Move x86 cpuid code from cpuid.hpp to src/arch/x86/cpuid.cpp

CMakeLists.txt

@@ -154,6 +154,14 @@ else()
     set(LIB_SRC ${LIB_SRC} src/gourdon/AC.cpp)
 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)
+endif()
+
 # Enable __float128 support (requires libquadmath) ###################
 
 if(WITH_FLOAT128)

ChangeLog

@@ -1,5 +1,6 @@
-Changes in primecount-7.14, 2024-06-20
+Changes in primecount-7.14, 2024-06-21
 
+* Move x86 cpuid code from cpuid.hpp to src/arch/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

@@ -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/arch/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);

cmake/x86_cpuid.cmake

@@ -0,0 +1,29 @@
+# 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("
+    #include <src/arch/x86/cpuid.cpp>
+    #include <iostream>
+
+    int main()
+    {
+        int abcd[4];
+        primecount::run_cpuid(1, 0, abcd);
+        int ecx = abcd[2];
+
+        if (ecx & (1 << 23)) == (1 << 23))
+            std::cout << \"CPU supports POPCNT!\" << std::endl;
+        else
+            std::cout << \"CPU does not support POPCNT!\" << std::endl;
+
+        return 0;
+    }
+" x86_cpuid)
+
+cmake_pop_check_state()

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
 

include/cpu_supports_popcnt.hpp

@@ -11,7 +11,9 @@
 #ifndef CPU_SUPPORTS_POPCNT_HPP
 #define CPU_SUPPORTS_POPCNT_HPP
 
-// Enable CPUID on x86 and x86-64 CPUs
+// 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) || \
@@ -23,6 +25,7 @@
 // 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.
@@ -31,29 +34,22 @@
 #endif
 
 #if !defined(HAS_POPCNT)
-
-#include <cpuid.hpp>
 #define ENABLE_CPUID_POPCNT
 
-namespace {
+namespace primecount {
 
-inline bool run_cpuid_supports_popcnt()
-{
-  int abcd[4];
-  run_cpuid(1, 0, abcd);
+bool has_cpuid_popcnt();
 
-  // %ecx POPCNT bit flag
-  // https://en.wikipedia.org/wiki/CPUID
-  int bit_POPCNT = 1 << 23;
-  return (abcd[2] & bit_POPCNT) == bit_POPCNT;
-}
+} // namespace
+
+namespace {
 
 /// 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 // x86 or x86-64
 
 #endif