sdrangel/cmake/Modules/DetectArchitecture.cmake

include_guard(GLOBAL)

include(CheckCXXCompilerFlag)
include(CheckSymbolExists)
include(CMakePushCheckState)

# Detect current compilation architecture and create standard definitions
macro(detect_architecture symbol arch)
    if (NOT DEFINED ARCHITECTURE)
        check_symbol_exists("${symbol}" "" ARCHITECTURE_${arch})

        if (ARCHITECTURE_${arch})
            set(ARCHITECTURE ${arch})
            set(ARCHITECTURE_${arch} TRUE)
            add_compile_definitions(ARCHITECTURE_${arch})
        endif()
    endif()
endmacro()

macro(force_ext_available extension)
  message(STATUS "Looking for __${extension}__ - forced found")
  set(HAS_${extension} 1 CACHE INTERNAL "")
endmacro()

function(detect_extensions extension)
  unset(HAS_${extension})
  if (ARGC EQUAL 2 AND (${ARGV1})) # force available
    force_ext_available(${extension})
  endif()
  check_symbol_exists("__${extension}__" "" HAS_${extension})
  if (HAS_${extension})
    add_compile_definitions(USE_${extension})
  endif()
endfunction()

function(detect_msvc_native_opt)
  set(TEST_DIR ${PROJECT_SOURCE_DIR}/cmake/test)

  try_run(RUN_AVX512 COMPILE_AVX512 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx512.cxx" COMPILE_DEFINITIONS /arch:AVX512)
  if (COMPILE_AVX512 AND RUN_AVX512 EQUAL 0)
    set(ARCH_OPT "AVX512" PARENT_SCOPE)
    return()
  endif()
  try_run(RUN_AVX2 COMPILE_AVX2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx2.cxx" COMPILE_DEFINITIONS /arch:AVX2)
  if (COMPILE_AVX2 AND RUN_AVX2 EQUAL 0)
    set(ARCH_OPT "AVX2" PARENT_SCOPE)
    return()
  endif()
  try_run(RUN_AVX COMPILE_AVX "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_avx.cxx" COMPILE_DEFINITIONS /arch:AVX)
  if (COMPILE_AVX AND RUN_AVX EQUAL 0)
    set(ARCH_OPT "AVX" PARENT_SCOPE)
    return()
  endif()

  # Supporting 32-bit x86, what year is it?
  set(COMPILE_DEF "")
  set(ARCH_OPT "" PARENT_SCOPE)
  if (ARCHITECTURE_x86)
    set(COMPILE_DEF "/arch:SSE2")
    set(ARCH_OPT "SSE2" PARENT_SCOPE)
  endif()

  try_run(RUN_SSE42 COMPILE_SSE42 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse42.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
  if (COMPILE_SSE42 AND RUN_SSE42 EQUAL 0)
    force_ext_available(SSE4_2)
    return()
  endif()
  try_run(RUN_SSE41 COMPILE_SSE41 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse41.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
  if (COMPILE_SSE41 AND RUN_SSE41 EQUAL 0)
    force_ext_available(SSE4_1)
    return()
  endif()
  try_run(RUN_SSSE3 COMPILE_SSSE3 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_ssse3.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
  if (COMPILE_SSSE3 AND RUN_SSSE3 EQUAL 0)
    force_ext_available(SSSE3)
    return()
  endif()
  try_run(RUN_SSE3 COMPILE_SSE3 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse3.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
  if (COMPILE_SSE3 AND RUN_SSE3 EQUAL 0)
    force_ext_available(SSE3)
    return()
  endif()
  try_run(RUN_SSE2 COMPILE_SSE2 "${CMAKE_BINARY_DIR}/tmp" "${TEST_DIR}/test_x86_sse2.cxx" COMPILE_DEFINITIONS ${COMPILE_DEF})
  if (COMPILE_SSE2 AND RUN_SSE2 EQUAL 0)
    force_ext_available(SSE2)
    return()
  endif()

  if (ARCHITECTURE_x86)
    # At this point we might as well...
    set(ARCH_OPT "IA32" PARENT_SCOPE)
    return()
  endif()
endfunction()

if (MSVC)
    detect_architecture("_M_AMD64" x86_64)
    detect_architecture("_M_IX86" x86)
    detect_architecture("_M_ARM" ARM)
    detect_architecture("_M_ARM64" ARM64)
else()
    detect_architecture("__x86_64__" x86_64)
    detect_architecture("__i386__" x86)
    detect_architecture("__arm__" ARM)
    detect_architecture("__aarch64__" ARM64)
endif()
if (NOT DEFINED ARCHITECTURE)
  message(FATAL_ERROR "Not supported. Please add needed architecture detection.")
endif()

# Note: On x86 MSVC's /arch:SSE2 enables all SSE intrinsics support and is default option.
#       On x86_64 MSVC's SSE is supported and enabled, so only AVX selection is needed.

if (FORCE_SSSE3)
  message(WARNING "FORCE_SSSE3 flag is deprecated, please use ARCH_OPT option.")
  set(ARCH_OPT "")
  if (MSVC)
    if (ARCHITECTURE_x86)
      set(ARCH_OPT "SSE2")
    endif()
    force_ext_available(SSSE3)
  else()
    set(FORCE_OPT "-mssse3")
  endif()
endif()

if (FORCE_SSE41)
  message(WARNING "FORCE_SSE41 flag is deprecated, please use ARCH_OPT option.")
  set(ARCH_OPT "")
  if (MSVC)
    if (ARCHITECTURE_x86)
      set(ARCH_OPT "SSE2")
    else()
      force_ext_available(SSE4_1)
    endif()
  else()
    set(FORCE_OPT "-msse4.1")
  endif()
endif()

if (MSVC)
  # Glue to make ARCH_OPT more flexible for MSVC
  if (ARCH_OPT STREQUAL "native")
    # Some compilers simulating MSVC supports the march flag, so use it if we can
    check_cxx_compiler_flag("-march=native" MARCH_NATIVE_SUPPORTED)
    if (NOT MARCH_NATIVE_SUPPORTED)
      detect_msvc_native_opt()
      FILE(REMOVE_RECURSE ${CMAKE_BINARY_DIR}/tmp)
    endif()
  elseif(ARCH_OPT STREQUAL "SSE4_2")
    force_ext_available(SSE4_2)
    set(ARCH_OPT "")
  elseif(ARCH_OPT STREQUAL "SSE4_1")
    force_ext_available(SSE4_1)
    set(ARCH_OPT "")
  elseif(ARCH_OPT STREQUAL "SSSE3")
    force_ext_available(SSSE3)
    set(ARCH_OPT "")
    elseif(ARCH_OPT STREQUAL "SSE3")
    force_ext_available(SSE3)
    set(ARCH_OPT "")
  elseif(ARCH_OPT STREQUAL "SSE2")
    force_ext_available(SSE2)
    set(ARCH_OPT "")
  endif()
endif()

message(STATUS "Target architecture: ${ARCHITECTURE}-${ARCH_OPT}")

cmake_push_check_state(RESET)
if (ARCH_OPT)
  if(MSVC AND NOT MARCH_NATIVE_SUPPORTED)
    set(CMAKE_REQUIRED_FLAGS "/arch:${ARCH_OPT}")
    add_compile_options(${CMAKE_REQUIRED_FLAGS})
  else()
    set(CMAKE_REQUIRED_FLAGS "-march=${ARCH_OPT}")
    add_compile_options(-march=${ARCH_OPT})
  endif()
elseif(FORCE_SSSE3 OR FORCE_SSE41)
  if (NOT MSVC)
    set(CMAKE_REQUIRED_FLAGS ${FORCE_OPT})
    add_compile_options(${FORCE_OPT})
  endif()
endif()

check_cxx_compiler_flag("${CMAKE_REQUIRED_FLAGS}" FLAG_SUPPORTED)
if (NOT FLAG_SUPPORTED)
  message(FATAL_ERROR "Flag '${CMAKE_REQUIRED_FLAGS}' rejected by compiler. Please adjust ARCH_OPT option.")
endif()

if (ARCHITECTURE_ARM)
  if (MSVC)
    force_ext_available(ARM_NEON)
  else()
    list(APPEND CMAKE_REQUIRED_FLAGS -mfpu=neon)
  endif()
endif()

# Extensions detection for ARM
check_symbol_exists("__ARM_NEON" "" HAS_NEON)
if (HAS_NEON)
  message(STATUS "ARM Neon extensions enabled")
  add_compile_definitions(USE_NEON)
endif()

# This is quite basic detection, can be extended if needed
detect_extensions(AVX512F)
detect_extensions(AVX2 HAS_AVX512F)
detect_extensions(AVX HAS_AVX2)
detect_extensions(SSE4_2 HAS_AVX)
detect_extensions(SSE4_1 HAS_SSE4_2)
detect_extensions(SSSE3 HAS_SSE4_1)
detect_extensions(SSE3 HAS_SSSE3)
detect_extensions(SSE2 HAS_SSE3)

cmake_pop_check_state()