Add host standard library detection

libcudacxx/include/cuda/__cccl_config

@@ -23,6 +23,7 @@
 #include <cuda/std/__cccl/exceptions.h> // IWYU pragma: export
 #include <cuda/std/__cccl/execution_space.h> // IWYU pragma: export
 #include <cuda/std/__cccl/extended_data_types.h> // IWYU pragma: export
+#include <cuda/std/__cccl/host_std_lib.h> // IWYU pragma: export
 #include <cuda/std/__cccl/os.h> // IWYU pragma: export
 #include <cuda/std/__cccl/preprocessor.h> // IWYU pragma: export
 #include <cuda/std/__cccl/ptx_isa.h> // IWYU pragma: export

libcudacxx/include/cuda/__complex/get_real_imag.h

@@ -24,10 +24,6 @@
 #include <cuda/__fwd/complex.h>
 #include <cuda/std/__fwd/complex.h>
 
-#if !_CCCL_COMPILER(NVRTC)
-#  include <complex>
-#endif // !_CCCL_COMPILER(NVRTC)
-
 #include <cuda/std/__cccl/prologue.h>
 
 _CCCL_BEGIN_NAMESPACE_CUDA

libcudacxx/include/cuda/std/__cccl/builtin.h

@@ -25,6 +25,7 @@
 
 #include <cuda/std/__cccl/attributes.h>
 #include <cuda/std/__cccl/extended_data_types.h>
+#include <cuda/std/__cccl/host_std_lib.h>
 
 //! This file consolidates all compiler builtin detection for CCCL.
 //!
@@ -607,55 +608,51 @@
 #  define _CCCL_BUILTIN_STRLEN(...) __builtin_strlen(__VA_ARGS__)
 #endif
 
-// Some compilers provide std::move/std::forward/etc as builtins
-#if defined(__cplusplus)
-// Bring in the feature test macros (needed for std::forward_like)
-#  if _CCCL_HAS_INCLUDE(<version>) // <version> should be the smallest include possible
-#    include <version>
-#  elif !_CCCL_COMPILER(NVRTC)
-#    include <ciso646> // otherwise go for the smallest possible header
-#  endif // !_CCCL_COMPILER(NVRTC)
-
-// Bring in the bits of the STL we need
-#  if defined(_GLIBCXX_VERSION)
-#    include <bits/move.h> // for move, forward, forward_like, and addressof
-#  elif defined(_LIBCPP_VERSION)
-#    include <__memory/addressof.h>
-#    include <__utility/as_const.h>
-#    include <__utility/forward.h>
-#    include <__utility/forward_like.h>
-#    include <__utility/move.h>
-#  endif
-
-#  if defined(_GLIBCXX_VERSION) || defined(_LIBCPP_VERSION)
-// std::move builtin
-#    if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 15)
-#      define _CCCL_HAS_BUILTIN_STD_MOVE() 1
-#    endif
-
-// std::forward builtin
-#    if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 15)
-#      define _CCCL_HAS_BUILTIN_STD_FORWARD() 1
-#    endif
-
-// std::addressof builtin
-#    if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 15)
-#      define _CCCL_HAS_BUILTIN_STD_ADDRESSOF() 1
-#    endif
-
-// std::as_const builtin
-#    if _CCCL_COMPILER(CLANG, >=, 15)
-#      define _CCCL_HAS_BUILTIN_STD_AS_CONST() 1
-#    endif
-
-// std::forward_like builtin
-// Leaving out MSVC for now because it is hard for forward-declare std::forward_like.
-#    if (_CCCL_COMPILER(CLANG, >=, 17) || _CCCL_COMPILER(GCC, >=, 15)) && defined(__cpp_lib_forward_like) \
-      && (__cpp_lib_forward_like >= 202217L)
-#      define _CCCL_HAS_BUILTIN_STD_FORWARD_LIKE() 1
-#    endif
-#  endif // defined(_GLIBCXX_VERSION) || defined(_LIBCPP_VERSION) || defined(_MSVC_STL_VERSION)
-#endif // defined(__cplusplus)
+// todo: re-enable std builtins
+
+// // Some compilers provide std::move/std::forward/etc as builtins
+// #if defined(__cplusplus)
+// // Bring in the bits of the STL we need
+// #  if _CCCL_HOST_STD_LIB(LIBSTDCXX)
+// #    include <bits/move.h> // for move, forward, forward_like, and addressof
+// #  elif _CCCL_HOST_STD_LIB(LIBCXX)
+// #    include <__memory/addressof.h>
+// #    include <__utility/as_const.h>
+// #    include <__utility/forward.h>
+// #    if __cpp_lib_forward_like >= 202217L
+// #      include <__utility/forward_like.h>
+// #    endif // __cpp_lib_forward_like >= 202217L
+// #    include <__utility/move.h>
+// #  endif
+
+// #  if _CCCL_HOST_STD_LIB(LIBSTDCXX) || _CCCL_HOST_STD_LIB(LIBCXX)
+// // std::move builtin
+// #    if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 15)
+// #      define _CCCL_HAS_BUILTIN_STD_MOVE() 1
+// #    endif
+
+// // std::forward builtin
+// #    if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 15)
+// #      define _CCCL_HAS_BUILTIN_STD_FORWARD() 1
+// #    endif
+
+// // std::addressof builtin
+// #    if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 15)
+// #      define _CCCL_HAS_BUILTIN_STD_ADDRESSOF() 1
+// #    endif
+
+// // std::as_const builtin
+// #    if _CCCL_COMPILER(CLANG, >=, 15)
+// #      define _CCCL_HAS_BUILTIN_STD_AS_CONST() 1
+// #    endif
+
+// // std::forward_like builtin
+// // Leaving out MSVC for now because it is hard for forward-declare std::forward_like.
+// #    if (_CCCL_COMPILER(CLANG, >=, 17) || _CCCL_COMPILER(GCC, >=, 15)) && __cpp_lib_forward_like >= 202217L
+// #      define _CCCL_HAS_BUILTIN_STD_FORWARD_LIKE() 1
+// #    endif
+// #  endif // _CCCL_HOST_STD_LIB(LIBSTDCXX) || _CCCL_HOST_STD_LIB(LIBCXX)
+// #endif // defined(__cplusplus)
 
 #ifndef _CCCL_HAS_BUILTIN_STD_MOVE
 #  define _CCCL_HAS_BUILTIN_STD_MOVE() 0

libcudacxx/include/cuda/std/__cccl/host_std_lib.h

@@ -0,0 +1,52 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_HOST_STD_LIB_H
+#define __CCCL_HOST_STD_LIB_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/preprocessor.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#define _CCCL_HOST_STD_LIB_LIBSTDCXX() 0
+#define _CCCL_HOST_STD_LIB_LIBCXX()    0
+#define _CCCL_HOST_STD_LIB_STL()       0
+
+// include a minimal header
+#if _CCCL_HAS_INCLUDE(<version>)
+#  include <version>
+#elif _CCCL_HAS_INCLUDE(<ciso646>)
+#  include <ciso646>
+#endif // ^^^ _CCCL_HAS_INCLUDE(<ciso646>) ^^^
+
+#if defined(_MSVC_STL_VERSION)
+#  undef _CCCL_HOST_STD_LIB_STL
+#  define _CCCL_HOST_STD_LIB_STL() 1
+#elif defined(__GLIBCXX__)
+#  undef _CCCL_HOST_STD_LIB_LIBSTDCXX
+#  define _CCCL_HOST_STD_LIB_LIBSTDCXX() 1
+#elif defined(_LIBCPP_VERSION)
+#  undef _CCCL_HOST_STD_LIB_LIBCXX
+#  define _CCCL_HOST_STD_LIB_LIBCXX() 1
+#endif // ^^^ _LIBCPP_VERSION ^^^
+
+#define _CCCL_HOST_STD_LIB(_X) _CCCL_HOST_STD_LIB_##_X()
+#define _CCCL_HAS_HOST_STD_LIB() \
+  (_CCCL_HOST_STD_LIB_LIBSTDCXX() || _CCCL_HOST_STD_LIB_LIBCXX() || _CCCL_HOST_STD_LIB_STL())
+
+#endif // __CCCL_HOST_STD_LIB_H

libcudacxx/include/cuda/std/__complex/complex.h

@@ -23,6 +23,7 @@
 
 #include <cuda/std/__complex/vector_support.h>
 #include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__fwd/complex.h>
 #include <cuda/std/__fwd/get.h>
 #include <cuda/std/__tuple_dir/tuple_element.h>
 #include <cuda/std/__tuple_dir/tuple_size.h>
@@ -36,9 +37,9 @@
 #include <cuda/std/limits>
 
 // Compatibility helpers for thrust to convert between `std::complex` and `cuda::std::complex`
+// todo: find a way to get rid of this include
 #if !_CCCL_COMPILER(NVRTC)
-#  include <complex>
-#  include <sstream> // for std::basic_ostringstream
+#  include <complex> // for std::complex stream operators
 
 #  define _LIBCUDACXX_ACCESS_STD_COMPLEX_REAL(__c) reinterpret_cast<const _Up(&)[2]>(__c)[0]
 #  define _LIBCUDACXX_ACCESS_STD_COMPLEX_IMAG(__c) reinterpret_cast<const _Up(&)[2]>(__c)[1]

libcudacxx/include/cuda/std/__complex/nvbf16.h

@@ -31,8 +31,9 @@
 #  include <cuda/std/__type_traits/enable_if.h>
 #  include <cuda/std/__type_traits/is_constructible.h>
 
+// todo: find a way to get rid of this include
 #  if !_CCCL_COMPILER(NVRTC)
-#    include <sstream> // for std::basic_ostringstream
+#    include <complex> // for std::complex stream operators
 #  endif // !_CCCL_COMPILER(NVRTC)
 
 #  include <cuda/std/__cccl/prologue.h>

libcudacxx/include/cuda/std/__complex/nvfp16.h

@@ -31,8 +31,9 @@
 #  include <cuda/std/__type_traits/enable_if.h>
 #  include <cuda/std/__type_traits/is_constructible.h>
 
+// todo: find a way to get rid of this include
 #  if !_CCCL_COMPILER(NVRTC)
-#    include <sstream> // for std::basic_ostringstream
+#    include <complex> // for std::complex stream operators
 #  endif // !_CCCL_COMPILER(NVRTC)
 
 #  include <cuda/std/__cccl/prologue.h>
@@ -294,7 +295,7 @@ struct __get_complex_impl<__half>
   }
 };
 
-#  if !defined(_LIBCUDACXX_HAS_NO_LOCALIZATION) && !_CCCL_COMPILER(NVRTC)
+#  if !_CCCL_COMPILER(NVRTC)
 template <class _CharT, class _Traits>
 ::std::basic_istream<_CharT, _Traits>& operator>>(::std::basic_istream<_CharT, _Traits>& __is, complex<__half>& __x)
 {
@@ -310,7 +311,7 @@ operator<<(::std::basic_ostream<_CharT, _Traits>& __os, const complex<__half>& _
 {
   return __os << complex<float>{__x};
 }
-#  endif // !_LIBCUDACXX_HAS_NO_LOCALIZATION && !_CCCL_COMPILER(NVRTC)
+#  endif // !_CCCL_COMPILER(NVRTC)
 
 _CCCL_END_NAMESPACE_CUDA_STD
 

libcudacxx/include/cuda/std/__fwd/allocator.h

@@ -23,6 +23,19 @@
 
 #include <cuda/std/__cccl/prologue.h>
 
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _Tp>
+class allocator;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
 template <class _Tp>

libcudacxx/include/cuda/std/__fwd/char_traits.h

@@ -23,6 +23,19 @@
 
 #include <cuda/std/__cccl/prologue.h>
 
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _CharT>
+struct char_traits;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
 template <class _CharT>

libcudacxx/include/cuda/std/__fwd/complex.h

@@ -20,12 +20,21 @@
 #  pragma system_header
 #endif // no system header
 
-#if !_CCCL_COMPILER(NVRTC)
-#  include <complex>
-#endif // !_CCCL_COMPILER(NVRTC)
-
 #include <cuda/std/__cccl/prologue.h>
 
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class>
+class complex;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
 template <class _Tp>

libcudacxx/include/cuda/std/__fwd/pair.h

@@ -22,6 +22,19 @@
 
 #include <cuda/std/__cccl/prologue.h>
 
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class, class>
+struct pair;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
 template <class, class>

libcudacxx/include/cuda/std/__fwd/string.h

@@ -26,6 +26,28 @@
 
 #include <cuda/std/__cccl/prologue.h>
 
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+// libstdc++ puts basic_string to inline cxx11 namespace
+#  if _GLIBCXX_USE_CXX11_ABI
+inline _GLIBCXX_BEGIN_NAMESPACE_CXX11
+#  endif // _GLIBCXX_USE_CXX11_ABI
+
+  template <class _CharT, class _Traits, class _Alloc>
+  class basic_string;
+
+#  if _GLIBCXX_USE_CXX11_ABI
+_GLIBCXX_END_NAMESPACE_CXX11
+#  endif // _GLIBCXX_USE_CXX11_ABI
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
 #if 0 // we don't support these features

libcudacxx/include/cuda/std/__fwd/string_view.h

@@ -24,6 +24,19 @@
 
 #include <cuda/std/__cccl/prologue.h>
 
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB() && __cpp_lib_string_view >= 201606L
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _CharT, class _Traits>
+class basic_string_view;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB() && __cpp_lib_string_view >= 201606L
+
+// cuda::std:: forward declarations
+
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
 template <class _CharT, class _Traits = char_traits<_CharT>>
@@ -52,6 +65,7 @@ _CCCL_PREFERRED_NAME(wstring_view)
 #endif // _CCCL_HAS_WCHAR_T()
       basic_string_view;
 // clang-format on
+
 _CCCL_END_NAMESPACE_CUDA_STD
 
 #include <cuda/std/__cccl/epilogue.h>