POPCNT runtime detection using CPUID for x86 CPUs

CMakeLists.txt

@@ -46,6 +46,10 @@ if(NOT isMultiConfig AND NOT CMAKE_BUILD_TYPE)
     "Choose the type of build, options are: None Debug Release RelWithDebInfo MinSizeRel." FORCE)
 endif()
 
+if(CMAKE_BUILD_TYPE STREQUAL "Debug")
+    set(ENABLE_ASSERT "ENABLE_ASSERT")
+endif()
+
 # primesieve binary source files #####################################
 
 set(BIN_SRC src/app/CmdOptions.cpp
@@ -106,8 +110,10 @@ endif()
 # Check if compiler supports x64 multiarch ###########################
 
 if(WITH_MULTIARCH)
-    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_popcnt_bmi.cmake")
-    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_avx512.cmake")
+    include("${PROJECT_SOURCE_DIR}/cmake/multiarch_avx512_vbmi2.cmake")
+    if(multiarch_avx512_vbmi2)
+        set(MULTIARCH_AVX512 "MULTIARCH_AVX512")
+    endif()
 endif()
 
 # libprimesieve (shared library) #####################################
@@ -123,16 +129,7 @@ if(BUILD_SHARED_LIBS)
     set_target_properties(libprimesieve PROPERTIES SOVERSION ${PRIMESIEVE_SOVERSION_MAJOR})
     set_target_properties(libprimesieve PROPERTIES VERSION ${PRIMESIEVE_SOVERSION})
     target_compile_options(libprimesieve PRIVATE ${FTREE_VECTORIZE_FLAG} ${FVECT_COST_MODEL_FLAG})
-
-    if(multiarch_popcnt_bmi)
-        target_compile_definitions(libprimesieve PRIVATE "MULTIARCH_POPCNT_BMI")
-    endif()
-    if(multiarch_avx512)
-        target_compile_definitions(libprimesieve PRIVATE "MULTIARCH_AVX512")
-    endif()
-    if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-        target_compile_definitions(libprimesieve PRIVATE "ENABLE_ASSERT")
-    endif()
+    target_compile_definitions(libprimesieve PRIVATE "${ENABLE_ASSERT}" "${MULTIARCH_AVX512}")
 
     if(WIN32_MSVC_COMPATIBLE)
         # On Windows the shared library will be named primesieve.dll
@@ -170,16 +167,7 @@ if(BUILD_STATIC_LIBS)
     set_target_properties(libprimesieve-static PROPERTIES OUTPUT_NAME primesieve)
     target_link_libraries(libprimesieve-static PRIVATE Threads::Threads ${LIBATOMIC})
     target_compile_options(libprimesieve-static PRIVATE ${FTREE_VECTORIZE_FLAG} ${FVECT_COST_MODEL_FLAG})
-
-    if(multiarch_popcnt_bmi)
-        target_compile_definitions(libprimesieve-static PRIVATE "MULTIARCH_POPCNT_BMI")
-    endif()
-    if(multiarch_avx512)
-        target_compile_definitions(libprimesieve-static PRIVATE "MULTIARCH_AVX512")
-    endif()
-    if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-        target_compile_definitions(libprimesieve-static PRIVATE "ENABLE_ASSERT")
-    endif()
+    target_compile_definitions(libprimesieve-static PRIVATE "${ENABLE_ASSERT}" "${MULTIARCH_AVX512}")
 
     if(WITH_MSVC_CRT_STATIC)
         set_target_properties(libprimesieve-static PROPERTIES MSVC_RUNTIME_LIBRARY "MultiThreaded")
@@ -234,16 +222,10 @@ endif()
 if(BUILD_PRIMESIEVE)
     add_executable(primesieve ${BIN_SRC})
     target_link_libraries(primesieve primesieve::primesieve Threads::Threads)
+    target_compile_definitions(primesieve PRIVATE "${ENABLE_ASSERT}")
     target_compile_features(primesieve PRIVATE cxx_auto_type)
     install(TARGETS primesieve DESTINATION ${CMAKE_INSTALL_BINDIR})
 
-    if(multiarch_avx512)
-        target_compile_definitions(primesieve PRIVATE "MULTIARCH_AVX512")
-    endif()
-    if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-        target_compile_definitions(primesieve PRIVATE "ENABLE_ASSERT")
-    endif()
-
     if(WITH_MSVC_CRT_STATIC)
         set_target_properties(primesieve PROPERTIES MSVC_RUNTIME_LIBRARY "MultiThreaded")
     endif()

ChangeLog

@@ -1,3 +1,10 @@
+Changes in version 12.3, 04/04/2024
+===================================
+
+* Add runtime POPCNT detection using CPUID for x86 CPUs.
+* Improve GCC/Clang multiarch preprocessor logic.
+* CMakeLists.txt: Remove POPCNT/BMI check for x86 CPUs.
+
 Changes in version 12.2, 30/03/2024
 ===================================
 

cmake/multiarch_avx512.cmake → cmake/multiarch_avx512_vbmi2.cmake

@@ -11,15 +11,15 @@ check_cxx_source_compiles("
         public:
         __attribute__ ((target (\"default\")))
         void fillNextPrimes(uint64_t* primes64);
-        __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2,popcnt\")))
+        __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2\")))
         void fillNextPrimes(uint64_t* primes64);
     };
     __attribute__ ((target (\"default\")))
     void PrimeGenerator::fillNextPrimes(uint64_t* primes64)
     {
         primes64[0] = 2;
     }
-    __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2,popcnt\")))
+    __attribute__ ((target (\"avx512f,avx512vbmi,avx512vbmi2\")))
     void PrimeGenerator::fillNextPrimes(uint64_t* primes64)
     {
         __m512i bytes_0_to_7 = _mm512_setr_epi64(0, 1, 2, 3, 4, 5, 6, 7);
@@ -36,4 +36,4 @@ check_cxx_source_compiles("
         p.fillNextPrimes(primes);
         return 0;
     }
-" multiarch_avx512)
+" multiarch_avx512_vbmi2)

include/primesieve/CPUID.hpp

@@ -0,0 +1,94 @@
+///
+/// @file  CPUID.hpp
+/// @brief POPCNT detection fo x86 and x86-64 CPUs.
+///
+/// Copyright (C) 2024 Kim Walisch, <[email protected]>
+///
+/// This file is distributed under the BSD License. See the COPYING
+/// file in the top level directory.
+///
+
+#ifndef CPUID_HPP
+#define CPUID_HPP
+
+// Enable 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
+
+#if defined(_MSC_VER)
+  #include <intrin.h>
+#endif
+
+namespace {
+
+inline void run_CPUID(int eax, int ecx, int* abcd)
+{
+#if defined(_MSC_VER)
+  __cpuidex(abcd, eax, ecx);
+#else
+  int ebx = 0;
+  int edx = 0;
+
+  #if defined(__i386__) && \
+      defined(__PIC__)
+    /* in case of PIC under 32-bit EBX cannot be clobbered */
+    __asm__ ("movl %%ebx, %%edi;"
+             "cpuid;"
+             "xchgl %%ebx, %%edi;"
+             : "=D" (ebx),
+               "+a" (eax),
+               "+c" (ecx),
+               "=d" (edx));
+  #else
+    __asm__ ("cpuid;"
+             : "+b" (ebx),
+               "+a" (eax),
+               "+c" (ecx),
+               "=d" (edx));
+  #endif
+
+  abcd[0] = eax;
+  abcd[1] = ebx;
+  abcd[2] = ecx;
+  abcd[3] = edx;
+#endif
+}
+
+#if !defined(HAS_POPCNT)
+#define ENABLE_CPUID_POPCNT
+
+inline bool run_CPUID_POPCNT()
+{
+  // %ecx POPCNT bit flag
+  int bit_POPCNT = 1 << 23;
+  int abcd[4];
+
+  run_CPUID(1, 0, abcd);
+  return (abcd[2] & bit_POPCNT) == bit_POPCNT;
+}
+
+/// Initialized at startup
+const bool HAS_CPUID_POPCNT = run_CPUID_POPCNT();
+
+#endif // ENABLE_CPUID_POPCNT
+
+} // namespace
+
+#endif // x86 CPU
+
+#endif

include/primesieve/PrimeGenerator.hpp

@@ -6,7 +6,7 @@
 ///         returns the primes. When there are no more primes left in
 ///         the vector PrimeGenerator generates new primes.
 ///
-/// Copyright (C) 2023 Kim Walisch, <[email protected]>
+/// Copyright (C) 2024 Kim Walisch, <[email protected]>
 ///
 /// This file is distributed under the BSD License. See the COPYING
 /// file in the top level directory.
@@ -23,6 +23,21 @@
 #include <stdint.h>
 #include <cstddef>
 
+#if defined(MULTIARCH_AVX512)
+  // GCC/Clang function multiversioning for AVX512 is not needed if
+  // the user compiles with -mavx512f -mavx512vbmi -mavx512vbmi2.
+  // GCC/Clang function multiversioning generally causes a minor
+  // overhead, hence we disable it if it is not needed.
+  #if defined(__AVX512__) || (defined(__AVX512F__) && \
+                              defined(__AVX512VBMI__) && \
+                              defined(__AVX512VBMI2__))
+    #undef MULTIARCH_AVX512
+  #else
+    #define MULTIARCH_TARGET_DEFAULT
+    #define MULTIARCH_TARGET_AVX512
+  #endif
+#endif
+
 namespace primesieve {
 
 class PreSieve;
@@ -34,22 +49,15 @@ class PrimeGenerator : public Erat
   void fillPrevPrimes(Vector<uint64_t>& primes, std::size_t* size);
   static uint64_t maxCachedPrime();
 
-#if defined(MULTIARCH_POPCNT_BMI)
-  #define MULTIARCH
-  __attribute__ ((target ("popcnt,bmi")))
-  void fillNextPrimes(Vector<uint64_t>& primes, std::size_t* size);
+#if defined(MULTIARCH_TARGET_DEFAULT)
+  __attribute__ ((target ("default")))
 #endif
-
-#if defined(MULTIARCH_AVX512)
-  #define MULTIARCH
-  __attribute__ ((target ("avx512f,avx512vbmi,avx512vbmi2,popcnt")))
   void fillNextPrimes(Vector<uint64_t>& primes, std::size_t* size);
-#endif
 
-#if defined(MULTIARCH)
-  __attribute__ ((target ("default")))
-#endif
+#if defined(MULTIARCH_TARGET_AVX512)
+  __attribute__ ((target ("avx512f,avx512vbmi,avx512vbmi2")))
   void fillNextPrimes(Vector<uint64_t>& primes, std::size_t* size);
+#endif
 
 private:
   bool isInit_ = false;