|
| 1 | +if (USE_CPU_NATIVE) |
| 2 | + # Detect best GNUCC_ARCH to tune for |
| 3 | + if (CMAKE_COMPILER_IS_GNUCC) |
| 4 | + message(STATUS "gcc version ${CMAKE_C_COMPILER_VERSION}") |
| 5 | + |
| 6 | + # If gcc doesn't recognise the host cpu, then mtune=native becomes |
| 7 | + # generic, which isn't very good in some cases. march=native looks at |
| 8 | + # cpuid info and then chooses the best microarch it can (and replaces |
| 9 | + # the flag), so use that for tune. |
| 10 | + |
| 11 | + set(TUNE_FLAG "mtune") |
| 12 | + set(GNUCC_TUNE "") |
| 13 | + message(STATUS "ARCH_FLAG '${ARCH_FLAG}' '${GNUCC_ARCH}', TUNE_FLAG '${TUNE_FLAG}' '${GNUCC_TUNE}' ") |
| 14 | + |
| 15 | + # arg1 might exist if using ccache |
| 16 | + string (STRIP "${CMAKE_C_COMPILER_ARG1}" CC_ARG1) |
| 17 | + set (EXEC_ARGS ${CC_ARG1} -c -Q --help=target -${ARCH_FLAG}=native -${TUNE_FLAG}=native) |
| 18 | + execute_process(COMMAND ${CMAKE_C_COMPILER} ${EXEC_ARGS} |
| 19 | + OUTPUT_VARIABLE _GCC_OUTPUT) |
| 20 | + set(_GCC_OUTPUT_TUNE ${_GCC_OUTPUT}) |
| 21 | + string(FIND "${_GCC_OUTPUT}" "${ARCH_FLAG}=" POS) |
| 22 | + string(SUBSTRING "${_GCC_OUTPUT}" ${POS} -1 _GCC_OUTPUT) |
| 23 | + string(REGEX REPLACE "${ARCH_FLAG}=[ \t]*([^ \n]*)[ \n].*" "\\1" GNUCC_ARCH "${_GCC_OUTPUT}") |
| 24 | + |
| 25 | + string(FIND "${_GCC_OUTPUT_TUNE}" "${TUNE_FLAG}=" POS_TUNE) |
| 26 | + string(SUBSTRING "${_GCC_OUTPUT_TUNE}" ${POS_TUNE} -1 _GCC_OUTPUT_TUNE) |
| 27 | + string(REGEX REPLACE "${TUNE_FLAG}=[ \t]*([^ \n]*)[ \n].*" "\\1" GNUCC_TUNE "${_GCC_OUTPUT_TUNE}") |
| 28 | + |
| 29 | + message(STATUS "ARCH_FLAG '${ARCH_FLAG}' '${GNUCC_ARCH}', TUNE_FLAG '${TUNE_FLAG}' '${GNUCC_TUNE}' ") |
| 30 | + |
| 31 | + # test the parsed flag |
| 32 | + set (EXEC_ARGS ${CC_ARG1} -E - -${ARCH_FLAG}=${GNUCC_ARCH} -${TUNE_FLAG}=${GNUCC_TUNE}) |
| 33 | + execute_process(COMMAND ${CMAKE_C_COMPILER} ${EXEC_ARGS} |
| 34 | + OUTPUT_QUIET ERROR_QUIET |
| 35 | + INPUT_FILE /dev/null |
| 36 | + RESULT_VARIABLE GNUCC_TUNE_TEST) |
| 37 | + |
| 38 | + if (NOT GNUCC_TUNE_TEST EQUAL 0) |
| 39 | + message(WARNING "Something went wrong determining gcc tune: -mtune=${GNUCC_TUNE} not valid, falling back to -mtune=native") |
| 40 | + set(GNUCC_TUNE native) |
| 41 | + else() |
| 42 | + set(GNUCC_TUNE ${GNUCC_TUNE}) |
| 43 | + message(STATUS "gcc will tune for ${GNUCC_ARCH}, ${GNUCC_TUNE}") |
| 44 | + endif() |
| 45 | + elseif (CMAKE_COMPILER_IS_CLANG) |
| 46 | + if (ARCH_IA32 OR ARCH_X86_64) |
| 47 | + set(GNUCC_ARCH x86_64_v2) |
| 48 | + set(TUNE_FLAG generic) |
| 49 | + elseif(ARCH_AARCH64) |
| 50 | + if (BUILD_SVE2_BITPERM) |
| 51 | + set(GNUCC_ARCH ${SVE2_BITPERM_ARCH}) |
| 52 | + elseif (BUILD_SVE2) |
| 53 | + set(GNUCC_ARCH ${SVE2_ARCH}) |
| 54 | + elseif (BUILD_SVE) |
| 55 | + set(GNUCC_ARCH ${SVE_ARCH}) |
| 56 | + else () |
| 57 | + set(GNUCC_ARCH ${ARMV8_ARCH}) |
| 58 | + endif() |
| 59 | + set(TUNE_FLAG generic) |
| 60 | + elseif(ARCH_ARM32) |
| 61 | + set(GNUCC_ARCH armv7a) |
| 62 | + set(TUNE_FLAG generic) |
| 63 | + else() |
| 64 | + set(GNUCC_ARCH native) |
| 65 | + set(TUNE_FLAG generic) |
| 66 | + endif() |
| 67 | + message(STATUS "clang will tune for ${GNUCC_ARCH}, ${TUNE_FLAG}") |
| 68 | + endif() |
| 69 | +else() |
| 70 | + if (ARCH_IA32 OR ARCH_X86_64) |
| 71 | + set(GNUCC_ARCH native) |
| 72 | + set(TUNE_FLAG generic) |
| 73 | + elseif(ARCH_AARCH64) |
| 74 | + if (BUILD_SVE2_BITPERM) |
| 75 | + set(GNUCC_ARCH ${SVE2_BITPERM_ARCH}) |
| 76 | + elseif (BUILD_SVE2) |
| 77 | + set(GNUCC_ARCH ${SVE2_ARCH}) |
| 78 | + elseif (BUILD_SVE) |
| 79 | + set(GNUCC_ARCH ${SVE_ARCH}) |
| 80 | + else () |
| 81 | + set(GNUCC_ARCH ${ARMV8_ARCH}) |
| 82 | + endif() |
| 83 | + set(TUNE_FLAG generic) |
| 84 | + elseif(ARCH_ARM32) |
| 85 | + set(GNUCC_ARCH armv7a) |
| 86 | + set(TUNE_FLAG generic) |
| 87 | + else() |
| 88 | + set(GNUCC_ARCH power9) |
| 89 | + set(TUNE_FLAG power9) |
| 90 | + endif() |
| 91 | +endif() |
0 commit comments