Using std::numeric_limits<__fp16>::lowest() causes a compilation error

[AnthonyBarbier]

Description

|| In file included from /work/projects/jenkins/scripts/../toolchains//aarch64-linux-android-ndk-r18-beta1/bin/../lib/gcc/aarch64-linux-android/4.9.x/../../../../include/c++/4.9.x/vector:270:
|| In file included from /work/projects/jenkins/scripts/../toolchains//aarch64-linux-android-ndk-r18-beta1/bin/../lib/gcc/aarch64-linux-android/4.9.x/../../../../include/c++/4.9.x/__bit_reference:15:
|| In file included from /work/projects/jenkins/scripts/../toolchains//aarch64-linux-android-ndk-r18-beta1/bin/../lib/gcc/aarch64-linux-android/4.9.x/../../../../include/c++/4.9.x/algorithm:644:
|| In file included from /work/projects/jenkins/scripts/../toolchains//aarch64-linux-android-ndk-r18-beta1/bin/../lib/gcc/aarch64-linux-android/4.9.x/../../../../include/c++/4.9.x/memory:659:
/work/projects/jenkins/toolchains/aarch64-linux-android-ndk-r18-beta1/include/c++/4.9.x/limits|189 col 59| error: invalid operands to binary expression ('float' and 'int')
||     static _LIBCPP_CONSTEXPR const _Tp value = _Tp(_Tp(1) << __digits);
||                                                    ~~~~~~ ^  ~~~~~~~~
/work/projects/jenkins/toolchains/aarch64-linux-android-ndk-r18-beta1/include/c++/4.9.x/limits|210 col 49| note: in instantiation of template class 'std::__ndk1::__libcpp_compute_min<__fp16, 15, true>' requested here
||     static _LIBCPP_CONSTEXPR const type __min = __libcpp_compute_min<type, digits, is_signed>::value;
||                                                 ^
/work/projects/jenkins/toolchains/aarch64-linux-android-ndk-r18-beta1/include/c++/4.9.x/limits|443 col 15| note: in instantiation of template class 'std::__ndk1::__libcpp_numeric_limits<__fp16, true>' requested here
||     : private __libcpp_numeric_limits<typename remove_cv<_Tp>::type>
||               ^
src/core/NEON/kernels/NESoftmaxLayerKernel.cpp|395 col 54| note: in instantiation of template class 'std::__ndk1::numeric_limits<__fp16>' requested here
||         auto vec_max = vdup_n<vec_16_byte_t<T>>(std::numeric_limits<T>::lowest());

Environment Details

ndk r18-beta1 on Ubuntu 16.04 x86 64bit cross compiling for -march=armv8.2-a+fp16 Using a standalone toolchain generated by doing:

$NDK/build/tools/make_standalone_toolchain.py --arch arm64 --install-dir $TOOLCHAINS/aarch64-linux-android-ndk-r18-beta1 --stl libc++ --api 21

[DanAlbert]

I think the only issue is that libc++ is missing a specialization for this (technically the standard doesn't specify one, but neither does it specify that fp16 exists, so it shouldn't be hard to get added).

[DanAlbert]

Not sure when this was fixed upstream, but the following code compiles now:

#include <limits>

int main() {
  std::numeric_limits<__fp16>::lowest();
}