cmake/nc_optimisation.cmake

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

set (__nc_optimisation TRUE)


###############################################################################
include ("${CMAKE_CURRENT_LIST_DIR}/compile_flag.cmake")
include ("${CMAKE_CURRENT_LIST_DIR}/link_flag.cmake")
include ("${CMAKE_CURRENT_LIST_DIR}/canonical_host.cmake")


###############################################################################
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")
    # HACK: we shouldn't use 'auto' as it will likely lead to oversubscription
    #   but I'm sick of holding GCC's hand today.
    append_first_link_flag("-flto=auto" "-flto")

    # Enable LTO on the compilation side, but try very hard to avoid
    # situations where we may accidentally use regular/fat objects.
    #
    # GCC 12 will emit a warning about serial compilation of LTRANS jobs if an
    #   explicit parameter is not passed. It must be an integer, auto, or
    #   jobserver.
    # Clang will only accept 'full', 'thin', or no argument.
    #
    # HACK: we shouldn't use 'auto' as it will likely lead to oversubscription
    #   but I'm sick of holding GCC's hand today.
    append_first_compile_flag("-flto=auto" "-flto")
    append_compile_flag("-fno-fat-lto-objects")
    append_compile_flag("-flto-jobs=0")

    # Try to squeeze out some more diagnostics via LTO
    append_compile_flag("-flto-odr-type-merging")

    # GCC: Attempt to use a more aggressive whole-program style of LTO
    # NOTE: It's unclear if this _actually_ buys us performance or if it just
    #       slows down linking (by a substantial factor).
    #append_compile_flag("-flto-partition=none")

    # 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")
    # Explicitly qualify the CPU as the minimum viable architecture, but tune
    # for common platforms. This might prevent some invalid instructions being
    # emitted.
    append_compile_flag(-march=x86-64)
    append_compile_flag(-mtune=generic)

    append_compile_flag(-msse)
    append_compile_flag(-msse2)
    append_compile_flag(-mssse3)
    append_compile_flag(-msse3)
    append_compile_flag(-mavx)

    append_compile_flag(-mno-avx2)

    # These instructions should always be available on reasonable platforms.
    # They're almost 20 years old at this point.
    #append_compile_flag(-msse)
    #append_compile_flag(-msse2)
    # append_compile_flag(-mssse3)
    # 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 (WARNING "Unknown architecture. Not attempting performance options")
endif ()


###############################################################################
if (CMAKE_BUILD_TYPE MATCHES Debug)
else()
    # unsafe maths
    # append_compile_flag(-funsafe-math-optimizations)
    # append_compile_flag(-ffast-math)
    # append_compile_flag(-fno-finite-math-only)

    # vectorisation
    append_compile_flag(-ftree-vectorize)
    append_compile_flag(-fvectorize)
    append_compile_flag(-fslp-vectorize)

    # 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)

    # GCC 8.2 encounters an ICE in LTO linking with ipa-pta enabled
    append_compile_flag (-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)
    set (_DEFAULT_RUNTIME_DEBUGGING ON)
else()
    set (_DEFAULT_RUNTIME_DEBUGGING OFF)
endif()

#message(FATAL_ERROR "RUNTIME DEBUGGING ${_DEFAULT_RUNTIME_DEBUGGING}")
option(
    RUNTIME_DEBUGGING
    "enable debugging features that impact runtime and ABI"
    ${_DEFAULT_RUNTIME_DEBUGGING}
)

if (RUNTIME_DEBUGGING)
    if (CMAKE_BUILD_TYPE MATCHES Debug)
    else()
        message(FATAL_ERROR "RUNTIME_DEBUGGING is only supported under a DEBUG build")
    endif()
endif ()


if (CMAKE_BUILD_TYPE MATCHES Debug)
    add_definitions(-DENABLE_DEBUGGING)

    if (RUNTIME_DEBUGGING)
        if ("${STDLIB}" STREQUAL "libstdc++")
            add_definitions(-D_GLIBCXX_DEBUG)
        endif()
    endif()

    # Don't make any commits that use -O0 by default. Instead, either add an
    # argument that allows one to choose, or temporarily switch the lines
    # below locally. -O0 is just too slow for computationally heavy projects.
    #append_compile_flag(-Og)
    append_compile_flag(-O0)

    append_compile_flag(-Werror)

    # stack protection tends to cause segfaults in the generated binary
    # occuring in the standard library under msys2 and related compilers.
    #
    # it's easily shown by simply creating a local std::string. rather than
    # get to the root cause we just disable the protection under windows.
    if (NOT WIN32)
        append_compile_flag(-fstack-protector)
    endif()
else ()
    append_compile_flag(-O2)
    append_compile_flag(-fno-rtti)
    add_definitions(-DNO_RTTI)
    add_definitions(-DNDEBUG)
endif ()


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

# gcc#: -ggdb tends to trigger an ICE in 'trunc_int_for_mode' under GCC with
# libcruft-util's coord code. It's not strictly necessary so disable it for
# now.
#append_compile_flag(-ggdb)