cmake/nc_optimisation.cmake

###############################################################################
if (__nc_optimisation)
    return()
endif ()

set (__nc_optimisation TRUE)


###############################################################################
include (compile_flag)
include (link_flag)
include (canonical_host)


###############################################################################
option(LTO "enable link-time optimisation" OFF)


##-----------------------------------------------------------------------------
## Enable link-time optimisation.
##
## The INTERPROCEDURAL_OPTIMISATION flag is silently ignored for almost all
## compilers. Fuck you too CMake. We have to make up for its deficiencies
## ourselves.
if (LTO)
    # Add the linker flags first otherwise the linker may not recognise the
    # object format
    append_link_flag("-fuse-linker-plugin")
    append_link_flag("-flto")

    # Actually enable LTO object creation, but try very hard to avoid
    # situations where we may accidentally use regular/fat objects.
    append_compile_flag("-flto")
    append_compile_flag("-fno-fat-lto-objects")

    # Throw in some optimisation flags that are LTO specific. We don't
    # particularly care about checking Debug/Release here because LTO is
    # pretty heavyweight anyway.
    append_compile_flag("-fdevirtualize-at-ltrans")

    # If we're using GCC we probably need to use gcc-{ar,nm,ranlib} so that
    # plugin support works. Ideally we'd actually do some compilation tests,
    # but... it's a royal PITA with CMake.
    if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
        set (CMAKE_AR "gcc-ar")
        set (CMAKE_NM "gcc-nm")
        set (CMAKE_RANLIB "gcc-ranlib")
    endif ()
endif ()


###############################################################################
option(FRAMEPOINTER "retain the framepointer even if optimising" OFF)


##-----------------------------------------------------------------------------
if (FRAMEPOINTER)
    append_compile_flag("-fno-omit-frame-pointer")
endif ()


###############################################################################
canonical_host()


##-----------------------------------------------------------------------------
if (${host_cpu} STREQUAL "x86_64")
    append_compile_flag(-mtune=generic)
    append_compile_flag(-msse)
    append_compile_flag(-msse2)
    append_compile_flag(-msahf)
elseif (${host_cpu} STREQUAL "i686")
    append_compile_flag(-march=prescott)
    append_compile_flag(-mtune=generic)
    append_compile_flag(-mcmov)
    append_compile_flag(-mfpmath=sse)
else ()
    message (FATAL_ERROR "unknown HOST_CPU '${host_cpu}'")
endif ()


###############################################################################
if (CMAKE_BUILD_TYPE MATCHES Debug)
else()
    append_compile_flag(-ftree-vectorize)

    # clang vectorisation
    append_compile_flag(-fvectorize)
    append_compile_flag(-fslp-vectorize)
    append_compile_flag(-fslp-vectorize-aggressive)

    # loop hosting/distribution
    append_compile_flag(-floop-nest-optimize)

    append_compile_flag(-ftree-loop-distribution)
    append_compile_flag(-ftree-loop-distribute-patterns)
    append_compile_flag(-ftree-loop-im)
    append_compile_flag(-ftree-loop-if-convert-stores)

    append_compile_flag(-fivopts)

    append_compile_flag(-funsafe-loop-optimizations)
    append_compile_flag(-floop-interchange)

    append_compile_flag(-fpredictive-commoning)
    append_compile_flag(-funswitch-loops)

    # options that require substantial time/memory
    # GCC: ipa-pta can produce broken code when using LTO
    #AS_IF([test "x$ax_cv_cxx_compiler_vendor" != "xgnu"   ||
    #       test "x$enable_lto"                != "xyes"], [
    #    AX_APPEND_COMPILE_FLAGS([-fipa-pta])
    #])

    # safety removal for performance
    append_compile_flag(-fno-stack-protector)
endif()


###############################################################################
if (CMAKE_BUILD_TYPE MATCHES Debug)
else ()
    append_compile_flag(-fdevirtualize)
    append_compile_flag(-fdevirtualize-speculatively)

    check_link_flag(TEST_GC_SECTIONS, "-Wl,--gc-sections")
    if (TEST_LD_GC_SECTIONS)
        append_compile_flag(-fdata-sections)
        append_compile_flag(-ffunction-sections)
        append_link_flag("-Wl,--gc-sections")
    endif ()
endif ()


###############################################################################
if (CMAKE_BUILD_TYPE MATCHES Debug)
    add_definitions(-DENABLE_DEBUGGING)
    add_definitions(-D_GLIBCXX_DEBUG)

    append_compile_flag(-Og)
    append_compile_flag(-Werror)
    # this was protected for xmingw32, find out why
    append_compile_flag(-fstack-protector)
else ()
    append_compile_flag(-O2)
    append_compile_flag(-fno-rtti)
    add_definitions(-DNO_RTTI)
    add_definitions(-DNDEBUG)
endif ()


###############################################################################
append_compile_flag(-g)
append_compile_flag(-ggdb)