代码

#include <x86intrin.h>
#pragma GCC optimize("Ofast")
#pragma GCC target("avx,sse2,sse3,sse4,mmx")

void my_memset(char *a, char c, unsigned n) {
  __m256i v = _mm256_set1_epi8(c);
  while ((long)a & 127) *a++ = c, --n;
  unsigned i = 0;
  for (; i + 127 < n; i += 128) {
    _mm256_store_si256((__m256i *)(a + i + 00), v);
    _mm256_store_si256((__m256i *)(a + i + 32), v);
    _mm256_store_si256((__m256i *)(a + i + 64), v);
    _mm256_store_si256((__m256i *)(a + i + 96), v);
  }
  for (; i < n; ++i) {
    a[i] = c;
  }
}

评测结果