skipnlsse2.c

#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>

#ifdef __SSE2__
#include <emmintrin.h>
#endif

static const char *SkipToNewline(const char *CurPtr, const char *BufferEnd) {
  unsigned char C = *CurPtr;
  if (CurPtr + 24 < BufferEnd) {
    // While not aligned to a 16-byte boundary.
    while (C != '\n' && C != '\r' && ((intptr_t)CurPtr & 0x0F) != 0)
      C = *CurPtr++;

    if (C == '\n' || C == '\r') return CurPtr;

#ifdef __SSE2__
    __m128i NLs = _mm_set1_epi8('\n');
    __m128i CRs = _mm_set1_epi8('\r');
    while (CurPtr+16 <= BufferEnd) {
      __m128 Data = *(__m128i*)CurPtr;
      __m128 CmpNL = _mm_cmpeq_epi8(Data, NLs);
      __m128 CmpCR = _mm_cmpeq_epi8(Data, CRs);
      __m128 Or = _mm_or_si128(CmpNL, CmpCR);
      if (_mm_movemask_epi8(Or) != 0)
        break;
      CurPtr += 16;
    }
#else
    while (CurPtr+4 < BufferEnd &&
           CurPtr[0] != '\n' && CurPtr[0] != '\r' &&
           CurPtr[1] != '\n' && CurPtr[1] != '\r' &&
           CurPtr[2] != '\n' && CurPtr[2] != '\r' &&
           CurPtr[3] != '\n' && CurPtr[3] != '\r')
      CurPtr += 4;
#endif

    // It has to be one of the bytes scanned, increment to it and read one.
    C = *CurPtr++;
  }

  // Loop to scan the remainder.
  while (C != '\n' && C != '\r' && C != '\0')
    C = *CurPtr++;

  return CurPtr;
}

int main(int argc, char *argv[]) {
  FILE *fd = fopen(argv[1], "r");
  fseek(fd, 0, SEEK_END);
  unsigned long len = ftell(fd);
  fseek(fd, 0, SEEK_SET);
  char *buf = malloc(len+1);
  char *end = buf + len;
  fread(buf, 1, len, fd);
  *end = 0;
  const char *ptr = buf-1;
  unsigned lines = 0;
  while ((ptr = SkipToNewline(ptr+1, end)) != end) {
    printf("%zu\n", ptr-buf);
    lines++;
  }
  printf("\n%u\n", lines);
}