基于 intel sse4指令集的crc32c实现

demo.c

#include <stdint.h>

#if defined(__x86_64__) || defined(__i386__)
  static inline uint32_t _mm_crc32_u8(uint32_t crc, uint8_t v) { 
    __asm__ volatile("crc32b %1, %0" : "+r"(crc) : "rm"(v)); 
    return crc; 
  }

  static inline uint32_t _mm_crc32_u16(uint32_t crc, uint16_t v) { 
    __asm__ volatile("crc32w %1, %0" : "+r"(crc) : "rm"(v)); 
    return crc; 
  }

  static inline uint32_t _mm_crc32_u32(uint32_t crc, uint32_t v) { 
    __asm__ volatile("crc32l %1, %0" : "+r"(crc) : "rm"(v)); 
    return crc;
  }

  static inline uint32_t _mm_crc32_u64(uint32_t crc, uint64_t v) { 
    uint64_t result = crc;
    __asm__ volatile("crc32q %1, %0" : "+r"(result) : "rm"(v)); 
    return result;
  }
#elif defined (__arm__) || defined (__aarch64__)
  static inline uint32_t _mm_crc32_u8(uint32_t crc, uint8_t v) { 
    __asm__ volatile("crc32cb %w[c], %w[c], %w[v]":[c]"+r"(crc):[v]"r"(v)); 
    return crc;
  }

  static inline uint32_t _mm_crc32_u16(uint32_t crc, uint16_t v) { 
    __asm__ volatile("crc32ch %w[c], %w[c], %w[v]":[c]"+r"(crc):[v]"r"(v)); 
    return crc;
  }

  static inline uint32_t _mm_crc32_u32(uint32_t crc, uint32_t v) { 
    __asm__ volatile("crc32cw %w[c], %w[c], %w[v]":[c]"+r"(crc):[v]"r"(v)); 
    return crc;
  }

  static inline uint32_t _mm_crc32_u64(uint32_t crc, uint64_t v) { 
    uint64_t result = crc;
    __asm__ volatile("crc32cx %w[c], %w[c], %x[v]":[c]"+r"(result):[v]"r"(v)); 
    return result;
  }
#else
  #error "Unsupported platforms"
#endif

static inline unsigned int luaS_hash (const char *str, size_t l, unsigned int seed) {
  unsigned int hash = seed ^ cast_uint(l);

  while (l >= 8)
  {
    hash = _mm_crc32_u64(hash, *(uint64_t*)str);
    str += 8; l -= 8;
  }

  while (l >= 4)
  {
    hash = _mm_crc32_u32(hash, *(uint32_t*)str);
    str += 4; l -= 4;
  }

  while (l >= 2)
  {
    hash = _mm_crc32_u16(hash, *(uint16_t*)str);
    str += 2; l -= 2;
  }
  
  while (l--)
    hash = _mm_crc32_u8(hash, *(uint8_t*)str++);

  return hash;
}

test.lua

local mrandom = math.random
local tconcat = table.concat

local str = "administratoradministrator"

local a, b = 1 << 32, 1 << 60

for i = 1, 1000000 do
  -- local v = str .. mrandom(a, b) .. str .. mrandom(a, b) .. str .. mrandom(a, b)
  local v = tconcat { str , mrandom(a, b) , str , mrandom(a, b) , str , mrandom(a, b)}
end