fasthash-sys-fork 0.4.2

/*
 *  Copyright (c) 2001-2020 Leonid Yuriev <leo@yuriev.ru>,
 *  The "Mera" library.
 *
 *  This software is provided 'as-is', without any express or implied
 *  warranty. In no event will the authors be held liable for any damages
 *  arising from the use of this software.
 *
 *  Permission is granted to anyone to use this software for any purpose,
 *  including commercial applications, and to alter it and redistribute it
 *  freely, subject to the following restrictions:
 *
 *  1. The origin of this software must not be misrepresented; you must not
 *     claim that you wrote the original software. If you use this software
 *     in a product, an acknowledgement in the product documentation would be
 *     appreciated but is not required.
 *  2. Altered source versions must be plainly marked as such, and must not be
 *     misrepresented as being the original software.
 *  3. This notice may not be removed or altered from any source distribution.
 */

#ifndef NOMINMAX
#define NOMINMAX
#endif

#ifndef _ISOC99_SOURCE
#define _ISOC99_SOURCE 1
#endif

#ifndef _ISOC11_SOURCE
#define _ISOC11_SOURCE 1
#endif

#ifndef _POSIX_C_SOURCE
#define _POSIX_C_SOURCE 200809L
#endif

#ifndef _XOPEN_SOURCE
#define _XOPEN_SOURCE 700
#endif

#ifndef _GNU_SOURCE
#define _GNU_SOURCE 1
#endif

#ifndef __STDC_LIMIT_MACROS
#define __STDC_LIMIT_MACROS 1
#endif

#ifndef _THREAD_SAFE
#define _THREAD_SAFE 1
#endif

#ifndef _REENTRANT
#define _REENTRANT 1
#endif

#if defined(_MSC_VER)
#pragma warning(disable : 4711) /* function 'xyz' selected for                 \
                                   automatic inline expansion */
#pragma warning(disable : 4127) /* conditional expression is constant */
#pragma warning(disable : 4702) /* unreachable code */
#if _MSC_VER < 1900
#define snprintf _snprintf
#pragma warning(disable : 4996) /* '_snprintf': This function or variable      \
                                   may be unsafe */
#endif
#if _MSC_VER > 1800
#pragma warning(disable : 4464) /* relative include path contains '..' */
#endif
#endif /* MSVC */

/* OS's includes for time/clock */
#if defined(__linux__) || defined(__gnu_linux__)
#include <linux/hw_breakpoint.h>
#include <linux/perf_event.h>
#include <sched.h>
#include <sys/prctl.h>
#include <sys/syscall.h>
#endif /* Linux */

#if defined(__FreeBSD__)
#include <sys/time.h>
#include <unistd.h>
#endif /* FreeBSD */

#if defined(EMSCRIPTEN)
#include <emscripten.h>
#elif defined(__APPLE__) || defined(__MACH__)
#include <mach/mach_time.h>
#include <mach/thread_policy.h>
#endif

#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) ||              \
    defined(__WINDOWS__)
#include <time.h>
#include <windows.h>
#include <winnt.h>
#else
#include <errno.h>
#include <fcntl.h>
#include <setjmp.h>
#include <signal.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/time.h>
#include <time.h>
#include <unistd.h>
#endif /* OS */

#include "mera.h"
#include <memory.h>
#include <stdio.h>
#include <stdlib.h>

#include "common.h"

/*****************************************************************************/

/* Compiler's includes for builtins/intrinsics */
#if defined(_MSC_VER)
#include <intrin.h>
#elif defined(__INTEL_COMPILER)
#include <ia32intrin.h>
#elif __GNUC_PREREQ(4, 4) || defined(__clang__)
#if defined(__ia32__) || defined(__e2k__)
#include <x86intrin.h>
#endif /* __ia32__ */
#if defined(__ia32__)
#include <cpuid.h>
#endif /* __ia32__ */
#elif defined(__SUNPRO_C) || defined(__sun) || defined(sun)
#include <mbarrier.h>
#elif (defined(_HPUX_SOURCE) || defined(__hpux) || defined(__HP_aCC)) &&       \
    (defined(HP_IA64) || defined(__ia64))
#include <machine/sys/inline.h>
#elif defined(__IBMC__) && defined(__powerpc)
#include <atomic.h>
#elif defined(_AIX)
#include <builtins.h>
#include <sys/atomic_op.h>
#elif (defined(__osf__) && defined(__DECC)) || defined(__alpha)
#include <c_asm.h>
#include <machine/builtins.h>
#elif defined(__MWERKS__)
/* CodeWarrior - troubles ? */
#pragma gcc_extensions
#elif defined(__SNC__)
/* Sony PS3 - troubles ? */
#elif defined(__hppa__) || defined(__hppa)
#include <machine/inline.h>
#else
#error Unsupported C compiler, please use GNU C 4.4 or newer
#endif /* Compiler */

static __inline void compiler_barrier(void) {
#if defined(__clang__) || defined(__GNUC__)
  __asm__ __volatile__("" ::: "memory");
#elif defined(_MSC_VER)
  _ReadWriteBarrier();
#elif defined(__INTEL_COMPILER) /* LY: Intel Compiler may mimic GCC and MSC */
  __memory_barrier();
  if (type > MDBX_BARRIER_COMPILER)
#if defined(__ia64__) || defined(__ia64) || defined(_M_IA64)
    __mf();
#elif defined(__i386__) || defined(__x86_64__)
    _mm_mfence();
#else
#error "Unknown target for Intel Compiler, please report to us."
#endif
#elif defined(__SUNPRO_C) || defined(__sun) || defined(sun)
  __compiler_barrier();
#elif (defined(_HPUX_SOURCE) || defined(__hpux) || defined(__HP_aCC)) &&       \
    (defined(HP_IA64) || defined(__ia64))
  _Asm_sched_fence(/* LY: no-arg meaning 'all expect ALU', e.g. 0x3D3D */);
#elif defined(_AIX) || defined(__ppc__) || defined(__powerpc__) ||             \
    defined(__ppc64__) || defined(__powerpc64__)
  __fence();
#else
#error "Could not guess the kind of compiler, please report to us."
#endif
}

#ifndef likely
#if defined(__GNUC__) || defined(__clang__)
#define likely(cond) __builtin_expect(!!(cond), 1)
#else
#define likely(x) (x)
#endif
#endif /* likely */

#ifndef unlikely
#if defined(__GNUC__) || defined(__clang__)
#define unlikely(cond) __builtin_expect(!!(cond), 0)
#else
#define unlikely(x) (x)
#endif
#endif /* unlikely */

/*****************************************************************************/

#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) ||              \
    defined(__WINDOWS__)
static unsigned seh_filter(unsigned exception_code) {
  switch (exception_code) {
  case EXCEPTION_ILLEGAL_INSTRUCTION:
  case EXCEPTION_PRIV_INSTRUCTION:
  case EXCEPTION_ACCESS_VIOLATION:
    return EXCEPTION_EXECUTE_HANDLER;
  }
  return EXCEPTION_CONTINUE_SEARCH;
}
#else
static sigjmp_buf sigaction_jump;
static void sigaction_handler(int signum, siginfo_t *info, void *context) {
  (void)context;
  (void)info;
  siglongjmp(sigaction_jump, signum);
}
#endif

static void minimal_sleep(void) {
#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__)
  Sleep(1);
#else
  struct timespec ts;
  ts.tv_sec = 0;
  ts.tv_nsec = 54321;
  nanosleep(&ts, 0);
#endif
}

/* LY: dedicated function to avoid clobber args by ‘longjmp’ */
static int do_probe(unsigned (*start)(timestamp_t *),
                    unsigned (*finish)(timestamp_t *)) {
#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) ||              \
    defined(__WINDOWS__)
  __try {
#else
  struct sigaction act, prev_sigsegv, prev_sigill, prev_sigbus;
  memset(&act, 0, sizeof(act));
  act.sa_sigaction = sigaction_handler;
  if (sigaction(SIGSEGV, &act, &prev_sigsegv)) {
    perror(MERA_PERROR_PREFIX "sigaction(SIGSEGV)");
    return -1;
  }
  if (sigaction(SIGILL, &act, &prev_sigill)) {
    perror(MERA_PERROR_PREFIX "sigaction(SIGILL)");
    return -1;
  }
  if (sigaction(SIGBUS, &act, &prev_sigbus)) {
    perror(MERA_PERROR_PREFIX "sigaction(SIGBUS)");
    return -1;
  }

  if (sigsetjmp(sigaction_jump, 1) != 0) {
    sigaction(SIGSEGV, &prev_sigsegv, NULL);
    sigaction(SIGILL, &prev_sigill, NULL);
    sigaction(SIGBUS, &prev_sigbus, NULL);
    return -2;
  }
#endif

    for (unsigned n = 0; n < 42; ++n) {
      timestamp_t timestamp_start, timestamp_finish;
      unsigned coreid = start(&timestamp_start);
      minimal_sleep();
      if (coreid != finish(&timestamp_finish))
        continue;
      if (timestamp_finish > timestamp_start)
        return 1;
      if (timestamp_finish == timestamp_start || n > 5)
        break;
    }

#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) ||              \
    defined(__WINDOWS__)
  } __except (seh_filter(GetExceptionCode())) {
    return -2;
  }
#else
  sigaction(SIGSEGV, &prev_sigsegv, NULL);
  sigaction(SIGILL, &prev_sigill, NULL);
  sigaction(SIGBUS, &prev_sigbus, NULL);
#endif
  return 0;
}

static bool probe(unsigned (*start)(timestamp_t *),
                  unsigned (*finish)(timestamp_t *),
                  double (*convert)(timestamp_t), unsigned flags,
                  const char *source_name, const char *time_units) {

  if (is_option_set(bench_verbose)) {
    printf(" - probe for %s", source_name);
    fflush(stdout);
  }

  flags |= timestamp_clock_have;
  if (mera.flags >= flags) {
    if (is_option_set(bench_verbose))
      printf(": Skip (already have)\n");
    return false;
  }

  int rc = do_probe(start, finish);
  switch (rc) {
  case 1:
    if (is_option_set(bench_verbose))
      printf(": Ok\n");
    mera.start = start;
    mera.finish = finish;
    mera.source = source_name;
    mera.convert = convert;
    if (flags & timestamp_cycles)
      mera.units = "cycle";
    else if (flags & timestamp_ticks)
      mera.units = "tick";
    else
      mera.units = time_units;
    mera.flags = flags;
    return true;
  case 0:
    if (is_option_set(bench_verbose))
      printf(": Doesnt work\n");
    break;
  case -2:
    if (is_option_set(bench_verbose))
      printf(": Not available (SIGSEGV/SIGILL)\n");
    break;
  }
  return false;
}

/*****************************************************************************/

static int set_single_affinity(void) {
#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) ||              \
    defined(__WINDOWS__)
  return -1;
#elif defined(__GLIBC__) || defined(__GNU_LIBRARY__) || defined(__ANDROID__)
  const int current_cpu = sched_getcpu();
  if (current_cpu < 0) {
    perror(MERA_PERROR_PREFIX "sched_getcpu()");
    return -1;
  }
  const int ncpu = sysconf(_SC_NPROCESSORS_CONF);
  const unsigned cpuset_size = CPU_ALLOC_SIZE(ncpu);
  cpu_set_t *affinity = CPU_ALLOC(ncpu);
  if (!affinity) {
    perror(MERA_PERROR_PREFIX "CPU_ALLOC()");
    return -1;
  }
  CPU_ZERO_S(cpuset_size, affinity);
  CPU_SET_S(current_cpu, cpuset_size, affinity);
  if (sched_setaffinity(0, cpuset_size, affinity)) {
    perror(MERA_PERROR_PREFIX "sched_setaffinity()");
    CPU_FREE(affinity);
    return -1;
  }
  CPU_FREE(affinity);
  return current_cpu;
#elif defined(__APPLE__) || defined(__MACH__)
  return -1;
#else
  return -1;
#endif
}

/*****************************************************************************/

union timestamp {
  uint64_t u64;
  struct {
#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
    uint32_t h, l;
#else
    uint32_t l, h;
#endif
  } u32;
};

#ifndef __has_builtin
#define __has_builtin(x) (0)
#endif

#if __has_builtin(__builtin_readcyclecounter)
static unsigned clock_readcyclecounter(timestamp_t *now) {
  *now = __builtin_readcyclecounter();
  return 0;
}
#endif /* __builtin_readcyclecounter */

#if defined(EMSCRIPTEN)
static unsigned clock_emscripten(timestamp_t *now) {
  compiler_barrier();
  *now = (timestamp_t)(emscripten_get_now() * 1e6);
  compiler_barrier();
  return 0;
}
#endif /* EMSCRIPTEN */

#if defined(__APPLE__) || defined(__MACH__)
static unsigned clock_mach(timestamp_t *now) {
  compiler_barrier();
  *now = mach_absolute_time();
  compiler_barrier();
  return 0;
}

static double convert_mach(timestamp_t timestamp) {
  static double ratio /* from mach_absolute_time() to seconds */;
  if (!ratio) {
    mach_timebase_info_data_t ti;
    if (mach_timebase_info(&ti) != 0) {
      perror(MERA_PERROR_PREFIX "mach_timebase_info()");
      return -1;
    }
    ratio = (double)ti.numer / ti.denom;
  }
  return ratio * (double)timestamp;
}
#endif /* defined(__APPLE__) || defined(__MACH__) */

#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) ||              \
    defined(__WINDOWS__)
static unsigned clock_windows(timestamp_t *now) {
  compiler_barrier();
  if (!QueryPerformanceCounter((LARGE_INTEGER *)now)) {
    perror(MERA_PERROR_PREFIX "QueryPerformanceCounter()");
    *now = 42;
  }
  compiler_barrier();
  return 0;
}

static double convert_windows(timestamp_t timestamp) {
  static double ratio /* from QueryPerformanceCounter() to seconds */;
  if (!ratio) {
    LARGE_INTEGER frequency;
    if (!QueryPerformanceFrequency(&frequency)) {
      perror(MERA_PERROR_PREFIX "QueryPerformanceFrequency()");
      return -1;
    }
    ratio = 1e9 / frequency.QuadPart;
  }
  return ratio * (double)timestamp;
}

#else /* Windows */

static unsigned clock_gettimeofday(timestamp_t *now) {
  compiler_barrier();
  struct timeval tv;
  if (gettimeofday(&tv, NULL)) {
    perror(MERA_PERROR_PREFIX "gettimeofday()");
    tv.tv_sec = tv.tv_usec = 0;
  }
  *now = tv.tv_sec * UINT64_C(1000000) + tv.tv_usec;
  compiler_barrier();
  return 0;
}

static double convert_us2ns(timestamp_t timestamp) { return 1e3 * timestamp; }

#endif /* ! Windows */

#if defined(TIMEBASE_SZ) || defined(__OS400__)
static unsigned clock_os400(timestamp_t *now) {
  compiler_barrier();
  timebasestruct_t tb;
  if (read_wall_time(&tb, TIMEBASE_SZ) != 0) {
    perror(MERA_PERROR_PREFIX "read_wall_time(TIMEBASE_SZ)");
    abort();
  }
  union timestamp *u = (union timestamp *)now;
  u->u32.h = tb.tb_high;
  u->u32.l = tb.tb_low;
  compiler_barrier();
  return 0;
}

static double convert_os400(timestamp_t timestamp) {
  static double ratio /* from read_wall_time() to seconds */;
  if (!ratio) {
    timebasestruct_t tb;
    tb.tb_high = 0x7fff;
    tb.tb_low = 0;
    if (time_base_to_time(&tb, TIMEBASE_SZ) != 0) {
      perror(MERA_PERROR_PREFIX "time_base_to_time()");
      abort();
    }
    ratio = (tb.tb_high * 1e9 + tb.tb_low) / UINT64_C(0x7fff00000000);
  }
  return ratio * (double)timestamp;
}
#endif /* __OS400__ */

#if defined(CLOCK_MONOTONIC) || defined(CLOCK_MONOTONIC_RAW) ||                \
    defined(CLOCK_SGI_CYCLE)
static clockid_t posix_clockid = CLOCK_REALTIME;
static unsigned clock_posix(timestamp_t *now) {
  compiler_barrier();
  struct timespec ts;
  if (clock_gettime(posix_clockid, &ts)) {
    perror(MERA_PERROR_PREFIX "clock_gettime()");
    ts.tv_sec = ts.tv_nsec = 0;
  }
  *now = ts.tv_sec * UINT64_C(1000000000) + ts.tv_nsec;
  compiler_barrier();
  return 0;
}
#endif /* CLOCK_MONOTONIC || CLOCK_MONOTONIC_RAW || CLOCK_SGI_CYCLE */

#if defined(__sun__) || defined(__sun)
static unsigned clock_solaris(timestamp_t *now) {
  compiler_barrier();
  *now = gethrtime();
  compiler_barrier();
}
#endif /* __sun__ */

/*****************************************************************************/

#if defined(__e2k__) || defined(__elbrus__)
static unsigned clock_elbrus(timestamp_t *now) {
  compiler_barrier();
  unsigned coreid;
  *now = __rdtscp(&coreid);
  compiler_barrier();
  return coreid;
}
#endif /* __e2k__ || __elbrus__ */

#if (defined(__powerpc64__) || defined(__ppc64__) || defined(__ppc64) ||       \
     defined(__powerpc64))
static unsigned clock_powerpc64_mfspr268(timestamp_t *now) {
  compiler_barrier();
#if defined(__GNUC__)
  uint64_t ticks;
  __asm __volatile("mfspr %0, 268" : "=r"(ticks));
  *now = ticks;
#else
  *now = 42 /* FIXME */;
#endif
  compiler_barrier();
  return 0;
}
#endif /* __powerpc64__  */

#if (defined(__powerpc__) || defined(__ppc__) || defined(__powerpc) ||         \
     defined(__ppc))
static unsigned clock_powerpc_mftb(timestamp_t *now) {
  /* A time-base timer, which is not always precisely a cycle-count. */
  compiler_barrier();
#if UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul
#if defined(__GNUC__)
  uint64_t ticks;
  __asm __volatile("mftb  %0" : "=r"(ticks));
  *now = ticks;
#else
  *now = 42 /* FIXME */;
#endif

#else

#if defined(__GNUC__)
  uint32_t low, high_before, high_after;
  __asm __volatile("mftbu %0; mftb  %1; mftbu %2"
                   : "=r"(high_before), "=r"(low), "=r"(high_after));
  union timestamp *u = (union timestamp *)now;
  u->u32.h = high_after;
  u->u32.l = low & /* zeroes if high part has changed */
             ~(high_before - high_after);
#else
  *now = 42 /* FIXME */;
#endif
#endif
  compiler_barrier();
  return 0;
}
#endif /* __powerpc__ */

#if defined(__sparc__) || defined(__sparc) || defined(__sparc64__) ||          \
    defined(__sparc64) || defined(__sparc_v8plus__) ||                         \
    defined(__sparc_v8plus) || defined(__sparc_v8plusa__) ||                   \
    defined(__sparc_v8plusa) || defined(__sparc_v9__) || defined(__sparc_v9)
static unsigned clock_sparc(timestamp_t *now) {
  compiler_barrier();
  union timestamp cycles;
#ifndef __GNUC__
#warning FIXME
#else

#if defined(__sparc_v8plus__) || defined(__sparc_v8plusa__) ||                 \
    defined(__sparc_v9__) || defined(__sparc_v8plus) ||                        \
    defined(__sparc_v8plusa) || defined(__sparc_v9)

#if UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul ||                  \
    defined(__sparc64__) || defined(__sparc64)
  __asm __volatile("rd %%tick, %0" : "=r"(cycles.u64));
#else
  __asm __volatile("rd %%tick, %1; srlx %1, 32, %0"
                   : "=r"(cycles.u32.h), "=r"(cycles.u32.l));
#endif /* __sparc64__ */

#else
  __asm __volatile(".byte 0x83, 0x41, 0x00, 0x00; mov %%g1, %0"
                   : "=r"(cycles.u64)
                   :
                   : "%g1");
#endif /* __sparc8plus__ || __sparc_v9__ */
#endif /* GCC */
  *now = cycles.u64;
  compiler_barrier();
  return 0;
}
#endif /* __sparc__ */

#if defined(__ia64__) || defined(__ia64)
static unsigned clock_ia64(timestamp_t *now) {
  compiler_barrier();
#if defined(__GNUC__)
  uint64_t ticks;
  __asm __volatile("mov %0 = ar.itc" : "=r"(ticks));
  *now = ticks;
#elif defined(__EDG_VERSION) || defined(__ECC) || defined(IA64_REG_AR_ITC)
  *now = __getReg(_IA64_REG_AR_ITC);
#elif defined(__hpux) || defined(_AREG_ITC)
  *now = _Asm_mov_from_ar(_AREG_ITC);
#else
  *now = 42 /* FIXME */;
#endif
  compiler_barrier();
  return 0;
}
#endif /* __ia64__ */

#if (defined(__hppa__) || defined(__hppa) || defined(__hppa64__) ||            \
     defined(__hppa64))
static unsigned clock_hppa(timestamp_t *now) {
  compiler_barrier();
  uint64_t cycles;
#ifdef __GNUC__
  __asm __volatile("mfctl 16, %0" : "=r"(cycles));
#else
  _MFCTL(16, ticks);
#endif
  *now = cycles;
  compiler_barrier();
  return 0;
}
#endif /* __hppa__ */

#if defined(__s390__) || defined(__s390) || defined(__zarch__) ||              \
    defined(__zarch)
static unsigned clock_stcke(timestamp_t *now) {
  compiler_barrier();
  uint8_t clk[16];
#ifdef __GNUC__
  __asm __volatile("stcke %0" : "=Q"(clk) : : "cc");
#else
#warning FIXME
#endif
  *now = *((unsigned long long *)&clk[1]) >> 2;
  compiler_barrier();
  return (clk[14] << 8) | (clk[15]);
}

static unsigned clock_stckf(timestamp_t *now) {
  compiler_barrier();
#ifdef __GNUC__
  __asm __volatile("stckf 0(%1)" : "=m"(*now) : "a"(now) : "cc");
#else
#warning FIXME
#endif
  compiler_barrier();
  return 0;
}

static unsigned clock_stck(timestamp_t *now) {
  compiler_barrier();
#ifdef __GNUC__
  __asm __volatile("stck 0(%1)" : "=m"(*now) : "a"(now) : "cc");
#else
#warning FIXME
#endif
  compiler_barrier();
  return 0;
}
#endif /* __s390__  */

#if defined(__alpha__) || defined(__alpha)
static unsigned clock_alpha(timestamp_t *now) {
  compiler_barrier();
#ifdef __GNUC__
  unsigned long cycles;
  __asm__ __volatile("rpcc %0" : "=r"(cycles));
  *now = cycles & 0xFFFFfffful;
#else
  *now = 42 /* FIXME */;
#endif
  compiler_barrier();
  return 0;
}
#endif /* __alpha__ */

/*****************************************************************************/

static double convert_1to1(timestamp_t timestamp) { return (double)timestamp; }

#if (defined(__ARM_ARCH) && __ARM_ARCH > 5 && __ARM_ARCH < 8) || defined(_M_ARM)
static unsigned clock_pmccntr(timestamp_t *now) {
  compiler_barrier();
#ifdef _M_ARM
  *now = __rdpmccntr64();
#else
  unsigned long pmccntr;
  __asm __volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(pmccntr));
  *now = (uint64_t)pmccntr;
#endif
  compiler_barrier();
  return 0;
}

static double convert_pmccntr_x64(timestamp_t timestamp) {
  /* The counter is set up to count every 64th cycle */
  return timestamp * 64.0;
}
#endif /* __ARM_ARCH >= 6 || _M_ARM */

#if defined(__aarch64__) || (defined(__ARM_ARCH) && __ARM_ARCH > 7) ||         \
    defined(_M_ARM64)
static unsigned clock_cntvct_el0(timestamp_t *now) {
  compiler_barrier();
/* System timer of ARMv8 runs at a different frequency than the CPU's.
 * The frequency is fixed, typically in the range 1-50MHz.  It can be
 * read at CNTFRQ special register.  We assume the OS has set up
 * the virtual timer properly. */
#ifdef _M_ARM64
  *now = _ReadStatusReg(42 /* FIXME: cntvct_el0 */);
#else
  uint64_t virtual_timer;
  __asm __volatile("mrs %0, cntvct_el0" : "=r"(virtual_timer));
  *now = virtual_timer;
#endif
  compiler_barrier();
  return 0;
}
#endif /* __aarch64__ || __ARM_ARCH > 7 || _M_ARM64 */

#if defined(__mips__) || defined(__mips) || defined(_R4000)

#if defined(PROT_READ) && defined(MAP_SHARED)
static volatile uint64_t *mips_tsc_addr;
static unsigned clock_zbustimer(timestamp_t *now) {
  compiler_barrier();
  *now = *mips_tsc_addr;
  compiler_barrier();
  return 0;
}
#endif /* PROT_READ && MAP_SHARED */

#if (defined(_MIPS_ISA) && defined(_MIPS_ISA_MIPS2) &&                         \
     _MIPS_ISA >= _MIPS_ISA_MIPS2) ||                                          \
    (defined(__mips) && __mips >= 2 && __mips < 16) || defined(_R4000) ||      \
    defined(__MIPS_ISA2) || defined(__MIPS_ISA3) || defined(__MIPS_ISA4) ||    \
    (defined(__mips_isa_rev) && __mips_isa_rev >= 2)

static unsigned clock_mfc0_25_1(timestamp_t *now) {
  compiler_barrier();
#if (defined(_MIPS_SIM) && defined(_ABI64) && _MIPS_SIM == _ABI64) ||          \
    (defined(_MIPS_SIM) && defined(_ABIO64) && _MIPS_SIM == _ABIO64) ||        \
    defined(__mips64) || defined(__mips64__) ||                                \
    (defined(__mips) && (__mips >= 64))
  uint64_t count;
  __asm __volatile("dmfc0 %0, $25, 1" : "=r"(count));
#else
  uint32_t count;
  __asm __volatile("mfc0 %0, $25, 1" : "=r"(count));
#endif
  *now = count;
  compiler_barrier();
  return 0;
}

static unsigned clock_mfc0_9_0(timestamp_t *now) {
  compiler_barrier();
#if (defined(_MIPS_SIM) && defined(_ABI64) && _MIPS_SIM == _ABI64) ||          \
    (defined(_MIPS_SIM) && defined(_ABIO64) && _MIPS_SIM == _ABIO64) ||        \
    defined(__mips64) || defined(__mips64__) ||                                \
    (defined(__mips) && (__mips >= 64))
  uint64_t count;
  __asm __volatile("dmfc0 %0, $9, 0" : "=r"(count));
#else
  uint32_t count;
  __asm __volatile("mfc0 %0, $9, 0" : "=r"(count));
#endif
  *now = count;
  compiler_barrier();
  return 0;
}

static unsigned mips_rdhwr_resolution;
static unsigned clock_rdhwr(timestamp_t *now) {
  compiler_barrier();
  unsigned count, coreid;
  __asm __volatile("rdhwr %0, $2; rdhwr %1, $0" : "=r"(count), "=r"(coreid));
  *now = count;
  compiler_barrier();
  return coreid;
}

static double convert_rdhwr(timestamp_t timestamp) {
  return (double)timestamp * mips_rdhwr_resolution;
}
#endif /* MIPS >= 2 */

#endif /* MIPS */

#if defined(__ia32__)

enum ia32_fixed_perfomance_counters {
  /* count of retired instructions on the current core in the low-order 48 bits
     of an unsigned 64-bit integer */
  ia32_COUNT_HW_INSTRUCTIONS = 1 << 30,

  /* count of actual CPU core cycles executed by the current core.  Core cycles
     are not accumulated while the processor is in the "HALT" state, which is
     used when the operating system has no task(s) to run on a processor core.
     */
  ia32_COUNT_HW_CPU_CYCLES = (1 << 30) + 1,

  /* count of "reference" (or "nominal") CPU core cycles executed by the current
     core.  This counts at the same rate as the TSC, but does not count when the
     core is in the "HALT" state.  If a timed section of code shows a larger
     change in TSC than in rdpmc_reference_cycles, the processor probably spent
     some time in a HALT state. */
  ia32_COUNT_HW_REF_CPU_CYCLES = (1 << 30) + 2,
};

static unsigned clock_rdpmc_start(timestamp_t *now) {
  compiler_barrier();
#if __GNUC__
  uint32_t low, high;
  __asm __volatile("cpuid" ::: "%eax", "%ebx", "%ecx", "%edx");
  __asm __volatile("rdpmc"
                   : "=a"(low), "=d"(high)
                   : "c"(ia32_COUNT_HW_CPU_CYCLES));
  union timestamp *u = (union timestamp *)now;
  u->u32.l = low;
  u->u32.h = high;
#elif defined(_MSC_VER)
  int unused[4];
  __cpuid(unused, 0);
  *now = __readpmc(ia32_COUNT_HW_CPU_CYCLES);
#else
#error "FIXME: Unsupported compiler"
#endif
  compiler_barrier();
  return 0;
}

static unsigned clock_rdpmc_finish(timestamp_t *now) {
  compiler_barrier();
#if __GNUC__
  uint32_t low, high;
  __asm __volatile("mov %2, %%ecx; rdpmc; mov %%eax, %0; mov %%edx, %1; cpuid"
                   : "=r"(low), "=r"(high)
                   : "i"(ia32_COUNT_HW_CPU_CYCLES)
                   : "%eax", "%ebx", "%ecx", "%edx");
  union timestamp *u = (union timestamp *)now;
  u->u32.l = low;
  u->u32.h = high;
#elif defined(_MSC_VER)
  *now = __readpmc(ia32_COUNT_HW_CPU_CYCLES);
  int unused[4];
  __cpuid(unused, 0);
#else
#error "FIXME: Unsupported compiler"
#endif
  return 0;
}

static unsigned clock_rdtscp_start(timestamp_t *now) {
  compiler_barrier();
  unsigned coreid;
  *now = __rdtscp(&coreid);
  return coreid;
}

static unsigned clock_rdtscp_finish(timestamp_t *now) {
  compiler_barrier();
#if __GNUC__
  uint32_t low, high, coreid;
  __asm __volatile("rdtscp; mov %%eax, %0; mov %%edx, %1; mov %%ecx, %2; cpuid"
                   : "=r"(low), "=r"(high), "=r"(coreid)
                   :
                   : "%eax", "%ebx", "%ecx", "%edx");
  union timestamp *u = (union timestamp *)now;
  u->u32.l = low;
  u->u32.h = high;
  return coreid;
#elif defined(_MSC_VER)
  unsigned coreid;
  *now = __rdtscp(&coreid);
  int unused[4];
  __cpuid(unused, 0);
  return coreid;
#else
#error "FIXME: Unsupported compiler"
#endif
}

static unsigned clock_rdtsc_start(timestamp_t *now) {
  compiler_barrier();
#if __GNUC__
  uint32_t low, high;
  __asm __volatile("cpuid; rdtsc" : "=a"(low), "=d"(high) : : "%ebx", "%ecx");
  union timestamp *u = (union timestamp *)now;
  u->u32.l = low;
  u->u32.h = high;
#elif defined(_MSC_VER)
  int unused[4];
  __cpuid(unused, 0);
  *now = __rdtsc();
#else
#error "FIXME: Unsupported compiler"
#endif
  compiler_barrier();
  return 0;
}

static unsigned clock_rdtsc_finish(timestamp_t *now) {
  compiler_barrier();
#if __GNUC__
  uint32_t low, high;
  __asm __volatile("rdtsc; mov %%eax, %0; mov %%edx, %1; cpuid"
                   : "=r"(low), "=r"(high)
                   :
                   : "%eax", "%ebx", "%ecx", "%edx");
  union timestamp *u = (union timestamp *)now;
  u->u32.l = low;
  u->u32.h = high;
#elif defined(_MSC_VER)
  int unused[4];
  __cpuid(unused, 0);
  *now = __rdtsc();
#else
#error "FIXME: Unsupported compiler"
#endif
  compiler_barrier();
  return 0;
}

ia32_cpu_features_t ia32_cpu_features;

/* Crutch for Intel Compiler (copy&paste from GCC's cpuid.h file */
#if defined(__INTEL_COMPILER) && defined(__GNUC__) && !defined(__cpuid)
#define __cpuid(level, a, b, c, d)                                             \
  __asm__("cpuid\n\t" : "=a"(a), "=b"(b), "=c"(c), "=d"(d) : "0"(level))

#define __cpuid_count(level, count, a, b, c, d)                                \
  __asm__("cpuid\n\t"                                                          \
          : "=a"(a), "=b"(b), "=c"(c), "=d"(d)                                 \
          : "0"(level), "2"(count))
static __inline unsigned int __get_cpuid_max(unsigned int __ext,
                                             unsigned int *__sig) {
  unsigned int __eax, __ebx, __ecx, __edx;

#ifndef __x86_64__
  /* See if we can use cpuid.  On AMD64 we always can.  */
  __asm__("pushf{l|d}\n\t"
          "pushf{l|d}\n\t"
          "pop{l}\t%0\n\t"
          "mov{l}\t{%0, %1|%1, %0}\n\t"
          "xor{l}\t{%2, %0|%0, %2}\n\t"
          "push{l}\t%0\n\t"
          "popf{l|d}\n\t"
          "pushf{l|d}\n\t"
          "pop{l}\t%0\n\t"
          "popf{l|d}\n\t"
          : "=&r"(__eax), "=&r"(__ebx)
          : "i"(0x00200000));

  if (!((__eax ^ __ebx) & 0x00200000))
    return 0;
#endif /* __x86_64__ */

  /* Host supports cpuid.  Return highest supported cpuid input value.  */
  __cpuid(__ext, __eax, __ebx, __ecx, __edx);

  if (__sig)
    *__sig = __ebx;

  return __eax;
}
#endif /* Crutch for Intel Compiler (copy&paste from GCC's cpuid.h file */

void ia32_fetch_cpu_features(void) {
  memset(&ia32_cpu_features, 0, sizeof(ia32_cpu_features));
#ifdef __GNUC__
  uint32_t unused_eax, unused_ebx, cpuid_max;

  cpuid_max = __get_cpuid_max(0, NULL);
  if (cpuid_max >= 1) {
    __cpuid_count(1, 0, unused_eax, ia32_cpu_features.basic.ebx,
                  ia32_cpu_features.basic.ecx, ia32_cpu_features.basic.edx);
    if (cpuid_max >= 7)
      __cpuid_count(7, 0, unused_eax, ia32_cpu_features.extended_7.ebx,
                    ia32_cpu_features.extended_7.ecx,
                    ia32_cpu_features.extended_7.edx);
  }
  cpuid_max = __get_cpuid_max(0x80000000, NULL);
  if (cpuid_max >= 0x80000001) {
    __cpuid_count(0x80000001, 0, unused_eax, unused_ebx,
                  ia32_cpu_features.extended_80000001.ecx,
                  ia32_cpu_features.extended_80000001.edx);
    if (cpuid_max >= 0x80000007)
      __cpuid_count(0x80000007, 0, unused_eax, unused_ebx,
                    ia32_cpu_features.extended_80000007.ecx,
                    ia32_cpu_features.extended_80000007.edx);
  }

#elif defined(_MSC_VER)
  int info[4];
  __cpuid(info, 0);
  unsigned cpuid_max = info[0];
  if (cpuid_max >= 1) {
    __cpuidex(info, 1, 0);
    ia32_cpu_features.basic.ebx = info[1];
    ia32_cpu_features.basic.ecx = info[2];
    ia32_cpu_features.basic.edx = info[3];
    if (cpuid_max >= 7) {
      __cpuidex(info, 7, 0);
      ia32_cpu_features.extended_7.ebx = info[1];
      ia32_cpu_features.extended_7.ecx = info[2];
      ia32_cpu_features.extended_7.edx = info[3];
    }
  }

  __cpuid(info, 0x80000000);
  cpuid_max = info[0];
  if (cpuid_max >= 0x80000001) {
    __cpuidex(info, 0x80000001, 0);
    ia32_cpu_features.extended_80000001.ecx = info[2];
    ia32_cpu_features.extended_80000001.edx = info[3];
    if (cpuid_max >= 0x80000007) {
      __cpuidex(info, 0x80000007, 0);
      ia32_cpu_features.extended_80000007.ecx = info[2];
      ia32_cpu_features.extended_80000007.edx = info[3];
    }
  }
#else
#error "FIXME: Unsupported compiler"
#endif
}

#endif /* __ia32__ */

/*****************************************************************************/

#ifdef __NR_perf_event_open
static int perf_fd, perf_error;
#if defined(__ia32__)
static const struct perf_event_mmap_page volatile *perf_page;
#else
#define perf_page NULL
#endif
static long perf_event_open(struct perf_event_attr *event_attr, pid_t pid,
                            int cpu, int group_fd, unsigned long flags) {
  return syscall(__NR_perf_event_open, event_attr, pid, cpu, group_fd, flags);
}

static unsigned clock_perf(timestamp_t *now) {
  *now = 42;
  return read(perf_fd, now, sizeof(timestamp_t));
}

static int perf_setup(void) {
#ifdef PR_TASK_PERF_EVENTS_ENABLE
  if (prctl(PR_TASK_PERF_EVENTS_ENABLE, 1, 0, 0, 0))
    perror(MERA_PERROR_PREFIX "prctl(PR_TASK_PERF_EVENTS_ENABLE)");
#endif /* PR_TASK_PERF_EVENTS_ENABLE */

  struct perf_event_attr attr;
  memset(&attr, 0, sizeof(struct perf_event_attr));
  attr.size = sizeof(struct perf_event_attr);
  attr.type = PERF_TYPE_HARDWARE;
  attr.config = PERF_COUNT_HW_CPU_CYCLES;
  attr.read_format = PERF_FORMAT_TOTAL_TIME_RUNNING;
  attr.disabled = 1;
  // attr.pinned = 1;
  attr.exclude_kernel = 1;
  attr.exclude_hv = 1;
#ifndef PERF_FLAG_FD_CLOEXEC /* Since 3.14 */
#define PERF_FLAG_FD_CLOEXEC 0
#endif
  perf_fd = perf_event_open(&attr, 0 /* current process */, -1 /* any cpu */,
                            -1 /* no group */, PERF_FLAG_FD_CLOEXEC);
  if (perf_fd < 0) {
    perf_error = errno;
    if (perf_error != EACCES /* will handle later */)
      perror(MERA_PERROR_PREFIX "perf_event_open()");
    return -1;
  }

#if defined(__ia32__)
  perf_page = (struct perf_event_mmap_page *)mmap(
      NULL, getpagesize(), PROT_WRITE | PROT_READ, MAP_SHARED, perf_fd, 0);
  if (perf_page == MAP_FAILED) {
    perf_error = errno;
    perror(MERA_PERROR_PREFIX "mmap(perf_event_mmap_page)");
    perf_page = NULL;
  }
#endif /* __ia32__ */

  if (ioctl(perf_fd, PERF_EVENT_IOC_ENABLE, 0) /* Start counters */) {
    perf_error = errno;
    perror(MERA_PERROR_PREFIX "ioctl(PERF_EVENT_IOC_ENABLE)");
    close(perf_fd);
    perf_fd = -1;
    return -1;
  }
  perf_error = 0;
  return 0;
}

#if defined(__ia32__)
static unsigned perf_rdpmc_index;
unsigned perf_rdpmc_start(timestamp_t *now) {
  compiler_barrier();
  uint32_t low, high;
  __asm __volatile("cpuid; mov %2, %%ecx; rdpmc"
                   : "=a"(low), "=d"(high)
                   : "m"(perf_rdpmc_index)
                   : "%ebx", "%ecx");
  union timestamp *u = (union timestamp *)now;
  u->u32.l = low;
  u->u32.h = high;
  return 0;
}

unsigned perf_rdpmc_finish(timestamp_t *now) {
  compiler_barrier();
  uint32_t low, high;
  __asm __volatile("mov %2, %%ecx; rdpmc; mov %%eax, %0; mov %%edx, %1; cpuid"
                   : "=r"(low), "=r"(high)
                   : "m"(perf_rdpmc_index)
                   : "%eax", "%ebx", "%ecx", "%edx");
  union timestamp *u = (union timestamp *)now;
  u->u32.l = low;
  u->u32.h = high;
  return 0;
}
#endif /* __ia32__ */

#else
#define perf_fd (-1)
#endif /* __NR_perf_event_open */

/*****************************************************************************/

bool mera_init(void) {
  mera.flags = 0;
  mera.cpunum = set_single_affinity();

#if defined(PR_SET_TSC) && defined(__ia32__)
  int tsc_mode = PR_TSC_SIGSEGV;
  if (prctl(PR_GET_TSC, &tsc_mode, 0, 0, 0))
    perror(MERA_PERROR_PREFIX "prctl(PR_GET_TSC)");
  else if (tsc_mode != PR_TSC_ENABLE &&
           prctl(PR_SET_TSC, PR_TSC_ENABLE, 0, 0, 0))
    perror(MERA_PERROR_PREFIX "prctl(PR_SET_TSC, PR_TSC_ENABLE)");
#endif /* PR_SET_TSC */

#if defined(EMSCRIPTEN)
  return probe(clock_emscripten, clock_emscripten, convert_1to1, 0,
               "emscripten_get_now()", "ns");
#endif

#if defined(TIMEBASE_SZ) || defined(__OS400__)
  probe(clock_os400, clock_os400, convert_os400, 0,
        "read_wall_time(TIMEBASE_SZ)", "ns");
#endif

#if defined(__APPLE__) || defined(__MACH__)
  probe(clock_mach, clock_mach, convert_mach, 0, "mach_absolute_time()", "ns");
#endif

#if defined(__sun__) || defined(__sun)
  probe(clock_solaris, clock_solaris, convert_1to1, 0, "gethrtime()", "ns");
#endif /* __sun__ */

#if defined(CLOCK_SGI_CYCLE)
  if (posix_clockid == CLOCK_REALTIME) {
    posix_clockid = CLOCK_SGI_CYCLE;
    if (!probe(clock_posix, clock_posix, convert_1to1, 0,
               "clock_gettime(CLOCK_SGI_CYCLE)", "ns"))
      posix_clockid = CLOCK_REALTIME;
  }
#endif /* CLOCK_SGI_CYCLE */
#if defined(CLOCK_MONOTONIC_RAW)
  if (posix_clockid == CLOCK_REALTIME) {
    posix_clockid = CLOCK_MONOTONIC_RAW;
    if (!probe(clock_posix, clock_posix, convert_1to1, 0,
               "clock_gettime(CLOCK_MONOTONIC_RAW)", "ns"))
      posix_clockid = CLOCK_REALTIME;
  }
#endif /* CLOCK_MONOTONIC_RAW */
#if defined(CLOCK_MONOTONIC)
  if (posix_clockid == CLOCK_REALTIME) {
    posix_clockid = CLOCK_MONOTONIC;
    if (!probe(clock_posix, clock_posix, convert_1to1, 0,
               "clock_gettime(CLOCK_MONOTONIC)", "ns"))
      posix_clockid = CLOCK_REALTIME;
  }
#endif /* CLOCK_MONOTONIC */

#if defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) ||              \
    defined(__WINDOWS__)
  probe(clock_windows, clock_windows, convert_windows, 0,
        "QueryPerformanceCounter()", "ns");
#else
  probe(clock_gettimeofday, clock_gettimeofday, convert_us2ns, 0,
        "gettimeofday()", "ns");
#endif /* Windows */

  /***************************************************************************/

#ifndef __native_client__
#if defined(__elbrus__) || defined(__e2k__)
  probe(clock_elbrus, clock_elbrus, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "Elbrus_TSCP", "cycle");
#endif /* __elbrus__ */

#if (defined(__powerpc64__) || defined(__ppc64__) || defined(__ppc64) ||       \
     defined(__powerpc64))
  probe(clock_powerpc64_mfspr268, clock_powerpc64_mfspr268, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "MFSPR(268)", "cycle");
#endif /* __powerpc64__ */

#if (defined(__powerpc__) || defined(__ppc__) || defined(__powerpc) ||         \
     defined(__ppc))
  probe(clock_powerpc_mftb, clock_powerpc_mftb, convert_1to1,
        timestamp_clock_cheap | timestamp_ticks, "MFTB", "tick");
#endif /* __powerpc__ */

#if defined(__sparc__) || defined(__sparc) || defined(__sparc64__) ||          \
    defined(__sparc64) || defined(__sparc_v8plus__) ||                         \
    defined(__sparc_v8plus) || defined(__sparc_v8plusa__) ||                   \
    defined(__sparc_v8plusa) || defined(__sparc_v9__) || defined(__sparc_v9)
  probe(clock_sparc, clock_sparc, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "tick_register", "cycle");
#endif /* __sparc__ */

#if defined(__ia64__) || defined(__ia64)
  probe(clock_ia64, clock_ia64, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "ITC", "cycle");
#endif /* __ia64__ */

#if (defined(__hppa__) || defined(__hppa) || defined(__hppa64__) ||            \
     defined(__hppa64))
  probe(clock_hppa, clock_hppa, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "MFCTL(16)", "cycle");
#endif /* __hppa__ */

#if defined(__s390__) || defined(__s390) || defined(__zarch__) ||              \
    defined(__zarch)
  probe(clock_stcke, clock_stcke, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "STCKE", "cycle");
  probe(clock_stckf, clock_stckf, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "STCKF", "cycle");
  probe(clock_stck, clock_stck, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "STCK", "cycle");
#endif /* __s390__ */

#if defined(__alpha__) || defined(__alpha)
  probe(clock_alpha, clock_alpha, convert_1to1,
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
        "RPCC", "cycle");
#endif /* __alpha__ */

#if (defined(__ARM_ARCH) && __ARM_ARCH > 5 && __ARM_ARCH < 8) || defined(_M_ARM)
  /* Read the user mode perf monitor counter access permissions. */
  uint32_t pmuseren;
#ifdef _M_ARM
  pmuseren = _MoveFromCoprocessor(15, 0, 9, 14, 0);
#else
  __asm("mrc p15, 0, %0, c9, c14, 0" : "=r"(pmuseren));
#endif
  if (1 & pmuseren /* Is it allowed for user mode code? */) {
    uint32_t pmcntenset;
#ifdef _M_ARM
    pmcntenset = _MoveFromCoprocessor(15, 0, 9, 12, 1);
#else
    __asm("mrc p15, 0, %0, c9, c12, 1" : "=r"(pmcntenset));
#endif
    if (pmcntenset & 0x80000000ul /* Is it counting? */)
      probe(clock_pmccntr, clock_pmccntr, convert_pmccntr_x64,
            timestamp_clock_stable | timestamp_cycles | timestamp_clock_cheap,
            "PMCCNTR", "clk");
    else {
      printf(" - suggest enable performance-counter\n");
    }
  } else {
    printf(" - suggest enable access to performance-counters from user-mode\n");
  }
#endif /* (__ARM_ARCH > 5 && __ARM_ARCH < 8) || _M_ARM */

#if defined(__aarch64__) || (defined(__ARM_ARCH) && __ARM_ARCH > 7) ||         \
    defined(_M_ARM64)
  /* System timer of ARMv8 runs at a different frequency than the CPU's.
   * The frequency is fixed, typically in the range 1-50MHz.  It can be
   * read at CNTFRQ special register.  We assume the OS has set up
   * the virtual timer properly. */
  probe(clock_cntvct_el0, clock_cntvct_el0, convert_1to1,
        timestamp_clock_stable | timestamp_ticks | timestamp_clock_cheap,
        "CNTVCT_EL0", "tick");
#endif /* __aarch64__ || __ARM_ARCH > 7 || _M_ARM64 */

#if defined(__mips__) || defined(__mips)

#if (defined(_MIPS_ISA) && defined(_MIPS_ISA_MIPS2) &&                         \
     _MIPS_ISA >= _MIPS_ISA_MIPS2) ||                                          \
    (defined(__mips) && __mips >= 2 && __mips < 16) || defined(_R4000) ||      \
    defined(__MIPS_ISA2) || defined(__MIPS_ISA3) || defined(__MIPS_ISA4) ||    \
    (defined(__mips_isa_rev) && __mips_isa_rev >= 2)

  probe(clock_mfc0_9_0, clock_mfc0_9_0, convert_1to1,
        timestamp_clock_stable | timestamp_clock_cheap | timestamp_cycles,
        "MFC0(9.0)", "cycle");

  if (probe(clock_rdhwr, clock_rdhwr, convert_rdhwr,
            timestamp_clock_stable | timestamp_clock_cheap | timestamp_cycles,
            "RDHWR(2)", "cycle")) {
    unsigned rdhwr_3;
    __asm("rdhwr %0, $3" : "=r"(rdhwr_3));
    mips_rdhwr_resolution = rdhwr_3;
    if (mips_rdhwr_resolution < 2)
      mera.convert = convert_1to1;
  }

  probe(clock_mfc0_25_1, clock_mfc0_25_1, convert_1to1,
        timestamp_clock_stable | timestamp_clock_cheap | timestamp_cycles,
        "MFC0(25.1)", "cycle");
#endif /* MIPS >= 2 */

#if defined(PROT_READ) && defined(MAP_SHARED)
  uint64_t *mips_tsc_addr;
  int mem_fd = open("/dev/mem", O_RDONLY | O_SYNC, 0);

  if (mem_fd < 0)
    if (errno == EACCES)
      printf(" - suggest run from super-user for access to /dev/mem "
             "(MIPS_ZBUS_TIMER)\n");
    else
      perror(MERA_PERROR_PREFIX "open(/dev/mem)");
  else {
    mips_tsc_addr = mmap(NULL, getpagesize(), PROT_READ, MAP_SHARED, mem_fd,
                         0x10030000 /* MIPS_ZBUS_TIMER */);
    if (mips_tsc_addr == MAP_FAILED) {
      perror(MERA_PERROR_PREFIX "mmap(MIPS_ZBUS_TIMER)");
      close(mem_fd);
    } else {
      close(mem_fd);
      if (!probe(clock_zbustimer, clock_zbustimer, convert_1to1,
                 timestamp_clock_stable | timestamp_clock_cheap |
                     timestamp_ticks,
                 "ZBUS-Timer(0x10030000)", "tick")) {

        munmap(mips_tsc_addr, getpagesize());
        mips_tsc_addr = NULL;
      }
    }
  }
#endif /* PROT_READ && MAP_SHARED */

#endif /* __mips__ */

#if defined(__ia32__)
  if (ia32_cpu_features.basic.edx == 0)
    ia32_fetch_cpu_features();
  if (ia32_cpu_features.basic.edx & (1 << 4)) {
    probe(clock_rdpmc_start, clock_rdpmc_finish, convert_1to1,
          timestamp_clock_stable | timestamp_clock_cheap | timestamp_cycles,
          "RDPMC_40000001", "cycle");
    const unsigned tsc_flags =
        (ia32_cpu_features.extended_80000007.edx & (1 << 8))
            /* The TSC rate is invariant, i.e. not always on CPU frequency ! */
            ? timestamp_clock_cheap | timestamp_cycles
            : timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable;

#ifdef F_OK
    if (!(tsc_flags & timestamp_clock_stable) &&
        !(mera.flags & timestamp_clock_stable) &&
        access("/sys/devices/cpu/rdpmc", F_OK) == 0) {
      printf(" - suggest enable rdpmc for usermode (echo 2 | sudo tee "
             "/sys/devices/cpu/rdpmc)\n");
    }
#endif /* F_OK */

    if (ia32_cpu_features.extended_80000001.edx & (1 << 27))
      probe(clock_rdtscp_start, clock_rdtscp_finish, convert_1to1, tsc_flags,
            "RDTSCP", NULL);
    probe(clock_rdtsc_start, clock_rdtsc_finish, convert_1to1, tsc_flags,
          "RDTSC", NULL);
  }
#endif /* __ia32__ */

#if defined(__NR_perf_event_open)
  if (perf_setup() == 0) {
    bool perf_used = probe(clock_perf, clock_perf, convert_1to1,
                           timestamp_cycles | timestamp_clock_stable,
                           "PERF_COUNT_HW_CPU_CYCLES", "cycle");
#if defined(__ia32__)
    if (perf_page) {
      bool perf_used_page = false;
      if (perf_page->cap_bit0_is_deprecated && perf_page->cap_user_rdpmc &&
          perf_page->index) {
        perf_rdpmc_index = perf_page->index - 1;
        perf_used_page = probe(
            perf_rdpmc_start, perf_rdpmc_finish, convert_1to1,
            timestamp_clock_stable | timestamp_clock_cheap | timestamp_cycles,
            "RDPMC_perf", "cycle");
      }
      if (perf_used_page)
        perf_used = true;
      else {
        munmap((void *)perf_page, getpagesize());
        perf_page = NULL;
      }
    }
#endif /* __ia32__ */
    if (!perf_used) {
      close(perf_fd);
      perf_fd = -1;
    }
  } else if (!(mera.flags & timestamp_clock_stable) && perf_error == EACCES &&
             access("/proc/sys/kernel/perf_event_paranoid", F_OK) == 0) {
    printf(" - suggest enable perf for non-admin users (echo 2 | sudo tee "
           "/proc/sys/kernel/perf_event_paranoid)\n");
  }
#endif /* __NR_perf_event_open */

#endif /* ! __native_client__ */

#if __has_builtin(__builtin_readcyclecounter)
  probe(clock_readcyclecounter, clock_readcyclecounter, convert_1to1,
#if defined(__ia32__)
        timestamp_clock_cheap | timestamp_cycles |
            (mera.flags & timestamp_clock_stable),
#else
        timestamp_clock_cheap | timestamp_cycles | timestamp_clock_stable,
#endif
        "__builtin_readcyclecounter()", "cycle");
#endif

  return (mera.flags & timestamp_clock_have) ? true : false;
}

static unsigned fuse_timestamp(timestamp_t *unused) {
  (void)unused;
  abort();
  return 0;
}

static double fuse_convert(timestamp_t unused) {
  (void)unused;
  abort();
  return 0;
}

mera_t mera = {
    fuse_timestamp, fuse_timestamp, fuse_convert, "void", "none", 0, -1};

/*****************************************************************************/

mera_bci_t mera_bci;

double mera_bench(MERA_BENCH_TARGET target, MERA_BENCH_SELF_ARGS) {
  const time_t timeout_fuse = time(NULL);
  unsigned target_loops = 1;
  unsigned retry_count = 0, restart_count = 0;

  timestamp_t overhead_best = INT64_MAX;
  timestamp_t overhead_gate = 0;
  unsigned overhead_loops_max = 0;

  restart_count -= 1;
restart_top:;
  timestamp_t overhead_sum = 0;
  unsigned overhead_total_count = 0;
  unsigned overhead_best_count = 1;
  unsigned overhead_worthless_loops = 0;
  unsigned overhead_accounted_loops = 0;

restart_middle:;
  timestamp_t target_best = INT64_MAX;
  timestamp_t target_gate = 0;
  unsigned tail_loops_max = 0;

restart_bottom:;
  timestamp_t target_brutto_sum = 0;
  unsigned target_overhead_count = 0;
  unsigned target_best_count = 1;
  unsigned target_total_count = 0;
  unsigned target_worthless_loops = 0;
  unsigned target_accounted_loops = 0;
  unsigned stable = 0;
  restart_count += 1;

  retry_count -= 1;
retry:
  retry_count += 1;

  while (true) {
    /* measure the overhead of measurement */
    unsigned coreid;
    {
      /* wait for edge of tick */
      timestamp_t snap, start, finish;
      coreid = mera.start(&snap);
      do {
        if (unlikely(coreid != mera.start(&start) || snap > start))
          goto retry;
      } while (snap == start);

      /* first iteration */
      unsigned loops = 1;
      if (unlikely(coreid != mera.finish(&finish) || start > finish))
        goto retry;

      /* loop until end of tick */
      while (start == finish) {
        loops += 1;
        if (unlikely(coreid != mera.start(&snap) || start > snap))
          goto retry;
        if (unlikely(coreid != mera.finish(&finish) || snap > finish))
          goto retry;
      }
      const timestamp_t elapsed = finish - start;
      if (unlikely(overhead_best > elapsed || overhead_loops_max < loops)) {
        if (overhead_best > elapsed) {
          overhead_gate = overhead_best + (overhead_best - elapsed + 1) / 2;
          if (overhead_gate > elapsed * 129 / 128)
            overhead_gate = elapsed * 129 / 128;
          if (overhead_gate < elapsed * 1025 / 1024 + 1)
            overhead_gate = elapsed * 1025 / 1024 + 1;
          overhead_best = elapsed;
        }
        overhead_loops_max =
            (overhead_loops_max > loops) ? overhead_loops_max : loops;
        goto restart_top;
      } else if (likely(elapsed <= overhead_gate &&
                        loops + 1 >= overhead_loops_max)) {
        if (elapsed == overhead_best && loops == overhead_loops_max)
          overhead_best_count += 1;
        overhead_sum += elapsed;
        overhead_total_count += loops;
        overhead_accounted_loops += 1;
      } else {
        overhead_worthless_loops += 1;
      }
    }

    /* measure the target */
    if (target) {
      /* wait for edge of tick */
      timestamp_t snap, start, finish;
      if (unlikely(coreid != mera.start(&snap)))
        goto retry;
      do {
        if (unlikely(coreid != mera.start(&start) || snap > start))
          goto retry;
      } while (snap == start);

      unsigned loops = 0;
      do
        target(MERA_BENCH_TARGET_ARGS);
      while (++loops < target_loops);

      loops = 1;
      if (unlikely(coreid != mera.finish(&finish) || snap > finish))
        goto retry;

      /* wait for next tick */
      while (true) {
        if (unlikely(coreid != mera.start(&snap) || finish > snap))
          goto retry;
        if (finish != snap)
          break;
        if (unlikely(coreid != mera.finish(&snap) || finish > snap))
          goto retry;
        if (finish != snap)
          break;
        loops += 1;
      }

      const timestamp_t elapsed = finish - start;
      if (unlikely(target_best > elapsed ||
                   (target_best == elapsed && tail_loops_max < loops))) {
        if (target_best > elapsed) {
          target_gate = target_best + (target_best - elapsed + 1) / 2;
          if (target_gate > elapsed * 129 / 128)
            target_gate = elapsed * 129 / 128;
          if (target_gate < elapsed * 1025 / 1024 + 1)
            target_gate = elapsed * 1025 / 1024 + 1;
          target_best = elapsed;
        }
        tail_loops_max = loops;
        goto restart_bottom;
      } else if (likely(elapsed <= target_gate &&
                        (tail_loops_max - loops /* overflow is ok */) < 2)) {
        if (elapsed == target_best && loops == tail_loops_max)
          target_best_count += 1;
        target_total_count += target_loops;
        target_brutto_sum += elapsed;
        target_overhead_count += loops;
        target_accounted_loops += 1;
      } else {
        target_worthless_loops += 1;
      }
    }

    /* checkpoint */
    if (unlikely((++stable & 1023) == 0)) {
      if (target) {
        const timestamp_t wanna = 1042 + overhead_best * overhead_loops_max;
        if (target_best < wanna) {
          target_loops += target_loops;
          goto restart_middle;
        }
        if (target_loops > 1 && target_best > wanna * 4) {
          target_loops >>= 1;
          goto restart_middle;
        }
      }

      const unsigned enough4fuse_seconds = 9;
      const unsigned enough4best =
          (mera.flags & timestamp_clock_stable) ? 499 : 1999;
      const unsigned enough4avg =
          (mera.flags & timestamp_clock_stable) ? 4999 : 29999;
      const unsigned enough4bailout =
          (mera.flags & timestamp_clock_cheap) ? 99999 : 59999;

      const unsigned spent_seconds = (unsigned)(time(NULL) - timeout_fuse);

      const bool enough4overhead = overhead_best_count > enough4best ||
                                   overhead_accounted_loops > enough4avg ||
                                   overhead_worthless_loops > enough4bailout ||
                                   spent_seconds > enough4fuse_seconds;

      const bool enough4target = target_best_count > enough4best ||
                                 target_accounted_loops > enough4avg ||
                                 target_worthless_loops > enough4bailout ||
                                 spent_seconds > enough4fuse_seconds;

      /* calculate results */
      if (enough4overhead && (!target || enough4target)) {
        memset(&mera_bci, 0, sizeof(mera_bci));
        mera_bci.retry_count = retry_count;
        mera_bci.restart_count = restart_count;
        mera_bci.spent_seconds = spent_seconds + 1;

        mera_bci.overhead_best = overhead_best;
        mera_bci.overhead_gate = overhead_gate;
        mera_bci.overhead_loops_max = overhead_loops_max;
        mera_bci.overhead_best_count = overhead_best_count;
        mera_bci.overhead_accounted_loops = overhead_accounted_loops;
        mera_bci.overhead_worthless_loops = overhead_worthless_loops;

        const double measured_overhead =
            (overhead_best_count > 2 || overhead_total_count < enough4avg / 2)
                ? mera.convert(overhead_best) / overhead_loops_max
                : mera.convert(overhead_sum) / overhead_total_count;
        if (!target)
          return measured_overhead;

        mera_bci.target_loops = target_loops;
        mera_bci.target_best = target_best;
        mera_bci.target_gate = target_gate;
        mera_bci.tail_loops_max = tail_loops_max;
        mera_bci.target_best_count = target_best_count;
        mera_bci.target_accounted_loops = target_accounted_loops;
        mera_bci.target_worthless_loops = target_worthless_loops;

        const double measured_target =
            (target_best_count > 2 || target_total_count < enough4avg / 2)
                ? (mera.convert(target_best) -
                   measured_overhead * tail_loops_max) /
                      target_loops
                : (mera.convert(target_brutto_sum) -
                   measured_overhead * target_overhead_count) /
                      target_total_count;
        return measured_target;
      }
    }
  }
}