Re: Speedup CSE by 5%

[Kazu Hirata <kazu at cs dot umass dot edu>]

Hi Jeff,

> I understand all that.  However, if you look at the first call to
> validate_change (inside an ASM_OPERANDS case) you'll see a case where
> I think we can modify x without copying it first.

I think the proposed patch is actually safe.  Consider

    case ASM_OPERANDS:
      for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
	validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
			 fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
      break;

validate_change enters a change into the CHANGES vector like so

  changes[num_changes].object = object;

apply_change_group starts out like so:

  for (i = 0; i < num_changes; i++)
    {
      rtx object = changes[i].object;

      /* If there is no object to test or if it is the same as the one we
         already tested, ignore it.  */
      if (object == 0 || object == last_validated)
	continue;

So, if we have changes[i].object == 0 for each i, apply_change_group
does nothing but set BB_DIRTY flag, which makes sense because we need
an insn to validate a change.

Just to make things clear, we might want to do something like this:

    case ASM_OPERANDS:
      if (insn)
	{
	  for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
	    validate_change (insn, &ASM_OPERANDS_INPUT (x, i),
			     fold_rtx (ASM_OPERANDS_INPUT (x, i), insn), 0);
	}
      break;

Kazu Hirata