Move x86 CPUID code from cpuid.hpp to cpuid.cpp (#73)

kimwalisch · Jun 22, 2024 · ae31111 · ae31111
1 parent f4f3c91
commit ae31111
Show file tree

Hide file tree

Showing 18 changed files with 268 additions and 246 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -91,7 +91,7 @@ if(NOT isMultiConfig AND NOT CMAKE_BUILD_TYPE)
 endif()
 
 if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-    set(ENABLE_ASSERT "ENABLE_ASSERT")
+    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "ENABLE_ASSERT")
 endif()
 
 # primecount binary source files #####################################
@@ -154,11 +154,26 @@ else()
     set(LIB_SRC ${LIB_SRC} src/gourdon/AC.cpp)
 endif()
 
+# Check if compiler supports CPU multiarch ###########################
+
+if(WITH_MULTIARCH)
+    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_x86_popcnt.cmake")
+    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_avx512_vpopcnt.cmake")
+
+    if(multiarch_x86_popcnt OR multiarch_avx512_vpopcnt)
+        set(LIB_SRC ${LIB_SRC} src/x86/cpuid.cpp)
+    endif()
+
+    if(NOT multiarch_avx512_vpopcnt)
+        include("${PROJECT_SOURCE_DIR}/cmake/multiarch_arm_sve.cmake")
+    endif()
+endif()
+
 # Enable __float128 support (requires libquadmath) ###################
 
 if(WITH_FLOAT128)
-    set(LIB_QUADMATH "quadmath")
-    set(HAVE_FLOAT128 "HAVE_FLOAT128")
+    list(APPEND PRIMECOUNT_LINK_LIBRARIES "quadmath")
+    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "HAVE_FLOAT128")
 endif()
 
 # Use 32-bit integer division ########################################
@@ -168,7 +183,7 @@ endif()
 # division. On most CPUs before 2020 this significantly
 # improves performance.
 if(WITH_DIV32)
-    set(ENABLE_DIV32 "ENABLE_DIV32")
+    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "ENABLE_DIV32")
 endif()
 
 # Use -Wno-uninitialized with GCC compiler ###########################
@@ -188,15 +203,6 @@ include("${PROJECT_SOURCE_DIR}/cmake/compiler_supports_cpp11.cmake")
 
 include("${PROJECT_SOURCE_DIR}/cmake/int128_t.cmake")
 
-# Check if compiler supports x64 multiarch ###########################
-
-if(WITH_MULTIARCH)
-    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_avx512_vpopcnt.cmake")
-    if(NOT multiarch_avx512_vpopcnt)
-        include("${PROJECT_SOURCE_DIR}/cmake/multiarch_arm_sve.cmake")
-    endif()
-endif()
-
 # Check for OpenMP ###################################################
 
 if(WITH_OPENMP)
@@ -246,8 +252,8 @@ if(BUILD_SHARED_LIBS)
     set_target_properties(libprimecount PROPERTIES SOVERSION ${PRIMECOUNT_VERSION_MAJOR})
     set_target_properties(libprimecount PROPERTIES VERSION ${PRIMECOUNT_VERSION})
     target_compile_options(libprimecount PRIVATE "${WNO_UNINITIALIZED}")
-    target_compile_definitions(libprimecount PRIVATE "${HAVE_FLOAT128}" "${DISABLE_INT128}" "${ENABLE_DIV32}" "${ENABLE_ASSERT}" "${ENABLE_MULTIARCH}" "${ENABLE_INT128_OPENMP_PATCH}")
-    target_link_libraries(libprimecount PRIVATE primesieve::primesieve "${LIB_OPENMP}" "${LIB_QUADMATH}" "${LIB_ATOMIC}")
+    target_compile_definitions(libprimecount PRIVATE ${PRIMECOUNT_COMPILE_DEFINITIONS})
+    target_link_libraries(libprimecount PRIVATE primesieve::primesieve ${PRIMECOUNT_LINK_LIBRARIES})
 
     target_compile_features(libprimecount
     PRIVATE
@@ -271,8 +277,8 @@ if(BUILD_STATIC_LIBS)
     add_library(libprimecount-static STATIC ${LIB_SRC})
     set_target_properties(libprimecount-static PROPERTIES OUTPUT_NAME primecount)
     target_compile_options(libprimecount-static PRIVATE "${WNO_UNINITIALIZED}")
-    target_compile_definitions(libprimecount-static PRIVATE "${HAVE_FLOAT128}" "${DISABLE_INT128}" "${ENABLE_DIV32}" "${ENABLE_ASSERT}" "${ENABLE_MULTIARCH}" "${ENABLE_INT128_OPENMP_PATCH}")
-    target_link_libraries(libprimecount-static PRIVATE primesieve::primesieve "${LIB_OPENMP}" "${LIB_QUADMATH}" "${LIB_ATOMIC}")
+    target_compile_definitions(libprimecount-static PRIVATE ${PRIMECOUNT_COMPILE_DEFINITIONS})
+    target_link_libraries(libprimecount-static PRIVATE primesieve::primesieve ${PRIMECOUNT_LINK_LIBRARIES})
 
     if(WITH_MSVC_CRT_STATIC)
         set_target_properties(libprimecount-static PROPERTIES MSVC_RUNTIME_LIBRARY "MultiThreaded")
@@ -307,7 +313,7 @@ endif()
 if(BUILD_PRIMECOUNT)
     add_executable(primecount ${BIN_SRC})
     target_link_libraries(primecount PRIVATE primecount::primecount primesieve::primesieve)
-    target_compile_definitions(primecount PRIVATE "${DISABLE_INT128}" "${ENABLE_DIV32}" "${ENABLE_ASSERT}" "${ENABLE_INT128_OPENMP_PATCH}")
+    target_compile_definitions(primecount PRIVATE ${PRIMECOUNT_COMPILE_DEFINITIONS})
     target_compile_features(primecount PRIVATE cxx_auto_type)
     install(TARGETS primecount DESTINATION ${CMAKE_INSTALL_BINDIR})
 

diff --git a/ChangeLog b/ChangeLog
@@ -1,9 +1,12 @@
-Changes in primecount-7.14, 2024-06-20
+Changes in primecount-7.14, 2024-06-22
 
+* Move x86 cpuid code from cpuid.hpp to src/x86/cpuid.cpp.
 * int128_t.hpp: Rename namespace port to pstd (portable std namespace).
-* popcnt.hpp: Improve GCC performance on x86 CPUs.
 * Sieve.hpp: Tune AVX512 code.
+* cpu_supports_popcnt.hpp: Simplify, move preprocessor checks to new multiarch_x86_popcnt.cmake.
 * multiarch_avx512_vpopcnt.cmake: Tune AVX512 code.
+* multiarch_x86_popcnt.cmake: Detect x86 POPCNT.
+* CMakeLists.txt: Use CMake list for all compile time definitions.
 
 Changes in primecount-7.13, 2024-04-15
 

diff --git a/cmake/OpenMP.cmake b/cmake/OpenMP.cmake
@@ -90,10 +90,12 @@ if(OpenMP_FOUND OR OpenMP_CXX_FOUND)
                 return 0;
             }" OpenMP_with_libatomic)
 
-        if(NOT OpenMP_with_libatomic)
+        if(OpenMP_with_libatomic)
+            list(APPEND PRIMECOUNT_LINK_LIBRARIES "${LIB_ATOMIC}")
+        else()
             set(LIB_ATOMIC "")
 
-            if (NOT DISABLE_INT128)
+            if(NOT DISABLE_INT128)
                 # As a last resort check if OpenMP supports int128_t if
                 # we include our <int128_OpenMP_patch.hpp> header.
                 # In this case OpenMP will use critical sections instead
@@ -118,7 +120,7 @@ if(OpenMP_FOUND OR OpenMP_CXX_FOUND)
                     }" OpenMP_int128_patch)
 
                 if(OpenMP_int128_patch)
-                    set(ENABLE_INT128_OPENMP_PATCH "ENABLE_INT128_OPENMP_PATCH")
+                    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "ENABLE_INT128_OPENMP_PATCH")
                 endif()
             endif()
         endif()
@@ -129,7 +131,7 @@ if(OpenMP_FOUND OR OpenMP_CXX_FOUND)
     # OpenMP has been tested successfully, enable it
     if(OpenMP OR OpenMP_with_libatomic OR OpenMP_int128_patch)
         if(TARGET OpenMP::OpenMP_CXX)
-            set(LIB_OPENMP "OpenMP::OpenMP_CXX")
+            list(APPEND PRIMECOUNT_LINK_LIBRARIES "OpenMP::OpenMP_CXX")
         else()
             set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
         endif()

diff --git a/cmake/int128_t.cmake b/cmake/int128_t.cmake
@@ -45,7 +45,7 @@ check_cxx_source_compiles("
     }" int128)
 
 if(NOT int128)
-    set(DISABLE_INT128 "DISABLE_INT128")
+    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "DISABLE_INT128")
 endif()
 
 cmake_pop_check_state()
diff --git a/cmake/multiarch_arm_sve.cmake b/cmake/multiarch_arm_sve.cmake
@@ -72,7 +72,7 @@ check_cxx_source_compiles("
 " multiarch_arm_sve)
 
 if(multiarch_arm_sve)
-    set(ENABLE_MULTIARCH "ENABLE_MULTIARCH_ARM_SVE")
+    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "ENABLE_MULTIARCH_ARM_SVE")
 endif()
 
 cmake_pop_check_state()
diff --git a/cmake/multiarch_avx512_vpopcnt.cmake b/cmake/multiarch_avx512_vpopcnt.cmake
@@ -7,7 +7,7 @@ include(CheckCXXSourceCompiles)
 include(CMakePushCheckState)
 
 cmake_push_check_state()
-set(CMAKE_REQUIRED_INCLUDES "${PROJECT_SOURCE_DIR}/include")
+set(CMAKE_REQUIRED_INCLUDES "${PROJECT_SOURCE_DIR}")
 
 check_cxx_source_compiles("
     // GCC/Clang function multiversioning for AVX512 is not needed if
@@ -20,7 +20,7 @@ check_cxx_source_compiles("
       Error: AVX512 BMI2 multiarch not needed!
     #endif
 
-    #include <cpu_supports_avx512_bmi2.hpp>
+    #include <src/x86/cpuid.cpp>
     #include <immintrin.h>
     #include <stdint.h>
 
@@ -65,7 +65,7 @@ check_cxx_source_compiles("
         uint64_t cnt = 0;
         Sieve sieve;
 
-        if (cpu_supports_avx512_bmi2)
+        if (primecount::has_cpuid_avx512_bmi2())
             cnt = sieve.count_avx512_bmi2(&array[0], 10);
         else
             cnt = sieve.count_default(&array[0], 10);
@@ -75,7 +75,7 @@ check_cxx_source_compiles("
 " multiarch_avx512_vpopcnt)
 
 if(multiarch_avx512_vpopcnt)
-    set(ENABLE_MULTIARCH "ENABLE_MULTIARCH_AVX512_BMI2")
+    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "ENABLE_MULTIARCH_AVX512_BMI2")
 endif()
 
 cmake_pop_check_state()
diff --git a/cmake/multiarch_x86_popcnt.cmake b/cmake/multiarch_x86_popcnt.cmake
@@ -0,0 +1,53 @@
+# On x86 CPUs we need to enable the use of cpuid.cpp.
+# If cpuid.cpp compiles we assume it is a x86 CPU.
+
+include(CheckCXXSourceCompiles)
+include(CMakePushCheckState)
+
+cmake_push_check_state()
+set(CMAKE_REQUIRED_INCLUDES "${PROJECT_SOURCE_DIR}")
+
+check_cxx_source_compiles("
+    // Enable CPUID for POPCNT on x86 and x86-64 CPUs.
+    // This is required because not all x86 and x86-64 CPUs
+    // support the POPCNT instruction.
+    #if !(defined(__x86_64__) || \
+          defined(__i386__) || \
+          defined(_M_X64) || \
+          defined(_M_IX86))
+      Error: x86 POPCNT multiarch not needed!
+    #endif
+
+    // Both GCC and Clang (even Clang on Windows) define the __POPCNT__
+    // macro if the user compiles with -mpopcnt. The __POPCNT__
+    // macro is even defined if the user compiles with other flags
+    // such as -mavx or -march=native.
+    #if defined(__POPCNT__)
+        Error: x86 POPCNT multiarch not needed!
+
+    // The MSVC compiler does not support a POPCNT macro, but if the user
+    // compiles with e.g. /arch:AVX or /arch:AVX512 then MSVC defines
+    // the __AVX__ macro and POPCNT is also supported.
+    #elif defined(_MSC_VER) && defined(__AVX__)
+        Error: x86 POPCNT multiarch not needed!
+    #endif
+
+    #include <src/x86/cpuid.cpp>
+    #include <iostream>
+
+    int main()
+    {
+        if (primecount::has_cpuid_popcnt())
+            std::cout << \"CPU supports POPCNT!\" << std::endl;
+        else
+            std::cout << \"CPU does not support POPCNT!\" << std::endl;
+
+        return 0;
+    }
+" multiarch_x86_popcnt)
+
+if(multiarch_x86_popcnt)
+    list(APPEND PRIMECOUNT_COMPILE_DEFINITIONS "ENABLE_MULTIARCH_x86_POPCNT")
+endif()
+
+cmake_pop_check_state()
diff --git a/include/cpu_supports_avx512_bmi2.hpp b/include/cpu_supports_avx512_bmi2.hpp
@@ -11,80 +11,16 @@
 #ifndef CPU_SUPPORTS_AVX512_BMI2_HPP
 #define CPU_SUPPORTS_AVX512_BMI2_HPP
 
-#include <cpuid.hpp>
-#include <stdint.h>
+namespace primecount {
 
-#if defined(_MSC_VER)
-  #include <immintrin.h>
-#endif
-
-// CPUID bits documentation:
-// https://en.wikipedia.org/wiki/CPUID
-
-// %ebx bit flags
-#define bit_BMI2    (1 << 8)
-#define bit_AVX512F (1 << 16)
+bool has_cpuid_avx512_bmi2();
 
-// %ecx bit flags
-#define bit_AVX512_VPOPCNTDQ (1 << 14)
-
-// xgetbv bit flags
-#define XSTATE_SSE (1 << 1)
-#define XSTATE_YMM (1 << 2)
-#define XSTATE_ZMM (7 << 5)
+} // namespace
 
 namespace {
 
-// Get Value of Extended Control Register
-inline uint64_t get_xcr0()
-{
-#if defined(_MSC_VER)
-  return _xgetbv(0);
-#else
-  uint32_t eax;
-  uint32_t edx;
-
-  __asm__ ("xgetbv" : "=a"(eax), "=d"(edx) : "c"(0));
-  return eax | (uint64_t(edx) << 32);
-#endif
-}
-
-inline bool run_cpuid_avx512_bmi2()
-{
-  int abcd[4];
-
-  run_cpuid(1, 0, abcd);
-
-  int osxsave_mask = (1 << 27);
-
-  // Ensure OS supports extended processor state management
-  if ((abcd[2] & osxsave_mask) != osxsave_mask)
-    return false;
-
-  uint64_t ymm_mask = XSTATE_SSE | XSTATE_YMM;
-  uint64_t zmm_mask = XSTATE_SSE | XSTATE_YMM | XSTATE_ZMM;
-  uint64_t xcr0 = get_xcr0();
-
-  // Check AVX OS support
-  if ((xcr0 & ymm_mask) != ymm_mask)
-    return false;
-
-  // Check AVX512 OS support
-  if ((xcr0 & zmm_mask) != zmm_mask)
-    return false;
-
-  run_cpuid(7, 0, abcd);
-
-  if ((abcd[1] & bit_BMI2) != bit_BMI2)
-    return false;
-
-  // AVX512F, AVX512VPOPCNTDQ
-  return ((abcd[1] & bit_AVX512F) == bit_AVX512F &&
-          (abcd[2] & bit_AVX512_VPOPCNTDQ) == bit_AVX512_VPOPCNTDQ);
-}
-
 /// Initialized at startup
-bool cpu_supports_avx512_bmi2 = run_cpuid_avx512_bmi2();
+bool cpu_supports_avx512_bmi2 = primecount::has_cpuid_avx512_bmi2();
 
 } // namespace
 

diff --git a/include/cpu_supports_popcnt.hpp b/include/cpu_supports_popcnt.hpp
@@ -11,49 +11,17 @@
 #ifndef CPU_SUPPORTS_POPCNT_HPP
 #define CPU_SUPPORTS_POPCNT_HPP
 
-// Enable CPUID on x86 and x86-64 CPUs
-#if defined(__x86_64__) || \
-    defined(__i386__) || \
-    defined(_M_X64) || \
-    defined(_M_IX86)
-
-// Both GCC and Clang (even Clang on Windows) define the __POPCNT__
-// macro if the user compiles with -mpopcnt. The __POPCNT__
-// macro is even defined if the user compiles with other flags
-// such as -mavx or -march=native.
-#if defined(__POPCNT__)
-  #define HAS_POPCNT
-// The MSVC compiler does not support a POPCNT macro, but if the user
-// compiles with e.g. /arch:AVX or /arch:AVX512 then MSVC defines
-// the __AVX__ macro and POPCNT is also supported.
-#elif defined(_MSC_VER) && defined(__AVX__)
-  #define HAS_POPCNT
-#endif
+namespace primecount {
 
-#if !defined(HAS_POPCNT)
+bool has_cpuid_popcnt();
 
-#include <cpuid.hpp>
-#define ENABLE_CPUID_POPCNT
+} // namespace
 
 namespace {
 
-inline bool run_cpuid_supports_popcnt()
-{
-  int abcd[4];
-  run_cpuid(1, 0, abcd);
-
-  // %ecx POPCNT bit flag
-  // https://en.wikipedia.org/wiki/CPUID
-  int bit_POPCNT = 1 << 23;
-  return (abcd[2] & bit_POPCNT) == bit_POPCNT;
-}
-
 /// Initialized at startup
-bool cpu_supports_popcnt = run_cpuid_supports_popcnt();
+bool cpu_supports_popcnt = primecount::has_cpuid_popcnt();
 
 } // namespace
 
-#endif // !defined(HAS_POPCNT)
-#endif // CPUID
-
 #endif