#define BLAKE3_USE_NEON 1

    #if defined(__ARM_BIG_ENDIAN)
      #define BLAKE3_USE_NEON 0
    #else
      #define BLAKE3_USE_NEON 1
    #endif

#include <Windows.h>

#if !defined(BLAKE3_ATOMICS)
#if defined(__has_include)
#if __has_include(<stdatomic.h>) && !defined(_MSC_VER)
#define BLAKE3_ATOMICS 1
#else
#define BLAKE3_ATOMICS 0
#endif /* __has_include(<stdatomic.h>) && !defined(_MSC_VER) */
#else
#define BLAKE3_ATOMICS 0
#endif /* defined(__has_include) */
#endif /* BLAKE3_ATOMICS */

#if BLAKE3_ATOMICS
#define ATOMIC_INT _Atomic int
#define ATOMIC_LOAD(x) x
#define ATOMIC_STORE(x, y) x = y
#elif defined(_MSC_VER)
#define ATOMIC_INT LONG
#define ATOMIC_LOAD(x) InterlockedOr(&x, 0)
#define ATOMIC_STORE(x, y) InterlockedExchange(&x, y)
#else
#define ATOMIC_INT int
#define ATOMIC_LOAD(x) x
#define ATOMIC_STORE(x, y) x = y
#endif

    enum cpu_feature g_cpu_features = UNDEFINED;

    ATOMIC_INT g_cpu_features = UNDEFINED;

  if (g_cpu_features != UNDEFINED) {
    return g_cpu_features;

  /* If TSAN detects a data race here, try compiling with -DBLAKE3_ATOMICS=1 */
  enum cpu_feature features = ATOMIC_LOAD(g_cpu_features);
  if (features != UNDEFINED) {
    return features;

    enum cpu_feature features = 0;

    features = 0;

    g_cpu_features = features;

    ATOMIC_STORE(g_cpu_features, features);

#define BLAKE3_VERSION_STRING "1.3.3"

#define BLAKE3_VERSION_STRING "1.5.0"

// wouldn't be able to implement exendable output.) Note that this function is

// wouldn't be able to implement extendable output.) Note that this function is