
+ if (__hy >= 0x7c00)
+ __builtin_nextafterf(__FLT_MAX__, __builtin_inff()); // overflow
+ else if (__hy < 0x0400)
+ __builtin_nextafterf(__FLT_MIN__, 0.0f); // underflow
+ __builtin_memcpy(&__x, &__hx, sizeof(__x));
+ return __x;
+ }
constexpr _Float16
remainder(_Float16 __x, _Float16 __y)
@@ -3426,7 +3471,52 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
nearbyint(__gnu_cxx::__bfloat16_t __x)
{ return __gnu_cxx::__bfloat16_t(__builtin_nearbyintf(__x)); }
- // nextafter not implemented so far.
+ inline __gnu_cxx::__bfloat16_t
+ nextafter(__gnu_cxx::__bfloat16_t __x, __gnu_cxx::__bfloat16_t __y)
+ {
+#ifdef __INT16_TYPE__
+ using __bfloat16_int_type = __INT16_TYPE__;
+#else
+ using __bfloat16_int_type = short int;
+#endif
+ __bfloat16_int_type __hx, __hy, __ix, __iy;
+ __builtin_memcpy(&__hx, &__x, sizeof(__x));
+ __builtin_memcpy(&__hy, &__y, sizeof(__x));
+ __ix = __hx & 0x7fff; // |x|
+ __iy = __hy & 0x7fff; // |y|
+ if (__ix > 0x7f80 || __iy > 0x7f80) // x or y is NaN
+ return __x + __y;
+ if (__x == __y)
+ return __y; // x == y, return y
+ if (__ix == 0) // x == 0
+ {
+ __hy = (__hy & 0x8000) | 1; // return +-__BFLT16_DENORM_MIN__
+ __builtin_memcpy(&__x, &__hy, sizeof(__x));
+ __builtin_nextafterf(0.0f, 1.0f); // raise underflow
+ return __x;
+ }
+ if (__hx >= 0) // x > 0
+ {
+ if (__hx > __hy) // x > y, x -= ulp
+ --__hx;
+ else // x < y, x += ulp
+ ++__hx;
+ }
+ else // x < 0
+ {
+ if (__hy >= 0 || __hx > __hy) // x < y, x -= ulp
+ --__hx;
+ else // x > y, x += ulp
+ ++__hx;
+ }
+ __hy = __hx & 0x7f80;
+ if (__hy >= 0x7f80)
+ __builtin_nextafterf(__FLT_MAX__, __builtin_inff()); // overflow
+ else if (__hy < 0x0080)
+ __builtin_nextafterf(__FLT_MIN__, 0.0f); // underflow
+ __builtin_memcpy(&__x, &__hx, sizeof(__x));
+ return __x;
+ }
constexpr __gnu_cxx::__bfloat16_t
remainder(__gnu_cxx::__bfloat16_t __x, __gnu_cxx::__bfloat16_t __y)
@@ -63,7 +63,7 @@ test_functions (T *p, int *q, long int *
p[36] = std::fmod (p[36], p[103]);
p[37] = std::remainder (p[37], p[104]);
p[38] = std::copysign (p[38], p[105]);
-// p[39] = std::nextafter (p[39], p[106]);
+ p[39] = std::nextafter (p[39], p[106]);
p[40] = std::fdim (p[40], p[107]);
p[41] = std::fmax (p[41], p[108]);
p[42] = std::fmin (p[42], p[109]);
@@ -102,13 +102,13 @@ test_functions (T *p, int *q, long int *
p[75] = std::riemann_zeta (p[75]);
p[76] = std::sph_bessel (q[12], p[76]);
p[77] = std::sph_legendre (q[13], q[14], p[77]);
- p[78] = std::sph_neumann (q[15], q[16], p[78]);
+ p[78] = std::sph_neumann (q[15], p[78]);
}
int
main ()
{
- int q[17] = {};
+ int q[16] = {};
long int r[16] = {};
long long int s[16] = {};
#if defined(__STDCPP_FLOAT16_T__) && defined(_GLIBCXX_FLOAT_IS_IEEE_BINARY32)
@@ -0,0 +1,124 @@
+// Copyright (C) 2022 Free Software Foundation, Inc.
+//
+// This file is part of the GNU ISO C++ Library. This library is free
+// software; you can redistribute it and/or modify it under the
+// terms of the GNU General Public License as published by the
+// Free Software Foundation; either version 3, or (at your option)
+// any later version.
+
+// This 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 General Public License for more details.
+
+// You should have received a copy of the GNU General Public License along
+// with this library; see the file COPYING3. If not see
+// <http://www.gnu.org/licenses/>.
+
+// { dg-do run { target c++23 } }
+
+#include <stdfloat>
+#include <cmath>
+#include <limits>
+#include <testsuite_hooks.h>
+
+template <typename T>
+void
+test ()
+{
+ using lim = std::numeric_limits<T>;
+ T t0 = std::nextafter(T(-0.0), T(2.0));
+ VERIFY(t0 == lim::denorm_min());
+ T t1 = std::nextafter(T(), T(1.0));
+ VERIFY(t1 == lim::denorm_min());
+ T t2 = std::nextafter(T(), T());
+ VERIFY(t2 == T() && !std::signbit(t2));
+ T t3 = std::nextafter(lim::denorm_min(), T(-2.0));
+ VERIFY(t3 == T() && !std::signbit(t3));
+ T t4 = std::nextafter(lim::min(), T(-0.0));
+ VERIFY(std::fpclassify(t4) == FP_SUBNORMAL && t4 > T());
+ T t5 = std::nextafter(t4, T(1.0));
+ VERIFY(t5 == lim::min());
+ T t6 = std::nextafter(lim::min(), lim::infinity());
+ VERIFY(std::fpclassify(t6) == FP_NORMAL && t6 > lim::min());
+ T t7 = std::nextafter(t6, -lim::infinity());
+ VERIFY(t7 == lim::min());
+ T t8 = std::nextafter(T(16.0), T(16.5));
+ VERIFY(t8 > t7);
+ T t9 = std::nextafter(t8, T(15.5));
+ VERIFY(t9 == T(16.0));
+ T t10 = std::nextafter(lim::max(), T(-0.5));
+ VERIFY(std::fpclassify(t10) == FP_NORMAL && t10 < lim::max());
+ T t11 = std::nextafter(t10, lim::infinity());
+ VERIFY(t11 == lim::max());
+ T t12 = std::nextafter(t11, lim::infinity());
+ VERIFY(std::fpclassify(t12) == FP_INFINITE && !std::signbit(t12));
+ T t13 = std::nextafter(lim::infinity(), t12);
+ VERIFY(t13 == t12);
+ T t14 = std::nextafter(t13, T(1.0));
+ VERIFY(t14 == lim::max());
+ T t15 = std::nextafter(lim::quiet_NaN(), T());
+ VERIFY(std::fpclassify(t15) == FP_NAN);
+ T t16 = std::nextafter(T(17.0), lim::quiet_NaN());
+ VERIFY(std::fpclassify(t16) == FP_NAN);
+ T t17 = std::nextafter(T(), T(-0.0));
+ VERIFY(t17 == T() && std::signbit(t17));
+ T t20 = std::nextafter(T(-0.0), T(-2.0));
+ VERIFY(t20 == -lim::denorm_min());
+ T t21 = std::nextafter(T(), T(-1.0));
+ VERIFY(t21 == -lim::denorm_min());
+ T t22 = std::nextafter(T(-0.0), T(-0.0));
+ VERIFY(t22 == T() && std::signbit(t22));
+ T t23 = std::nextafter(-lim::denorm_min(), T(2.0));
+ VERIFY(t23 == T() && std::signbit(t23));
+ T t24 = std::nextafter(-lim::min(), T());
+ VERIFY(std::fpclassify(t24) == FP_SUBNORMAL && t24 < T());
+ T t25 = std::nextafter(t24, T(-1.0));
+ VERIFY(t25 == -lim::min());
+ T t26 = std::nextafter(-lim::min(), -lim::infinity());
+ VERIFY(std::fpclassify(t26) == FP_NORMAL && t26 < -lim::min());
+ T t27 = std::nextafter(t26, lim::infinity());
+ VERIFY(t27 == -lim::min());
+ T t28 = std::nextafter(T(-16.0), T(-16.5));
+ VERIFY(t28 < t27);
+ T t29 = std::nextafter(t28, T(-15.5));
+ VERIFY(t29 == T(-16.0));
+ T t30 = std::nextafter(-lim::max(), T(0.5));
+ VERIFY(std::fpclassify(t30) == FP_NORMAL && t30 > -lim::max());
+ T t31 = std::nextafter(t30, -lim::infinity());
+ VERIFY(t31 == -lim::max());
+ T t32 = std::nextafter(t31, -lim::infinity());
+ VERIFY(std::fpclassify(t32) == FP_INFINITE && std::signbit(t32));
+ T t33 = std::nextafter(-lim::infinity(), t32);
+ VERIFY(t33 == t32);
+ T t34 = std::nextafter(t33, T(-1.0));
+ VERIFY(t34 == -lim::max());
+ T t35 = std::nextafter(-lim::quiet_NaN(), T());
+ VERIFY(std::fpclassify(t35) == FP_NAN);
+ T t36 = std::nextafter(T(-17.0), lim::quiet_NaN());
+ VERIFY(std::fpclassify(t36) == FP_NAN);
+ T t37 = std::nextafter(T(-0.0), T());
+ VERIFY(t37 == T() && !std::signbit(t37));
+}
+
+int
+main ()
+{
+#if defined(__STDCPP_FLOAT16_T__) && defined(_GLIBCXX_FLOAT_IS_IEEE_BINARY32)
+ test <std::float16_t> ();
+#endif
+#if defined(__STDCPP_FLOAT32_T__) && defined(_GLIBCXX_FLOAT_IS_IEEE_BINARY32)
+ test <std::float32_t> ();
+#endif
+#if defined(__STDCPP_FLOAT64_T__) && defined(_GLIBCXX_DOUBLE_IS_IEEE_BINARY64)
+ test <std::float64_t> ();
+#endif
+#if defined(__STDCPP_FLOAT128_T__) \
+ && (defined(_GLIBCXX_DOUBLE_IS_IEEE_BINARY128) \
+ || defined(_GLIBCXX_HAVE_FLOAT128_MATH))
+ test <std::float128_t> ();
+#endif
+#if defined(__STDCPP_BFLOAT16_T__) && defined(_GLIBCXX_FLOAT_IS_IEEE_BINARY32)
+ test <std::bfloat16_t> ();
+#endif
+}