AMReX_Extension.H

Go to the documentation of this file.

#ifndef AMREX_EXTENSION_H_
#define AMREX_EXTENSION_H_
#include <AMReX_Config.H>
 
#if defined(__INTEL_COMPILER) || (defined(__INTEL_LLVM_COMPILER) && !defined(SYCL_LANGUAGE_VERSION))
#define AMREX_CXX_INTEL
#elif defined(_CRAYC) || defined(__cray__)
#define AMREX_CXX_CRAY
#elif defined(__PGI)
#define AMREX_CXX_PGI
#elif defined(__NVCOMPILER)
#define AMREX_CXX_NVHPC
#elif defined(__NEC__)
#define AMREX_CXX_NEC
#elif defined(__ibmxl__)
#define AMREX_CXX_IBM
#elif defined(__clang__)
#define AMREX_CXX_CLANG
#elif defined(__GNUC__)
#define AMREX_CXX_GCC
#endif
 
#if !defined(BL_LANG_FORT)
 
// restrict
 
#ifdef __cplusplus
 
#ifdef _WIN32
#define AMREX_RESTRICT __restrict
#else
#define AMREX_RESTRICT __restrict__
#endif
 
#else
 
#define AMREX_RESTRICT restrict
 
#endif
 
// simd
 
#if defined(AMREX_DEBUG)
#define AMREX_PRAGMA_SIMD
 
#elif defined(__CUDA_ARCH__) && defined(AMREX_USE_CUDA)
#define AMREX_PRAGMA_SIMD
 
#elif defined(__HIP_DEVICE_COMPILE__) && defined(AMREX_USE_HIP)
#define AMREX_PRAGMA_SIMD
 
//#elif defined(AMREX_USE_OMP) && defined(_OPENMP) && (_OPENMP >= 201307) && !defined(__PGI)
//#define AMREX_PRAGMA_SIMD _Pragma("omp simd")
 
#elif defined(AMREX_CXX_INTEL)
#define AMREX_PRAGMA_SIMD _Pragma("ivdep")
 
#elif defined(AMREX_CXX_CRAY)
#define AMREX_PRAGMA_SIMD _Pragma("ivdep")
 
#elif defined(AMREX_CXX_PGI)
#define AMREX_PRAGMA_SIMD _Pragma("loop ivdep")
 
#elif defined(AMREX_CXX_NVHPC)
#define AMREX_PRAGMA_SIMD _Pragma("loop ivdep")
 
#elif defined(AMREX_CXX_NEC)
#define AMREX_PRAGMA_SIMD
 
#elif defined(AMREX_CXX_IBM)
#define AMREX_PRAGMA_SIMD _Pragma("ibm independent_loop")
 
#elif defined(__clang__)
#define AMREX_PRAGMA_SIMD
 
#elif defined(__GNUC__)
#define AMREX_PRAGMA_SIMD _Pragma("GCC ivdep")
 
#else
#define AMREX_PRAGMA_SIMD
 
#endif /* simd */
 
// force inline
#if defined(__CUDA_ARCH__) && defined(AMREX_USE_CUDA)
#define AMREX_FORCE_INLINE __forceinline__
 
#elif defined(__HIP_DEVICE_COMPILE__) && defined(AMREX_USE_HIP)
#define AMREX_FORCE_INLINE __forceinline__
 
#elif defined(AMREX_CXX_INTEL)
#define AMREX_FORCE_INLINE inline __attribute__((always_inline))
 
#elif defined(AMREX_CXX_CRAY)
#define AMREX_FORCE_INLINE inline
 
#elif defined(AMREX_CXX_PGI)
#define AMREX_FORCE_INLINE inline
 
#elif defined(AMREX_CXX_NVHPC)
#define AMREX_FORCE_INLINE inline
 
#elif defined(AMREX_CXX_NEC)
#define AMREX_FORCE_INLINE inline
 
#elif defined(AMREX_CXX_IBM)
#define AMREX_FORCE_INLINE inline __attribute__((always_inline))
 
#elif defined(__clang__)
#define AMREX_FORCE_INLINE inline __attribute__((always_inline))
 
#elif defined(__GNUC__)
#define AMREX_FORCE_INLINE inline __attribute__((always_inline))
 
#elif defined(_MSC_VER)
#define AMREX_FORCE_INLINE inline __forceinline
 
#else
#define AMREX_FORCE_INLINE inline
 
#endif /* force inline */
 
 
#ifdef AMREX_USE_FORCE_INLINE
#define AMREX_INLINE AMREX_FORCE_INLINE
#else
#define AMREX_INLINE inline
#endif
 
// no inline
#if defined(_MSC_VER)
#define AMREX_NO_INLINE __declspec(noinline)
#elif (defined(__GNUC__) || defined(__clang__) || defined(__CUDACC__) || defined(__HIP__) || defined(__INTEL_CLANG_COMPILER))
#define AMREX_NO_INLINE __attribute__((noinline))
#else
#define AMREX_NO_INLINE
#endif
 
// flatten
#if defined(_MSC_VER)
#define AMREX_FLATTEN [[msvc::flatten]]
#elif defined(__clang__) || defined(__GNUC__)
#define AMREX_FLATTEN __attribute__((flatten))
#else
#define AMREX_FLATTEN
#endif
 
#ifdef AMREX_USE_FLATTEN_FOR
#define AMREX_ATTRIBUTE_FLATTEN_FOR AMREX_FLATTEN
#else
#define AMREX_ATTRIBUTE_FLATTEN_FOR
#endif
 
// unroll loop
#define AMREX_TO_STRING_HELPER(X)   #X
#define AMREX_TO_STRING(X)          AMREX_TO_STRING_HELPER(X)
 
#if defined(__clang__) || defined(__CUDACC__) || defined(__HIP__) || defined(__INTEL_CLANG_COMPILER)
#define AMREX_UNROLL_LOOP(n)  _Pragma(AMREX_TO_STRING(unroll n))
#elif defined(__GNUC__)
#define AMREX_UNROLL_LOOP(n)  _Pragma(AMREX_TO_STRING(GCC unroll n))
#else
#define AMREX_UNROLL_LOOP(n)
#endif
 
// __COUNTER__
//
// __COUNTER__ has been a de-facto extension in all compilers AMReX supports for
// a very long time, but it was only recently adopted into C2y.  Clang 22 and
// later therefore diagnose it as a C2y extension under -Wpedantic, even in C++
// (https://github.com/llvm/llvm-project/issues/196557).  Bracket a use of
// __COUNTER__ with these macros to silence that diagnostic.  The nested
// -Wunknown-warning-option suppression keeps older Clang, which does not know
// -Wc2y-extensions, quiet as well.
//
// Each AMREX_WARNING_PUSH_* macro silences a specific diagnostic, while the
// single AMREX_WARNING_POP restores the previous state of whatever was pushed
// last.  The naming is intentionally asymmetric: the pops are not tied to a
// particular warning, so pushes and pops must nest, not interleave.
 
#if defined(__clang__)
#define AMREX_WARNING_PUSH_COUNTER                                        \
    _Pragma("clang diagnostic push")                                      \
    _Pragma("clang diagnostic ignored \"-Wunknown-warning-option\"")      \
    _Pragma("clang diagnostic ignored \"-Wc2y-extensions\"")
#define AMREX_WARNING_POP                                                 \
    _Pragma("clang diagnostic pop")
#else
#define AMREX_WARNING_PUSH_COUNTER
#define AMREX_WARNING_POP
#endif
 
// __attribute__((weak))
 
#if defined(AMREX_TYPECHECK)
#define AMREX_ATTRIBUTE_WEAK
#elif defined(_WIN32)
#define AMREX_ATTRIBUTE_WEAK
#elif defined(__clang__) && defined(__apple_build_version__)
#define AMREX_ATTRIBUTE_WEAK __attribute__((weak_import))
#else
#define AMREX_ATTRIBUTE_WEAK __attribute__((weak))
#endif
 
// public globals
// https://stackoverflow.com/questions/54560832/cmake-windows-export-all-symbols-does-not-cover-global-variables/54568678#54568678
#if defined(_MSC_VER)
#   if defined(AMREX_IS_DLL)
#       if defined(AMREX_IS_DLL_BUILDING)
#           define AMREX_EXPORT __declspec(dllexport)
#       else
#           define AMREX_EXPORT __declspec(dllimport)
#       endif
#   else
#       define AMREX_EXPORT
#   endif
#else
#   define AMREX_EXPORT
#endif
 
#if defined(__cplusplus) && defined(_WIN32)
#include <ciso646>
#endif
 
#if defined(__INTEL_COMPILER) && defined(__EDG__) && (__cplusplus < 201703L)
// Classical EDG based Intel compiler does not support fallthrough when std=c++14
#    define AMREX_FALLTHROUGH ((void)0)
#elif defined(__has_cpp_attribute) && __has_cpp_attribute(fallthrough) >= 201603L
#    define AMREX_FALLTHROUGH [[fallthrough]]
#elif defined(__clang__)
#    define AMREX_FALLTHROUGH [[clang::fallthrough]]
#elif defined(__GNUC__) && (__GNUC__ >= 7)
#    define AMREX_FALLTHROUGH [[gnu::fallthrough]]
#else
#    define AMREX_FALLTHROUGH ((void)0)
#endif
 
// Note: following compilers support [[likely]] and [[unlikely]]
//   - Clang >= 12.0
//   - GCC >= 9.0
//   - Intel >= 2021.7
//   - MSVC >= 19.26
//   - nvcc >= 12
#if defined(__has_cpp_attribute) && __has_cpp_attribute(likely) >= 201803L
#    define AMREX_LIKELY [[likely]]
#    define AMREX_UNLIKELY [[unlikely]]
#else
#    define AMREX_LIKELY
#    define AMREX_UNLIKELY
#endif
 
// Note: following compilers support assumptions, at least using builtin functions:
//   - Clang >= 3.7
//   - GCC >= 5.1
//   - MSVC >= 19.20
//   - nvcc >= 11.1.0
//   - icx >= 2021.1.2
//
// [[assume]] itself is C++23.  Clang >= 19 and GCC >= 13 advertise it through
// __has_cpp_attribute in older language modes as well, and Clang then diagnoses
// every use as a C++23 extension under -Wpedantic
// (-Wc++23-attribute-extensions).  So only take the attribute when the language
// mode really is C++23 and rely on the compiler-specific spellings below
// otherwise; they carry the same information to the optimizer.
#if defined(__has_cpp_attribute) && __has_cpp_attribute(assume) >= 202207L && \
    defined(__cplusplus) && __cplusplus >= 202302L
#   define AMREX_ASSUME(ASSUMPTION) [[assume(ASSUMPTION)]]
#else
#   if defined(__CUDA_ARCH__) && defined(__CUDACC__)
#       define AMREX_ASSUME(ASSUMPTION) __builtin_assume(ASSUMPTION)
#   elif defined(AMREX_CXX_INTEL) || defined(__clang__)
#       define AMREX_ASSUME(ASSUMPTION) __builtin_assume(ASSUMPTION)
#   elif defined(_MSC_VER)
#       define AMREX_ASSUME(ASSUMPTION) __assume(ASSUMPTION)
#   elif defined(__GNUC__) && __GNUC__ >= 13
        // like [[assume]], this does not evaluate ASSUMPTION
#       define AMREX_ASSUME(ASSUMPTION) __attribute__((assume(ASSUMPTION)))
#   elif defined(__GNUC__)
#       define AMREX_ASSUME(ASSUMPTION) \
            do { if (!(ASSUMPTION)) { __builtin_unreachable(); } } while (0)
#   else
#       define AMREX_ASSUME(ASSUMPTION)
#   endif
#endif
 
// CI uses -Werror -Wc++17-extension, thus we need to add the __cplusplus clause
#if !defined(AMREX_NO_NODISCARD) && defined(__has_cpp_attribute) && __has_cpp_attribute(nodiscard) >= 201603L
#   define AMREX_NODISCARD [[nodiscard]]
#else
#   define AMREX_NODISCARD
#endif
 
// Note: following compilers support [[no_unique_address]]
//   - Clang >= 9.0
//   - GCC >= 9.0
//   - MSVC >= 19.26
// Using no unique address makes empty base class optimization for multiple policies much easier
#if !defined(AMREX_NO_NO_UNIQUE_ADDRESS) && defined(__has_cpp_attribute) && __has_cpp_attribute(no_unique_address) >= 201803L
#   define AMREX_NO_UNIQUE_ADDRESS [[no_unique_address]]
#   define AMREX_HAS_NO_UNIQUE_ADDRESS 1
#else
#   define AMREX_NO_UNIQUE_ADDRESS
#endif
 
#if defined(__cpp_if_constexpr) && __cpp_if_constexpr >= 201606L
#   define AMREX_IF_CONSTEXPR if constexpr
#else
#   define AMREX_IF_CONSTEXPR if
#endif
 
#endif /* !BL_LANG_FORT */
 
#endif