fix build failures with clang on x86, make sure compilation works on other Power...

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 9c58fd465f8c2920cfeb41f3482230705684a69e..3485e5f8d414dc117d1c4d3e37fecd4146d39be2 100644 (file)
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -151,15 +151,16 @@ else ()
     set(ICELAKE_FLAG "-march=icelake-server")
 endif ()
 
+if(ARCH_PPC64EL)
+    set(ARCH_FLAG mcpu)
+else()
+    set(ARCH_FLAG march)
+endif()
+
 # Detect best GNUCC_ARCH to tune for
 if (CMAKE_COMPILER_IS_GNUCC AND NOT CROSS_COMPILE)
     message(STATUS "gcc version ${CMAKE_C_COMPILER_VERSION}")
 
-    if(ARCH_PPC64EL)
-        set(ARCH_FLAG mcpu)
-    else()
-        set(ARCH_FLAG march)
-    endif()
     # If gcc doesn't recognise the host cpu, then mtune=native becomes
     # generic, which isn't very good in some cases. march=native looks at
     # cpuid info and then chooses the best microarch it can (and replaces
@@ -185,23 +186,12 @@ if (CMAKE_COMPILER_IS_GNUCC AND NOT CROSS_COMPILE)
         set(TUNE_FLAG native)
     else()
         set(TUNE_FLAG ${GNUCC_ARCH})
-        message(STATUS "gcc will tune for ${GNUCC_ARCH}")
+        message(STATUS "gcc will tune for ${GNUCC_ARCH}, ${TUNE_FLAG}")
     endif()
 elseif (CMAKE_COMPILER_IS_CLANG AND NOT CROSS_COMPILE)
-    message(STATUS "clang will tune for ${TUNE_FLAG}")
     if (ARCH_IA32 OR ARCH_X86_64)
         set(GNUCC_ARCH native)
         set(TUNE_FLAG generic)
-        if (BUILD_AVX512)
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SKYLAKE_FLAG}")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SKYLAKE_FLAG}")
-        elseif (BUILD_AVX2)
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx2")
-        else()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -msse4.2")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.2")
-        endif()
     elseif(ARCH_AARCH64)
        set(GNUCC_ARCH armv8)
        set(TUNE_FLAG generic)
@@ -212,11 +202,30 @@ elseif (CMAKE_COMPILER_IS_CLANG AND NOT CROSS_COMPILE)
        set(GNUCC_ARCH native)
        set(TUNE_FLAG generic)
     endif()
+    message(STATUS "clang will tune for ${GNUCC_ARCH}, ${TUNE_FLAG}")
 elseif (CROSS_COMPILE)
     set(GNUCC_ARCH generic)
     set(TUNE_FLAG generic)
 endif()
 
+if (ARCH_IA32 OR ARCH_X86_64)
+    if (NOT FAT_RUNTIME)
+        if (BUILD_AVX512)
+            set(ARCH_C_FLAGS "${SKYLAKE_FLAG}")
+            set(ARCH_CXX_FLAGS "${SKYLAKE_FLAG}")
+        elseif (BUILD_AVX2)
+            set(ARCH_C_FLAGS "-mavx2")
+            set(ARCH_CXX_FLAGS "-mavx2")
+        else()
+            set(ARCH_C_FLAGS "-msse4.2")
+            set(ARCH_CXX_FLAGS "-msse4.2")
+        endif()
+    else()
+       set(ARCH_C_FLAGS "-msse4.2")
+       set(ARCH_CXX_FLAGS "-msse4.2")
+    endif()
+endif()
+
 if (ARCH_AARCH64)
     if (BUILD_SVE2_BITPERM)
         set(GNUCC_ARCH "${GNUCC_ARCH}+sve2-bitperm")
@@ -227,23 +236,26 @@ if (ARCH_AARCH64)
     endif ()
 endif(ARCH_AARCH64)
 
-if (ARCH_IA32 OR ARCH_X86_64 OR ARCH_ARM32 OR ARCH_AARCH64)
-    if (NOT CMAKE_C_FLAGS MATCHES .*march.* AND NOT CMAKE_C_FLAGS MATCHES .*mtune.*)
-        set(ARCH_C_FLAGS "-march=${GNUCC_ARCH} -mtune=${TUNE_FLAG}")
-    endif()
-    if (NOT CMAKE_CXX_FLAGS MATCHES .*march.* AND NOT CMAKE_CXX_FLAGS MATCHES .*mtune.*)
-        set(ARCH_CXX_FLAGS "-march=${GNUCC_ARCH} -mtune=${TUNE_FLAG}")
-    endif()
-endif()
+set(ARCH_C_FLAGS "-${ARCH_FLAG}=${GNUCC_ARCH} -mtune=${TUNE_FLAG} ${ARCH_C_FLAGS}")
+set(ARCH_CXX_FLAGS "-${ARCH_FLAG}=${GNUCC_ARCH} -mtune=${TUNE_FLAG} ${ARCH_CXX_FLAGS}")
+
+#if (ARCH_IA32 OR ARCH_X86_64 OR ARCH_ARM32 OR ARCH_AARCH64)
+#    if (NOT CMAKE_C_FLAGS MATCHES .*march.* AND NOT CMAKE_C_FLAGS MATCHES .*mtune.*)
+#        set(ARCH_C_FLAGS "-march=${GNUCC_ARCH} -mtune=${TUNE_FLAG}")
+#    endif()
+#    if (NOT CMAKE_CXX_FLAGS MATCHES .*march.* AND NOT CMAKE_CXX_FLAGS MATCHES .*mtune.*)
+#        set(ARCH_CXX_FLAGS "-march=${GNUCC_ARCH} -mtune=${TUNE_FLAG}")
+#    endif()
+#endif()
     
-if(ARCH_PPC64EL)
-    if (NOT CMAKE_C_FLAGS MATCHES .*march.* AND NOT CMAKE_C_FLAGS MATCHES .*mtune.*)
-        set(ARCH_C_FLAGS "-mtune=${TUNE_FLAG}")
-    endif()
-    if (NOT CMAKE_CXX_FLAGS MATCHES .*march.* AND NOT CMAKE_CXX_FLAGS MATCHES .*mtune.*)
-        set(ARCH_CXX_FLAGS "-mtune=${TUNE_FLAG}")
-    endif()
-endif()
+#if(ARCH_PPC64EL)
+#    if (NOT CMAKE_C_FLAGS MATCHES .*march.* AND NOT CMAKE_C_FLAGS MATCHES .*mtune.*)
+#        set(ARCH_C_FLAGS "-mtune=${TUNE_FLAG}")
+#    endif()
+#    if (NOT CMAKE_CXX_FLAGS MATCHES .*march.* AND NOT CMAKE_CXX_FLAGS MATCHES .*mtune.*)
+#        set(ARCH_CXX_FLAGS "-mtune=${TUNE_FLAG}")
+#    endif()
+#endif()
 
 # compiler version checks TODO: test more compilers
 if (CMAKE_COMPILER_IS_GNUCXX)
@@ -306,7 +318,6 @@ if (NOT(ARCH_IA32 AND RELEASE_BUILD))
     set(EXTRA_CXX_FLAGS "${EXTRA_CXX_FLAGS} -fno-omit-frame-pointer")
 endif()
 
-
 CHECK_INCLUDE_FILES(unistd.h HAVE_UNISTD_H)
 if (ARCH_IA32 OR ARCH_X86_64)
   CHECK_INCLUDE_FILES(intrin.h HAVE_C_INTRIN_H)
@@ -474,13 +485,12 @@ if(CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
     set(FREEBSD true)
 endif(CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
 
+
 if (FAT_RUNTIME)
     if (NOT (ARCH_IA32 OR ARCH_X86_64))
         message(FATAL_ERROR "Fat runtime is not supported on non-Intel architectures")
     else()
         message(STATUS "Building runtime for multiple microarchitectures")
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
     endif()
 else()
     if (CROSS_COMPILE)
@@ -488,9 +498,9 @@ else()
     else()
         message(STATUS "Building for current host CPU: ${ARCH_C_FLAGS}")
     endif()
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${ARCH_C_FLAGS}")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${ARCH_CXX_FLAGS}")
 endif()
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${ARCH_C_FLAGS}")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${ARCH_CXX_FLAGS}")
 
 add_subdirectory(util)
 add_subdirectory(doc/dev-reference)
@@ -1207,10 +1217,6 @@ if (NOT FAT_RUNTIME)
         set_target_properties(hs_runtime PROPERTIES LINKER_LANGUAGE C)
 
         add_library(hs_compile OBJECT ${hs_compile_SRCS})
-        if (ARCH_IA32)
-            set_target_properties(hs_compile PROPERTIES COMPILE_FLAGS "-msse4.2")
-        endif (ARCH_IA32)
-
         add_library(hs STATIC
             src/hs_version.c
             src/hs_valid_platform.c
@@ -1241,7 +1247,7 @@ else (FAT_RUNTIME)
         add_library(hs_exec_core2 OBJECT ${hs_exec_SRCS})
         list(APPEND RUNTIME_LIBS $<TARGET_OBJECTS:hs_exec_core2>)
         set_target_properties(hs_exec_core2 PROPERTIES
-            COMPILE_FLAGS "-march=core2"
+            COMPILE_FLAGS "-march=core2 -msse4.2"
             RULE_LAUNCH_COMPILE "${BUILD_WRAPPER} core2 ${CMAKE_MODULE_PATH}/keep.syms.in"
             )
 
@@ -1290,10 +1296,6 @@ else (FAT_RUNTIME)
             ${RUNTIME_LIBS})
         set_target_properties(hs_runtime PROPERTIES LINKER_LANGUAGE C)
         add_library(hs_compile OBJECT ${hs_compile_SRCS})
-        if (ARCH_IA32 OR ARCH_X86_64)
-            set_target_properties(hs_exec_common PROPERTIES COMPILE_FLAGS "-msse4.2")
-            set_target_properties(hs_compile PROPERTIES COMPILE_FLAGS "-msse4.2")
-        endif ()
 
         # we want the static lib for testing
         add_library(hs STATIC src/hs_version.c src/hs_valid_platform.c
@@ -1310,7 +1312,7 @@ else (FAT_RUNTIME)
         add_library(hs_exec_shared_core2 OBJECT ${hs_exec_SRCS})
         list(APPEND RUNTIME_SHLIBS $<TARGET_OBJECTS:hs_exec_shared_core2>)
         set_target_properties(hs_exec_shared_core2 PROPERTIES
-            COMPILE_FLAGS "-march=core2"
+            COMPILE_FLAGS "-march=core2 -msse4.2"
             POSITION_INDEPENDENT_CODE TRUE
             RULE_LAUNCH_COMPILE "${BUILD_WRAPPER} core2 ${CMAKE_MODULE_PATH}/keep.syms.in"
             )

diff --git a/src/util/simd_types.h b/src/util/simd_types.h
index 0deff7e58af470f9c01e4829a7b91e7d6a982555..4f0fd1a98739c9b795eb3b70054d4a0c132def0d 100644 (file)
--- a/src/util/simd_types.h
+++ b/src/util/simd_types.h
@@ -51,6 +51,7 @@ typedef struct ALIGN_AVX_DIRECTIVE {m128 lo; m128 hi;} m256;
 #endif
 
 typedef struct {m128 lo; m128 mid; m128 hi;} m384;
+
 #if !defined(m512) && !defined(HAVE_SIMD_512_BITS)
 typedef struct ALIGN_ATTR(64) {m256 lo; m256 hi;} m512;
 #endif

diff --git a/util/CMakeLists.txt b/util/CMakeLists.txt
index 82cee0ffa62148f013956b54dbb375a0ab0053c6..ea942ef1a5ced1871d5029e31588f106d9cd434b 100644 (file)
--- a/util/CMakeLists.txt
+++ b/util/CMakeLists.txt
@@ -33,9 +33,6 @@ SET(corpusomatic_SRCS
     ng_find_matches.cpp
 )
 add_library(corpusomatic STATIC ${corpusomatic_SRCS})
-if (ARCH_IA32 OR ARCH_X86_64)
-    set_target_properties(corpusomatic PROPERTIES COMPILE_FLAGS "-mssse3")
-endif ()
 
 set(databaseutil_SRCS
     database_util.cpp