Implement std::midpoint for floating point types

[mclow.lists]

Summary

finish up the implementation of std::midpoint - add the floating point implementation.
Add a new header file in test/support for doing FP testing.

Set the feature test macros; even though we don't do lerp yet - that's coming next.

[mclow.lists]

For some reason; the tests didn't get included.

[EricWF]

Can you update utils/generate_feature_test_macro_components.py and re-running it?

libcxx/include/numeric

573

Can you update the #endif to comment what it's closing?

libcxx/include/version

58

These changes should be made by updating utils/generate_feature_test_macro_components.py and re-running it.

[zoecarver]

libcxx/include/numeric

148

Is there any overlap between cmath and numeric which might confuse the compiler?

libcxx/include/version

226

nit: every other line has a space between # and define.

[zoecarver]

libcxx/test/std/numerics/numeric.ops/numeric.ops.midpoint/midpoint.float.pass.cpp

39 ↗

(On Diff #196241)

Another nit: maybe use ASSERT_SAME_TYPE?

[mclow.lists]

Updated the <version> header using the tool.
Fixed the check to use ASSERT_SAME_TYPE and ASSERT_NOEXCEPT

[mclow.lists]

libcxx/include/numeric

148

I don't think so. I played around a bit; ran the whole test suite.

[zoecarver]

LGTM :)

libcxx/include/numeric

148

Okay, nevermind then.

[EricWF]

This LGTM.

libcxx/include/numeric

568

I'm not sure it matters, but I would prefer if the calls to abs and isnormal were qualified.

libcxx/test/support/fp_compare.h

40 ↗

(On Diff #196332)

Should 100.0 be cast to T to avoid promotion?

[mclow.lists]

Committed as revision 359184

libcxx/test/support/fp_compare.h

40 ↗

(On Diff #196332)

Done.

[a_Tom]

libcxx/include/numeric

559

Shouldn't this function be constexpr?

[zoecarver]

libcxx/include/numeric

564

Also, maybe use _LIBCPP_CONSTEXPR.

[mclow.lists]

libcxx/include/numeric

564

We don't use _LIBCPP_CONSTEXPR in new code - code that requires C++20 (for example), because we assume that modern compilers support constexpr. _LIBCPP_CONSTEXPR is mostly useful for things that should be constexpr, but either (a) have to exist in C++03/11, or (b) for supporting compilers that have incomplete implementations of constexpr.

I am aware that we are not 100% consistent on this.

[mclow.lists]

libcxx/include/numeric

559

Yes, it should - and it is now.

libcxx/include/numeric

 #include <iterator>
 #include <limits> // for numeric_limits
 #include <functional>
+#include <cmath> // for isnormal
 #include <version>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)

 {
     return __a + _VSTD::midpoint(ptrdiff_t(0), __b - __a);
 }
+
+
+template <typename _Tp>
+int __sign(_Tp __val) {
+    return (_Tp(0) < __val) - (__val < _Tp(0));
+}
+
+template <class _Fp>
+_LIBCPP_INLINE_VISIBILITY constexpr
+enable_if_t<is_floating_point_v<_Fp>, _Fp>
+midpoint(_Fp __a, _Fp __b) noexcept
+{
+    return isnormal(__a) && isnormal(__b) 
+       && ((__sign(__a) != __sign(__b)) || ((numeric_limits<_Fp>::max() - abs(__a)) < abs(__b)))
+             ?  __a / 2 + __b / 2
+             : (__a + __b) / 2;
+}
 #endif
 
 _LIBCPP_END_NAMESPACE_STD

libcxx/include/version

 __cpp_lib_incomplete_container_elements                 201505L <forward_list> <list> <vector>
 __cpp_lib_integer_sequence                              201304L <utility>
 __cpp_lib_integral_constant_callable                    201304L <type_traits>
+__cpp_lib_interpolate                                   201902L <numeric>
 __cpp_lib_invoke                                        201411L <functional>
 __cpp_lib_is_aggregate                                  201703L <type_traits>
 __cpp_lib_is_constant_evaluated                         201811L <type_traits>

 // # define __cpp_lib_destroying_delete                    201806L
 # define __cpp_lib_erase_if                             201811L
 // # define __cpp_lib_generic_unordered_lookup             201811L
+#define __cpp_lib_interpolate                           201902L
 // # define __cpp_lib_is_constant_evaluated                201811L
 // # define __cpp_lib_list_remove_return_type              201806L
 // # define __cpp_lib_ranges                               201811L

libcxx/test/std/language.support/support.limits/support.limits.general/numeric.version.pass.cpp

 /*  Constant                        Value
     __cpp_lib_gcd_lcm               201606L [C++17]
     __cpp_lib_parallel_algorithm    201603L [C++17]
+    __cpp_lib_interpolate           201902L [C++20]
 */
 
 #include <numeric>

 
 #elif TEST_STD_VER > 17
 
+# ifndef __cpp_lib_interpolate
+#   error "__cpp_lib_interpolate should be defined in c++2a"
+# endif
+# if __cpp_lib_interpolate != 201902L
+#   error "__cpp_lib_interpolate should have the value 201902L in c++2a"
+# endif
+
 # ifndef __cpp_lib_gcd_lcm
 #   error "__cpp_lib_gcd_lcm should be defined in c++2a"
 # endif

libcxx/test/std/language.support/support.limits/support.limits.general/version.version.pass.cpp

     __cpp_lib_incomplete_container_elements        201505L [C++17]
     __cpp_lib_integer_sequence                     201304L [C++14]
     __cpp_lib_integral_constant_callable           201304L [C++14]
+    __cpp_lib_interpolate                          201902L [C++20]
     __cpp_lib_invoke                               201411L [C++17]
     __cpp_lib_is_aggregate                         201703L [C++17]
     __cpp_lib_is_constant_evaluated                201811L [C++2a]

 #   error "__cpp_lib_integral_constant_callable should have the value 201304L in c++2a"
 # endif
 
+# ifndef __cpp_lib_interpolate
+#   error "__cpp_lib_interpolate should be defined in c++2a"
+# endif
+# if __cpp_lib_interpolate != 201902L
+#   error "__cpp_lib_interpolate should have the value 201902L in c++2a"
+# endif
+
 # ifndef __cpp_lib_invoke
 #   error "__cpp_lib_invoke should be defined in c++2a"
 # endif