[Bug tree-optimization/78332] New: [ARM] Negative costs of ivopts groups

[ka6ash at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78332

            Bug ID: 78332
           Summary: [ARM] Negative costs of ivopts groups
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ka6ash at gmail dot com
                CC: amker.cheng at gmail dot com
  Target Milestone: ---

For the following code GCC generates six induction variables for each use
arr[{0-1}][{0-2}][k].

I think the problem is negative costs. The negative costs is appeared in the
following part of the function determine_group_iv_cost_address:

sum_cost = cost;
if (!sum_cost.infinite_cost_p () && cand->ainc_use == use)
  {
    if (can_autoinc)
      sum_cost -= cand->cost_step;
    /* If we generated the candidate solely for exploiting autoincrement
       opportunities, and it turns out it can't be used, set the cost to
       infinity to make sure we ignore it.  */
    else if (cand->pos == IP_AFTER_USE || cand->pos == IP_BEFORE_USE)
      sum_cost = infinite_cost;
  }

If we have a zero cost and autoinc is possible, we obtain the negative cost.
This situation is a little bit confusing because we also can have candidates
with a zero cost of casting and they are less preferable than candidates with
autoinc. It seems that there is a preference to use autoinc candidates. But the
testcase shows the preference could increase register pressure without any
reason.

I hotfixed it by zeroing negative costs and GCC started to use two iv.

So, are negative costs a part of the cost model and if it is, is it something
wrong with six ivs?

gcc btc.c -O2 -S -march=armv7-a -mcpu=cortex-a9

void foo(int amount, double arr[2][3][amount])
{
  int k;
  for (k = 1; k < amount; k++)
    {
      arr[0][0][k] = 1.0;
      arr[0][1][k] = 1.0;
      arr[0][2][k] = 1.0;

      arr[1][0][k] = 1.0;
      arr[1][1][k] = 1.0;
      arr[1][2][k] = 1.0;
    }
}

<Group-candidate Costs>:
Group 0:
  cand  cost    compl.  inv.ex. depends on
  0     10      2               1
...
  5     0       1
  6     -4      0
  7     -4      0
...
Group 1:
  cand  cost    compl.  inv.ex. depends on
  0     10      2       8
...
  9     0       1
  10    -4      0
  11    -4      0
...