[Bug c/93561] New: [bounds checking] memory overflow for spill_for

[zhongyunde at huawei dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93561

            Bug ID: 93561
           Summary: [bounds checking] memory overflow for spill_for
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zhongyunde at huawei dot com
  Target Milestone: ---

In funcion spill_for, there is following code:

  mode = PSEUDO_REGNO_MODE (regno);
  ...
  for (i = 0; i < rclass_size; i++)
    {
      hard_regno = ira_class_hard_regs[rclass][i];
      bitmap_clear (&spill_pseudos_bitmap);
      for (j = hard_regno_nregs[hard_regno][mode] - 1; j >= 0; j--)
        {
          if (try_hard_reg_pseudos_check[hard_regno + j] != curr_pseudo_check)
            continue;
          lra_assert (!bitmap_empty_p (&try_hard_reg_pseudos[hard_regno + j]));
          bitmap_ior_into (&spill_pseudos_bitmap,
                           &try_hard_reg_pseudos[hard_regno + j]);
        }
      /* Spill pseudos.  */

In our DSP chip, we have 32 1-bit hard regs(every register has only 1 -bit),
used to match data type similar as bool. so it need 64 such registers for
DImode reg to spill, and hard_regno_nregs[hard_regno][DImode] return 64, and
the value hard_regno + j larger than FIRST_PSEUDO_REGISTER bring into the array
accessing of memory overflow.

should we add the following code to avoid such issue ?
      for (j = hard_regno_nregs[hard_regno][mode] - 1; j >= 0; j--)
        {
+++          if ((hard_regno + j) >= FIRST_PSEUDO_REGISTER)
+++             break;
          if (try_hard_reg_pseudos_check[hard_regno + j] != curr_pseudo_check)
            continue;
          lra_assert (!bitmap_empty_p (&try_hard_reg_pseudos[hard_regno + j]));
          bitmap_ior_into (&spill_pseudos_bitmap,
                           &try_hard_reg_pseudos[hard_regno + j]);
        }