xref: /third_party/mesa3d/src/util/u_cpu_detect.c

/**************************************************************************
 *
 * Copyright 2008 Dennis Smit
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * on the rights to use, copy, modify, merge, publish, distribute, sub
 * license, and/or sell copies of the Software, and to permit persons to whom
 * the Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.  IN NO EVENT SHALL
 * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 **************************************************************************/

/**
 * @file
 * CPU feature detection.
 *
 * @author Dennis Smit
 * @author Based on the work of Eric Anholt <anholt@FreeBSD.org>
 */

#include "pipe/p_config.h"
#include "pipe/p_compiler.h"

#include "util/u_debug.h"
#include "u_cpu_detect.h"
#include "u_math.h"
#include "c11/threads.h"

#include <stdio.h>
#include <inttypes.h>

#if defined(PIPE_ARCH_PPC)
#if defined(PIPE_OS_APPLE)
#include <sys/sysctl.h>
#else
#include <signal.h>
#include <setjmp.h>
#endif
#endif

#if defined(PIPE_OS_BSD)
#include <sys/param.h>
#include <sys/sysctl.h>
#include <machine/cpu.h>
#endif

#if defined(PIPE_OS_FREEBSD)
#if __has_include(<sys/auxv.h>)
#include <sys/auxv.h>
#define HAVE_ELF_AUX_INFO
#endif
#endif

#if defined(PIPE_OS_LINUX)
#include <signal.h>
#include <fcntl.h>
#include <elf.h>
#endif

#ifdef PIPE_OS_UNIX
#include <unistd.h>
#endif

#if defined(HAS_ANDROID_CPUFEATURES)
#include <cpu-features.h>
#endif

#if defined(PIPE_OS_WINDOWS)
#include <windows.h>
#if defined(PIPE_CC_MSVC)
#include <intrin.h>
#endif
#endif

#if defined(HAS_SCHED_H)
#include <sched.h>
#endif

DEBUG_GET_ONCE_BOOL_OPTION(dump_cpu, "GALLIUM_DUMP_CPU", false)


struct util_cpu_caps_t util_cpu_caps;

#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
static int has_cpuid(void);
#endif


#if defined(PIPE_ARCH_PPC) && !defined(PIPE_OS_APPLE) && !defined(PIPE_OS_BSD) && !defined(PIPE_OS_LINUX)
static jmp_buf  __lv_powerpc_jmpbuf;
static volatile sig_atomic_t __lv_powerpc_canjump = 0;

static void
sigill_handler(int sig)
{
   if (!__lv_powerpc_canjump) {
      signal (sig, SIG_DFL);
      raise (sig);
   }

   __lv_powerpc_canjump = 0;
   longjmp(__lv_powerpc_jmpbuf, 1);
}
#endif

#if defined(PIPE_ARCH_PPC)
static void
check_os_altivec_support(void)
{
#if defined(__ALTIVEC__)
   util_cpu_caps.has_altivec = 1;
#endif
#if defined(__VSX__)
   util_cpu_caps.has_vsx = 1;
#endif
#if defined(__ALTIVEC__) && defined(__VSX__)
/* Do nothing */
#elif defined(PIPE_OS_APPLE) || defined(PIPE_OS_NETBSD) || defined(PIPE_OS_OPENBSD)
#ifdef HW_VECTORUNIT
   int sels[2] = {CTL_HW, HW_VECTORUNIT};
#else
   int sels[2] = {CTL_MACHDEP, CPU_ALTIVEC};
#endif
   int has_vu = 0;
   size_t len = sizeof (has_vu);
   int err;

   err = sysctl(sels, 2, &has_vu, &len, NULL, 0);

   if (err == 0) {
      if (has_vu != 0) {
         util_cpu_caps.has_altivec = 1;
      }
   }
#elif defined(PIPE_OS_FREEBSD) /* !PIPE_OS_APPLE && !PIPE_OS_NETBSD && !PIPE_OS_OPENBSD */
   unsigned long hwcap = 0;
#ifdef HAVE_ELF_AUX_INFO
   elf_aux_info(AT_HWCAP, &hwcap, sizeof(hwcap));
#else
   size_t len = sizeof(hwcap);
   sysctlbyname("hw.cpu_features", &hwcap, &len, NULL, 0);
#endif
   if (hwcap & PPC_FEATURE_HAS_ALTIVEC)
      util_cpu_caps.has_altivec = 1;
   if (hwcap & PPC_FEATURE_HAS_VSX)
      util_cpu_caps.has_vsx = 1;
#elif defined(PIPE_OS_LINUX) /* !PIPE_OS_FREEBSD */
#if defined(PIPE_ARCH_PPC_64)
    Elf64_auxv_t aux;
#else
    Elf32_auxv_t aux;
#endif
    int fd = open("/proc/self/auxv", O_RDONLY | O_CLOEXEC);
    if (fd >= 0) {
       while (read(fd, &aux, sizeof(aux)) == sizeof(aux)) {
          if (aux.a_type == AT_HWCAP) {
             char *env_vsx = getenv("GALLIVM_VSX");
             uint64_t hwcap = aux.a_un.a_val;
             util_cpu_caps.has_altivec = (hwcap >> 28) & 1;
             if (!env_vsx || env_vsx[0] != '0') {
                util_cpu_caps.has_vsx  = (hwcap >>  7) & 1;
             }
             break;
          }
       }
       close(fd);
    }
#else /* !PIPE_OS_APPLE && !PIPE_OS_BSD && !PIPE_OS_LINUX */
   /* not on Apple/Darwin or Linux, do it the brute-force way */
   /* this is borrowed from the libmpeg2 library */
   signal(SIGILL, sigill_handler);
   if (setjmp(__lv_powerpc_jmpbuf)) {
      signal(SIGILL, SIG_DFL);
   } else {
      boolean enable_altivec = TRUE;    /* Default: enable  if available, and if not overridden */
      boolean enable_vsx = TRUE;
#ifdef DEBUG
      /* Disabling Altivec code generation is not the same as disabling VSX code generation,
       * which can be done simply by passing -mattr=-vsx to the LLVM compiler; cf.
       * lp_build_create_jit_compiler_for_module().
       * If you want to disable Altivec code generation, the best place to do it is here.
       */
      char *env_control = getenv("GALLIVM_ALTIVEC");    /* 1=enable (default); 0=disable */
      if (env_control && env_control[0] == '0') {
         enable_altivec = FALSE;
      }
#endif
      /* VSX instructions can be explicitly enabled/disabled via GALLIVM_VSX=1 or 0 */
      char *env_vsx = getenv("GALLIVM_VSX");
      if (env_vsx && env_vsx[0] == '0') {
         enable_vsx = FALSE;
      }
      if (enable_altivec) {
         __lv_powerpc_canjump = 1;

         __asm __volatile
            ("mtspr 256, %0\n\t"
             "vand %%v0, %%v0, %%v0"
             :
             : "r" (-1));

         util_cpu_caps.has_altivec = 1;

         if (enable_vsx) {
            __asm __volatile("xxland %vs0, %vs0, %vs0");
            util_cpu_caps.has_vsx = 1;
         }
         signal(SIGILL, SIG_DFL);
      } else {
         util_cpu_caps.has_altivec = 0;
      }
   }
#endif /* !PIPE_OS_APPLE && !PIPE_OS_LINUX */
}
#endif /* PIPE_ARCH_PPC */


#if defined(PIPE_ARCH_X86) || defined (PIPE_ARCH_X86_64)
static int has_cpuid(void)
{
#if defined(PIPE_ARCH_X86)
#if defined(PIPE_OS_GCC)
   int a, c;

   __asm __volatile
      ("pushf\n"
       "popl %0\n"
       "movl %0, %1\n"
       "xorl $0x200000, %0\n"
       "push %0\n"
       "popf\n"
       "pushf\n"
       "popl %0\n"
       : "=a" (a), "=c" (c)
       :
       : "cc");

   return a != c;
#else
   /* FIXME */
   return 1;
#endif
#elif defined(PIPE_ARCH_X86_64)
   return 1;
#else
   return 0;
#endif
}


/**
 * @sa cpuid.h included in gcc-4.3 onwards.
 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh.aspx
 */
static inline void
cpuid(uint32_t ax, uint32_t *p)
{
#if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
   __asm __volatile (
     "xchgl %%ebx, %1\n\t"
     "cpuid\n\t"
     "xchgl %%ebx, %1"
     : "=a" (p[0]),
       "=S" (p[1]),
       "=c" (p[2]),
       "=d" (p[3])
     : "0" (ax)
   );
#elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
   __asm __volatile (
     "cpuid\n\t"
     : "=a" (p[0]),
       "=b" (p[1]),
       "=c" (p[2]),
       "=d" (p[3])
     : "0" (ax)
   );
#elif defined(PIPE_CC_MSVC)
   __cpuid(p, ax);
#else
   p[0] = 0;
   p[1] = 0;
   p[2] = 0;
   p[3] = 0;
#endif
}

/**
 * @sa cpuid.h included in gcc-4.4 onwards.
 * @sa http://msdn.microsoft.com/en-us/library/hskdteyh%28v=vs.90%29.aspx
 */
static inline void
cpuid_count(uint32_t ax, uint32_t cx, uint32_t *p)
{
#if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
   __asm __volatile (
     "xchgl %%ebx, %1\n\t"
     "cpuid\n\t"
     "xchgl %%ebx, %1"
     : "=a" (p[0]),
       "=S" (p[1]),
       "=c" (p[2]),
       "=d" (p[3])
     : "0" (ax), "2" (cx)
   );
#elif defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86_64)
   __asm __volatile (
     "cpuid\n\t"
     : "=a" (p[0]),
       "=b" (p[1]),
       "=c" (p[2]),
       "=d" (p[3])
     : "0" (ax), "2" (cx)
   );
#elif defined(PIPE_CC_MSVC)
   __cpuidex(p, ax, cx);
#else
   p[0] = 0;
   p[1] = 0;
   p[2] = 0;
   p[3] = 0;
#endif
}


static inline uint64_t xgetbv(void)
{
#if defined(PIPE_CC_GCC)
   uint32_t eax, edx;

   __asm __volatile (
     ".byte 0x0f, 0x01, 0xd0" // xgetbv isn't supported on gcc < 4.4
     : "=a"(eax),
       "=d"(edx)
     : "c"(0)
   );

   return ((uint64_t)edx << 32) | eax;
#elif defined(PIPE_CC_MSVC) && defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
   return _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
#else
   return 0;
#endif
}


#if defined(PIPE_ARCH_X86)
PIPE_ALIGN_STACK static inline boolean sse2_has_daz(void)
{
   alignas(16) struct {
      uint32_t pad1[7];
      uint32_t mxcsr_mask;
      uint32_t pad2[128-8];
   } fxarea;

   fxarea.mxcsr_mask = 0;
#if defined(PIPE_CC_GCC)
   __asm __volatile ("fxsave %0" : "+m" (fxarea));
#elif defined(PIPE_CC_MSVC) || defined(PIPE_CC_ICL)
   _fxsave(&fxarea);
#else
   fxarea.mxcsr_mask = 0;
#endif
   return !!(fxarea.mxcsr_mask & (1 << 6));
}
#endif

#endif /* X86 or X86_64 */

#if defined(PIPE_ARCH_ARM)
static void
check_os_arm_support(void)
{
   /*
    * On Android, the cpufeatures library is preferred way of checking
    * CPU capabilities. However, it is not available for standalone Mesa
    * builds, i.e. when Android build system (Android.mk-based) is not
    * used. Because of this we cannot use PIPE_OS_ANDROID here, but rather
    * have a separate macro that only gets enabled from respective Android.mk.
    */
#if defined(__ARM_NEON) || defined(__ARM_NEON__)
   util_cpu_caps.has_neon = 1;
#elif defined(PIPE_OS_FREEBSD) && defined(HAVE_ELF_AUX_INFO)
   unsigned long hwcap = 0;
   elf_aux_info(AT_HWCAP, &hwcap, sizeof(hwcap));
   if (hwcap & HWCAP_NEON)
      util_cpu_caps.has_neon = 1;
#elif defined(HAS_ANDROID_CPUFEATURES)
   AndroidCpuFamily cpu_family = android_getCpuFamily();
   uint64_t cpu_features = android_getCpuFeatures();

   if (cpu_family == ANDROID_CPU_FAMILY_ARM) {
      if (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON)
         util_cpu_caps.has_neon = 1;
   }
#elif defined(PIPE_OS_LINUX)
    Elf32_auxv_t aux;
    int fd;

    fd = open("/proc/self/auxv", O_RDONLY | O_CLOEXEC);
    if (fd >= 0) {
       while (read(fd, &aux, sizeof(Elf32_auxv_t)) == sizeof(Elf32_auxv_t)) {
          if (aux.a_type == AT_HWCAP) {
             uint32_t hwcap = aux.a_un.a_val;

             util_cpu_caps.has_neon = (hwcap >> 12) & 1;
             break;
          }
       }
       close (fd);
    }
#endif /* PIPE_OS_LINUX */
}

#elif defined(PIPE_ARCH_AARCH64)
static void
check_os_arm_support(void)
{
    util_cpu_caps.has_neon = true;
}
#endif /* PIPE_ARCH_ARM || PIPE_ARCH_AARCH64 */

#if defined(PIPE_ARCH_MIPS64)
static void
check_os_mips64_support(void)
{
#if defined(PIPE_OS_LINUX)
    Elf64_auxv_t aux;
    int fd;

    fd = open("/proc/self/auxv", O_RDONLY | O_CLOEXEC);
    if (fd >= 0) {
       while (read(fd, &aux, sizeof(Elf64_auxv_t)) == sizeof(Elf64_auxv_t)) {
          if (aux.a_type == AT_HWCAP) {
             uint64_t hwcap = aux.a_un.a_val;

             util_cpu_caps.has_msa = (hwcap >> 1) & 1;
             break;
          }
       }
       close (fd);
    }
#endif /* PIPE_OS_LINUX */
}
#endif /* PIPE_ARCH_MIPS64 */


static void
get_cpu_topology(void)
{
   /* Default. This is OK if L3 is not present or there is only one. */
   util_cpu_caps.num_L3_caches = 1;

   memset(util_cpu_caps.cpu_to_L3, 0xff, sizeof(util_cpu_caps.cpu_to_L3));

#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
   /* AMD Zen */
   if (util_cpu_caps.family >= CPU_AMD_ZEN1_ZEN2 &&
       util_cpu_caps.family < CPU_AMD_LAST) {
      uint32_t regs[4];

      uint32_t saved_mask[UTIL_MAX_CPUS / 32] = {0};
      uint32_t mask[UTIL_MAX_CPUS / 32] = {0};
      bool saved = false;

      uint32_t L3_found[UTIL_MAX_CPUS] = {0};
      uint32_t num_L3_caches = 0;
      util_affinity_mask *L3_affinity_masks = NULL;

      /* Query APIC IDs from each CPU core.
       *
       * An APIC ID is a logical ID of the CPU with respect to the cache
       * hierarchy, meaning that consecutive APIC IDs are neighbours in
       * the hierarchy, e.g. sharing the same cache.
       *
       * For example, CPU 0 can have APIC ID 0 and CPU 12 can have APIC ID 1,
       * which means that both CPU 0 and 12 are next to each other.
       * (e.g. they are 2 threads belonging to 1 SMT2 core)
       *
       * We need to find out which CPUs share the same L3 cache and they can
       * be all over the place.
       *
       * Querying the APIC ID can only be done by pinning the current thread
       * to each core. The original affinity mask is saved.
       *
       * Loop over all possible CPUs even though some may be offline.
       */
      for (int16_t i = 0; i < util_cpu_caps.max_cpus && i < UTIL_MAX_CPUS; i++) {
         uint32_t cpu_bit = 1u << (i % 32);

         mask[i / 32] = cpu_bit;

         /* The assumption is that trying to bind the thread to a CPU that is
          * offline will fail.
          */
         if (util_set_current_thread_affinity(mask,
                                              !saved ? saved_mask : NULL,
                                              util_cpu_caps.num_cpu_mask_bits)) {
            saved = true;

            /* Query the APIC ID of the current core. */
            cpuid(0x00000001, regs);
            unsigned apic_id = regs[1] >> 24;

            /* Query the total core count for the CPU */
            uint32_t core_count = 1;
            if (regs[3] & (1 << 28))
               core_count = (regs[1] >> 16) & 0xff;

            core_count = util_next_power_of_two(core_count);

            /* Query the L3 cache count. */
            cpuid_count(0x8000001D, 3, regs);
            unsigned cache_level = (regs[0] >> 5) & 0x7;
            unsigned cores_per_L3 = ((regs[0] >> 14) & 0xfff) + 1;

            if (cache_level != 3)
               continue;

            unsigned local_core_id = apic_id & (core_count - 1);
            unsigned phys_id = (apic_id & ~(core_count - 1)) >> util_logbase2(core_count);
            unsigned local_l3_cache_index = local_core_id / util_next_power_of_two(cores_per_L3);
#define L3_ID(p, i) (p << 16 | i << 1 | 1);

            unsigned l3_id = L3_ID(phys_id, local_l3_cache_index);
            int idx = -1;
            for (unsigned c = 0; c < num_L3_caches; c++) {
               if (L3_found[c] == l3_id) {
                  idx = c;
                  break;
               }
            }
            if (idx == -1) {
               idx = num_L3_caches;
               L3_found[num_L3_caches++] = l3_id;
               L3_affinity_masks = realloc(L3_affinity_masks, sizeof(util_affinity_mask) * num_L3_caches);
               if (!L3_affinity_masks)
                  return;
               memset(&L3_affinity_masks[num_L3_caches - 1], 0, sizeof(util_affinity_mask));
            }
            util_cpu_caps.cpu_to_L3[i] = idx;
            L3_affinity_masks[idx][i / 32] |= cpu_bit;

         }
         mask[i / 32] = 0;
      }

      util_cpu_caps.num_L3_caches = num_L3_caches;
      util_cpu_caps.L3_affinity_mask = L3_affinity_masks;

      if (saved) {
         if (debug_get_option_dump_cpu()) {
            fprintf(stderr, "CPU <-> L3 cache mapping:\n");
            for (unsigned i = 0; i < util_cpu_caps.num_L3_caches; i++) {
               fprintf(stderr, "  - L3 %u mask = ", i);
               for (int j = util_cpu_caps.max_cpus - 1; j >= 0; j -= 32)
                  fprintf(stderr, "%08x ", util_cpu_caps.L3_affinity_mask[i][j / 32]);
               fprintf(stderr, "\n");
            }
         }

         /* Restore the original affinity mask. */
         util_set_current_thread_affinity(saved_mask, NULL,
                                          util_cpu_caps.num_cpu_mask_bits);
      } else {
         if (debug_get_option_dump_cpu())
            fprintf(stderr, "Cannot set thread affinity for any thread.\n");
      }
   }
#endif
}

static void
util_cpu_detect_once(void)
{
   int available_cpus = 0;
   int total_cpus = 0;

   memset(&util_cpu_caps, 0, sizeof util_cpu_caps);

   /* Count the number of CPUs in system */
#if defined(PIPE_OS_WINDOWS)
   {
      SYSTEM_INFO system_info;
      GetSystemInfo(&system_info);
      available_cpus = MAX2(1, system_info.dwNumberOfProcessors);
   }
#elif defined(PIPE_OS_UNIX)
#  if defined(HAS_SCHED_GETAFFINITY)
   {
      /* sched_setaffinity() can be used to further restrict the number of
       * CPUs on which the process can run.  Use sched_getaffinity() to
       * determine the true number of available CPUs.
       *
       * FIXME: The Linux manual page for sched_getaffinity describes how this
       * simple implementation will fail with > 1024 CPUs, and we'll fall back
       * to the _SC_NPROCESSORS_ONLN path.  Support for > 1024 CPUs can be
       * added to this path once someone has such a system for testing.
       */
      cpu_set_t affin;
      if (sched_getaffinity(getpid(), sizeof(affin), &affin) == 0)
         available_cpus = CPU_COUNT(&affin);
   }
#  endif

   /* Linux, FreeBSD, DragonFly, and Mac OS X should have
    * _SC_NOPROCESSORS_ONLN.  NetBSD and OpenBSD should have HW_NCPUONLINE.
    * This is what FFmpeg uses on those platforms.
    */
#  if defined(PIPE_OS_BSD) && defined(HW_NCPUONLINE)
   if (available_cpus == 0) {
      const int mib[] = { CTL_HW, HW_NCPUONLINE };
      int ncpu;
      size_t len = sizeof(ncpu);

      sysctl(mib, 2, &ncpu, &len, NULL, 0);
      available_cpus = ncpu;
   }
#  elif defined(_SC_NPROCESSORS_ONLN)
   if (available_cpus == 0) {
      available_cpus = sysconf(_SC_NPROCESSORS_ONLN);
      if (available_cpus == ~0)
         available_cpus = 1;
   }
#  elif defined(PIPE_OS_BSD)
   if (available_cpus == 0) {
      const int mib[] = { CTL_HW, HW_NCPU };
      int ncpu;
      int len = sizeof(ncpu);

      sysctl(mib, 2, &ncpu, &len, NULL, 0);
      available_cpus = ncpu;
   }
#  endif /* defined(PIPE_OS_BSD) */

   /* Determine the maximum number of CPUs configured in the system.  This is
    * used to properly set num_cpu_mask_bits below.  On BSDs that don't have
    * HW_NCPUONLINE, it was not clear whether HW_NCPU is the number of
    * configured or the number of online CPUs.  For that reason, prefer the
    * _SC_NPROCESSORS_CONF path on all BSDs.
    */
#  if defined(_SC_NPROCESSORS_CONF)
   total_cpus = sysconf(_SC_NPROCESSORS_CONF);
   if (total_cpus == ~0)
      total_cpus = 1;
#  elif defined(PIPE_OS_BSD)
   {
      const int mib[] = { CTL_HW, HW_NCPU };
      int ncpu;
      int len = sizeof(ncpu);

      sysctl(mib, 2, &ncpu, &len, NULL, 0);
      total_cpus = ncpu;
   }
#  endif /* defined(PIPE_OS_BSD) */
#endif /* defined(PIPE_OS_UNIX) */

   util_cpu_caps.nr_cpus = MAX2(1, available_cpus);
   total_cpus = MAX2(total_cpus, util_cpu_caps.nr_cpus);

   util_cpu_caps.max_cpus = total_cpus;
   util_cpu_caps.num_cpu_mask_bits = align(total_cpus, 32);

   /* Make the fallback cacheline size nonzero so that it can be
    * safely passed to align().
    */
   util_cpu_caps.cacheline = sizeof(void *);

#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
   if (has_cpuid()) {
      uint32_t regs[4];
      uint32_t regs2[4];

      util_cpu_caps.cacheline = 32;

      /* Get max cpuid level */
      cpuid(0x00000000, regs);

      if (regs[0] >= 0x00000001) {
         unsigned int cacheline;

         cpuid (0x00000001, regs2);

         util_cpu_caps.x86_cpu_type = (regs2[0] >> 8) & 0xf;
         /* Add "extended family". */
         if (util_cpu_caps.x86_cpu_type == 0xf)
             util_cpu_caps.x86_cpu_type += ((regs2[0] >> 20) & 0xff);

         switch (util_cpu_caps.x86_cpu_type) {
         case 0x17:
            util_cpu_caps.family = CPU_AMD_ZEN1_ZEN2;
            break;
         case 0x18:
            util_cpu_caps.family = CPU_AMD_ZEN_HYGON;
            break;
         case 0x19:
            util_cpu_caps.family = CPU_AMD_ZEN3;
            break;
         default:
            if (util_cpu_caps.x86_cpu_type > 0x19)
               util_cpu_caps.family = CPU_AMD_ZEN_NEXT;
         }

         /* general feature flags */
         util_cpu_caps.has_tsc    = (regs2[3] >>  4) & 1; /* 0x0000010 */
         util_cpu_caps.has_mmx    = (regs2[3] >> 23) & 1; /* 0x0800000 */
         util_cpu_caps.has_sse    = (regs2[3] >> 25) & 1; /* 0x2000000 */
         util_cpu_caps.has_sse2   = (regs2[3] >> 26) & 1; /* 0x4000000 */
         util_cpu_caps.has_sse3   = (regs2[2] >>  0) & 1; /* 0x0000001 */
         util_cpu_caps.has_ssse3  = (regs2[2] >>  9) & 1; /* 0x0000020 */
         util_cpu_caps.has_sse4_1 = (regs2[2] >> 19) & 1;
         util_cpu_caps.has_sse4_2 = (regs2[2] >> 20) & 1;
         util_cpu_caps.has_popcnt = (regs2[2] >> 23) & 1;
         util_cpu_caps.has_avx    = ((regs2[2] >> 28) & 1) && // AVX
                                    ((regs2[2] >> 27) & 1) && // OSXSAVE
                                    ((xgetbv() & 6) == 6);    // XMM & YMM
         util_cpu_caps.has_f16c   = ((regs2[2] >> 29) & 1) && util_cpu_caps.has_avx;
         util_cpu_caps.has_fma    = ((regs2[2] >> 12) & 1) && util_cpu_caps.has_avx;
         util_cpu_caps.has_mmx2   = util_cpu_caps.has_sse; /* SSE cpus supports mmxext too */
#if defined(PIPE_ARCH_X86_64)
         util_cpu_caps.has_daz = 1;
#else
         util_cpu_caps.has_daz = util_cpu_caps.has_sse3 ||
            (util_cpu_caps.has_sse2 && sse2_has_daz());
#endif

         cacheline = ((regs2[1] >> 8) & 0xFF) * 8;
         if (cacheline > 0)
            util_cpu_caps.cacheline = cacheline;
      }
      if (util_cpu_caps.has_avx && regs[0] >= 0x00000007) {
         uint32_t regs7[4];
         cpuid_count(0x00000007, 0x00000000, regs7);
         util_cpu_caps.has_avx2 = (regs7[1] >> 5) & 1;
      }

      // check for avx512
      if (((regs2[2] >> 27) & 1) && // OSXSAVE
          (xgetbv() & (0x7 << 5)) && // OPMASK: upper-256 enabled by OS
          ((xgetbv() & 6) == 6)) { // XMM/YMM enabled by OS
         uint32_t regs3[4];
         cpuid_count(0x00000007, 0x00000000, regs3);
         util_cpu_caps.has_avx512f    = (regs3[1] >> 16) & 1;
         util_cpu_caps.has_avx512dq   = (regs3[1] >> 17) & 1;
         util_cpu_caps.has_avx512ifma = (regs3[1] >> 21) & 1;
         util_cpu_caps.has_avx512pf   = (regs3[1] >> 26) & 1;
         util_cpu_caps.has_avx512er   = (regs3[1] >> 27) & 1;
         util_cpu_caps.has_avx512cd   = (regs3[1] >> 28) & 1;
         util_cpu_caps.has_avx512bw   = (regs3[1] >> 30) & 1;
         util_cpu_caps.has_avx512vl   = (regs3[1] >> 31) & 1;
         util_cpu_caps.has_avx512vbmi = (regs3[2] >>  1) & 1;
      }

      if (regs[1] == 0x756e6547 && regs[2] == 0x6c65746e && regs[3] == 0x49656e69) {
         /* GenuineIntel */
         util_cpu_caps.has_intel = 1;
      }

      cpuid(0x80000000, regs);

      if (regs[0] >= 0x80000001) {

         cpuid(0x80000001, regs2);

         util_cpu_caps.has_mmx  |= (regs2[3] >> 23) & 1;
         util_cpu_caps.has_mmx2 |= (regs2[3] >> 22) & 1;
         util_cpu_caps.has_3dnow = (regs2[3] >> 31) & 1;
         util_cpu_caps.has_3dnow_ext = (regs2[3] >> 30) & 1;

         util_cpu_caps.has_xop = util_cpu_caps.has_avx &&
                                 ((regs2[2] >> 11) & 1);
      }

      if (regs[0] >= 0x80000006) {
         /* should we really do this if the clflush size above worked? */
         unsigned int cacheline;
         cpuid(0x80000006, regs2);
         cacheline = regs2[2] & 0xFF;
         if (cacheline > 0)
            util_cpu_caps.cacheline = cacheline;
      }

      if (!util_cpu_caps.has_sse) {
         util_cpu_caps.has_sse2 = 0;
         util_cpu_caps.has_sse3 = 0;
         util_cpu_caps.has_ssse3 = 0;
         util_cpu_caps.has_sse4_1 = 0;
      }
   }
#endif /* PIPE_ARCH_X86 || PIPE_ARCH_X86_64 */

#if defined(PIPE_ARCH_ARM) || defined(PIPE_ARCH_AARCH64)
   check_os_arm_support();
#endif

#if defined(PIPE_ARCH_PPC)
   check_os_altivec_support();
#endif /* PIPE_ARCH_PPC */

#if defined(PIPE_ARCH_MIPS64)
   check_os_mips64_support();
#endif /* PIPE_ARCH_MIPS64 */

#if defined(PIPE_ARCH_S390)
   util_cpu_caps.family = CPU_S390X;
#endif

   get_cpu_topology();

   if (debug_get_option_dump_cpu()) {
      printf("util_cpu_caps.nr_cpus = %u\n", util_cpu_caps.nr_cpus);

      printf("util_cpu_caps.x86_cpu_type = %u\n", util_cpu_caps.x86_cpu_type);
      printf("util_cpu_caps.cacheline = %u\n", util_cpu_caps.cacheline);

      printf("util_cpu_caps.has_tsc = %u\n", util_cpu_caps.has_tsc);
      printf("util_cpu_caps.has_mmx = %u\n", util_cpu_caps.has_mmx);
      printf("util_cpu_caps.has_mmx2 = %u\n", util_cpu_caps.has_mmx2);
      printf("util_cpu_caps.has_sse = %u\n", util_cpu_caps.has_sse);
      printf("util_cpu_caps.has_sse2 = %u\n", util_cpu_caps.has_sse2);
      printf("util_cpu_caps.has_sse3 = %u\n", util_cpu_caps.has_sse3);
      printf("util_cpu_caps.has_ssse3 = %u\n", util_cpu_caps.has_ssse3);
      printf("util_cpu_caps.has_sse4_1 = %u\n", util_cpu_caps.has_sse4_1);
      printf("util_cpu_caps.has_sse4_2 = %u\n", util_cpu_caps.has_sse4_2);
      printf("util_cpu_caps.has_avx = %u\n", util_cpu_caps.has_avx);
      printf("util_cpu_caps.has_avx2 = %u\n", util_cpu_caps.has_avx2);
      printf("util_cpu_caps.has_f16c = %u\n", util_cpu_caps.has_f16c);
      printf("util_cpu_caps.has_popcnt = %u\n", util_cpu_caps.has_popcnt);
      printf("util_cpu_caps.has_3dnow = %u\n", util_cpu_caps.has_3dnow);
      printf("util_cpu_caps.has_3dnow_ext = %u\n", util_cpu_caps.has_3dnow_ext);
      printf("util_cpu_caps.has_xop = %u\n", util_cpu_caps.has_xop);
      printf("util_cpu_caps.has_altivec = %u\n", util_cpu_caps.has_altivec);
      printf("util_cpu_caps.has_vsx = %u\n", util_cpu_caps.has_vsx);
      printf("util_cpu_caps.has_neon = %u\n", util_cpu_caps.has_neon);
      printf("util_cpu_caps.has_msa = %u\n", util_cpu_caps.has_msa);
      printf("util_cpu_caps.has_daz = %u\n", util_cpu_caps.has_daz);
      printf("util_cpu_caps.has_avx512f = %u\n", util_cpu_caps.has_avx512f);
      printf("util_cpu_caps.has_avx512dq = %u\n", util_cpu_caps.has_avx512dq);
      printf("util_cpu_caps.has_avx512ifma = %u\n", util_cpu_caps.has_avx512ifma);
      printf("util_cpu_caps.has_avx512pf = %u\n", util_cpu_caps.has_avx512pf);
      printf("util_cpu_caps.has_avx512er = %u\n", util_cpu_caps.has_avx512er);
      printf("util_cpu_caps.has_avx512cd = %u\n", util_cpu_caps.has_avx512cd);
      printf("util_cpu_caps.has_avx512bw = %u\n", util_cpu_caps.has_avx512bw);
      printf("util_cpu_caps.has_avx512vl = %u\n", util_cpu_caps.has_avx512vl);
      printf("util_cpu_caps.has_avx512vbmi = %u\n", util_cpu_caps.has_avx512vbmi);
      printf("util_cpu_caps.num_L3_caches = %u\n", util_cpu_caps.num_L3_caches);
      printf("util_cpu_caps.num_cpu_mask_bits = %u\n", util_cpu_caps.num_cpu_mask_bits);
   }

   /* This must happen at the end as it's used to guard everything else */
   p_atomic_set(&util_cpu_caps.detect_done, 1);
}

static once_flag cpu_once_flag = ONCE_FLAG_INIT;

void
util_cpu_detect(void)
{
   call_once(&cpu_once_flag, util_cpu_detect_once);
}