Re: GCC 4.0 Status Report (2005-02-03) - SMS improvements

[Mostafa Hagog <MUSTAFA at il dot ibm dot com>]

> * Project Title
SMS (Modulo Scheduling) Improvements.

>
> * Project Contributors
Mostafa Hagog

>
> * Dependencies
No dependencies.

>
> * Delivery Date

Ready, currently committed to the autovect-branch.

>
> * Description
>
>   Describe the project *in detail*.
>
>   What will you be doing?

1. Make the tree loop-versioning usable in RTL cfg-layout mode by making
   it a cfg hook.
2. Move SMS to use cfg-layout mode.
3. Use loop information to detect simple loops.
4. Replace the loop versioning in SMS by using the RTL loop-versioning.
5. Several other improvements:
   1. Check if the SMSed loop kernel is more efficient in means of
      number of cycles if not undo the changes. We do this by feeding
      the loop kernel into the DFA and counting the number of cycles
      before and after SMS - if we didn't improve (there is a chance
      because of the register copies we add) we prefer the original
      loop.

   2. Ignore register anti-dependencies - use register copies instead.
   3. Add backtracking to the scheduling algorithm;
      When failing to find a cycle within a kernel of II cycles for a
      given node we used to restart the whole process with kernel of
      II + 1.  Now we try to unscheduled some of the nodes that the one
      we failed on depends on and schedule the failing node first, then
      try the other nodes.

>
>   What parts of the compiler will be modified?

tree loop-versioning, cfg-hooks, modulo scheduler (SMS).

>
>   What will the benefits of your change be?  (Be specific -- for
>   example, if you will be improving code quality, indicate what kind
>   of code, and, if possible, how great the improvement will be.)
1. More loops will be applicable to SMS.
2. SMS is undone when it is not profitable, so we don't increase code
   size in cases we don't gain much.
3. Better schedules - due to backtracking.

>
>   What risks do you foresee?  (If you say "none", you'll be asked to
>   resubmit...)  What will you be doing to mitigate those risks?

The only risk is to affect the tree level loop versioning and
the modulo-scheduling.