Comment 3 for bug 2001932

By extending your testing to check for more alignments and sizes:

  for (size_t s = 99; s <= 4 * VEC_SIZE; s++)
    for (size_t s1a = 31; s1a < 32; s1a++)
      for (size_t s2a = 30; s2a < 32; s2a++)
        {
          ret = strncmp (s1 + PAGE_SIZE - s - s1a,
                         s1 + PAGE_SIZE - s - s2a,
                         s);
          assert (ret == 0);
        }

It seems that another page cross also requires fixing:

580 xorl %r8d, %r8d
581 /* If ECX > VEC_SIZE * 2, skip ECX - (VEC_SIZE * 2) bytes. */
582 subl $(VEC_SIZE * 2), %ecx
583 jle 1f
584 /* Skip ECX bytes. */
585 shrq %cl, %rdi
586 /* R8 has number of bytes skipped. */
587 movl %ecx, %r8d
588 1:
589 /* Before jumping back to the loop, set ESI to the number of
590 VEC_SIZE * 4 blocks before page crossing. */
591 movl $(PAGE_SIZE / (VEC_SIZE * 4) - 1), %esi
592
593 testq %rdi, %rdi
594 je L(back_to_loop)
595 tzcntq %rdi, %rcx
596 addq %r10, %rcx
597 /* Adjust for number of bytes skipped. */

It should not jump back to loop if the ecx is negative (as some cases).