Refactor

CMakeLists.txt

@@ -154,12 +154,27 @@ 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()
+
 # On x86 CPUs compile cpuid.cpp ######################################
 
 include("${PROJECT_SOURCE_DIR}/cmake/x86_cpuid.cmake")
 
 if(x86_cpuid)
-    set(LIB_SRC ${LIB_SRC} src/arch/x86/cpuid.cpp)
+    set(LIB_SRC ${LIB_SRC} src/x86/cpuid.cpp)
 endif()
 
 # Enable __float128 support (requires libquadmath) ###################
@@ -196,15 +211,6 @@ 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)

ChangeLog

@@ -1,6 +1,6 @@
 Changes in primecount-7.14, 2024-06-21
 
-* Move x86 cpuid code from cpuid.hpp to src/arch/x86/cpuid.cpp.
+* 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.

cmake/multiarch_avx512_vpopcnt.cmake

@@ -20,7 +20,7 @@ check_cxx_source_compiles("
       Error: AVX512 BMI2 multiarch not needed!
     #endif
 
-    #include <src/arch/x86/cpuid.cpp>
+    #include <src/x86/cpuid.cpp>
     #include <immintrin.h>
     #include <stdint.h>
 
@@ -75,7 +75,7 @@ check_cxx_source_compiles("
 " multiarch_avx512_vpopcnt)
 
 if(multiarch_avx512_vpopcnt)
-    set(ENABLE_MULTIARCH "ENABLE_MULTIARCH_AVX512_BMI2")
+    set(ENABLE_MULTIARCH "${ENABLE_MULTIARCH} ENABLE_MULTIARCH_AVX512_BMI2")
 endif()
 
 cmake_pop_check_state()

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 "${PROJECT_SOURCE_DIR}")
+
+check_cxx_source_compiles("
+    // Enable CPUID for POPCNT on x86 and x86-64 CPUs.
+    // This is required because not all x86 and x86-64 CPUs
+    // support the POPCNT instruction.
+    #if !(defined(__x86_64__) || \
+          defined(__i386__) || \
+          defined(_M_X64) || \
+          defined(_M_IX86))
+      Error: x86 POPCNT multiarch not needed!
+    #endif
+
+    // Both GCC and Clang (even Clang on Windows) define the __POPCNT__
+    // macro if the user compiles with -mpopcnt. The __POPCNT__
+    // macro is even defined if the user compiles with other flags
+    // such as -mavx or -march=native.
+    #if defined(__POPCNT__)
+        Error: x86 POPCNT multiarch not needed!
+
+    // The MSVC compiler does not support a POPCNT macro, but if the user
+    // compiles with e.g. /arch:AVX or /arch:AVX512 then MSVC defines
+    // the __AVX__ macro and POPCNT is also supported.
+    #elif defined(_MSC_VER) && defined(__AVX__)
+        Error: x86 POPCNT multiarch not needed!
+    #endif
+
+    #include <src/x86/cpuid.cpp>
+    #include <iostream>
+
+    int main()
+    {
+        if (primecount::has_cpuid_popcnt())
+            std::cout << \"CPU supports POPCNT!\" << std::endl;
+        else
+            std::cout << \"CPU does not support POPCNT!\" << std::endl;
+
+        return 0;
+    }
+" multiarch_x86_popcnt)
+
+if(multiarch_x86_popcnt)
+    set(ENABLE_MULTIARCH "${ENABLE_MULTIARCH} ENABLE_MULTIARCH_x86_POPCNT")
+endif()
+
+cmake_pop_check_state()

include/cpu_supports_popcnt.hpp

@@ -11,31 +11,6 @@
 #ifndef CPU_SUPPORTS_POPCNT_HPP
 #define CPU_SUPPORTS_POPCNT_HPP
 
-// 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)
-
-// 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
-
-#if !defined(HAS_POPCNT)
-#define ENABLE_MULTIARCH_x86_POPCNT
-
 namespace primecount {
 
 bool has_cpuid_popcnt();
@@ -49,7 +24,4 @@ bool cpu_supports_popcnt = primecount::has_cpuid_popcnt();
 
 } // namespace
 
-#endif // !defined(HAS_POPCNT)
-#endif // x86 or x86-64
-
 #endif

include/popcnt.hpp

@@ -12,11 +12,13 @@
 #ifndef POPCNT_HPP
 #define POPCNT_HPP
 
-#include <cpu_supports_popcnt.hpp>
 #include <macros.hpp>
-
 #include <stdint.h>
 
+#if defined(ENABLE_MULTIARCH_x86_POPCNT)
+  #include <cpu_supports_popcnt.hpp>
+#endif
+
 namespace {
 
 /// This uses fewer arithmetic operations than any other known

src/arch/x86/cpuid.cpp → src/x86/cpuid.cpp

@@ -31,7 +31,7 @@
 #define XSTATE_YMM (1 << 2)
 #define XSTATE_ZMM (7 << 5)
 
-namespace primecount {
+namespace {
 
 void run_cpuid(int eax, int ecx, int* abcd)
 {
@@ -66,13 +66,6 @@ void run_cpuid(int eax, int ecx, int* abcd)
 #endif
 }
 
-bool has_cpuid_popcnt()
-{
-  int abcd[4];
-  run_cpuid(1, 0, abcd);
-  return (abcd[2] & bit_POPCNT) == bit_POPCNT;
-}
-
 // Get Value of Extended Control Register
 uint64_t get_xcr0()
 {
@@ -87,6 +80,17 @@ uint64_t get_xcr0()
 #endif
 }
 
+} // namespace
+
+namespace primecount {
+
+bool has_cpuid_popcnt()
+{
+  int abcd[4];
+  run_cpuid(1, 0, abcd);
+  return (abcd[2] & bit_POPCNT) == bit_POPCNT;
+}
+
 bool has_cpuid_avx512_bmi2()
 {
   int abcd[4];