[llvm-dev] mischeduler (pre-RA) experiments

[Andrew Trick via llvm-dev]

> On Nov 23, 2017, at 2:53 AM, Jonas Paulsson <paulsson at linux.vnet.ibm.com> wrote:
> 
> Hi,
> 
> I have been experimenting for a while with tryCandidate() method of the pre-RA mischeduler. I have by chance found some parameters that give quite good results on benchmarks on SystemZ (on average 1% improvement, some improvements of several percent and very little regressions). Basically, I add a "latency heuristic boost" just above processor resources checking:
> 
> tryCandidate() {
>    ...
> 
>    // Avoid increasing the max pressure of the entire region.
>    if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax,
>         Cand.RPDelta.CurrentMax, TryCand, Cand, RegMax, TRI, DAG->MF))
>      return;
> 
>    /// INSERTION POINT
> 
>    ...
> }
> 
> I had started to experiment with adding tryLatency() in various places, and found this to be the best spot for SystemZ/SPEC-2006. This gave noticeable improvements immediately that were to good to ignore, so I started figuring out things about the regressions that of course also showed up. Eventually I have come up after many iterations a combined heuristic that reads:
> 
> if (((TryCand.Latency >= 7 && "Longest latency of any SU in DAG" < 15) ||
>      "Number of SUnits in DAG" > 180)
>      &&
>      tryLatency(TryCand, Cand, *Zone))
>         return;
> 
> In English: do tryLatency either if the latency of the candidate is >= 7 and the DAG has no really long latency SUs (lat > 15), or alternatively always if the DAG is really big (>180 SUnits).
> 
> I am now looking for opinions on what to do next with this.
> 
> pros:
> 
> - Clearly beneficial on benchmarks.
> 
> - All the register pressure heuristics have been run, so there *should* not be increased spilling.
> 
> - This is mainly giving latency priority over resource balancing, which seems harmless if it shows to be beneficial.
> 
> - This gives higher ILP (register usage) and gives the SystemZ post-RA scheduler more freedom to improve decoder grouping etc.
> 
> cons:
> 
> - I am not sure if it is acceptable to have limits like this to control scheduling? Code generation can change in the years to come and who knows if those limits are safe then...
> 
> On the other hand, as I have been rebasing just recently, results have varied a bit but stayed stably beneficial. I have also managed to remove the sharp limit and improve on this concern a bit by having a zone from 170-190 that makes the change more gradual as the DAG becomes "big". The values of 7 and 15 could as well be 6/8 or 30, so it's not really hyper sensitive either at the moment, I'd say.
> 
> I don't know about any better way of getting these experimental results, but of course it would be nice to know more and be able to say "why" this works, but this is indeed complex given the high OOO nature of SystemZ in addition to the regalloc effects etc.
> 
> Then there is the matter of implementation - could this become some kind of "latencyBoost" hook in the generic scheduler (after all, other targets might benefit also), or would SystemZ have to derive its own SchedStrategy (which isn't very nice if you just want to change one thing and still get future improvements of common code)?
> 
> I would appreciate any help and suggestions!
> 
> Jonas

Your analysis makes sense and approach seems reasonable. Regarding the implementation...

I try to avoid using hooks for heuristics. We do have some hooks for enabling/disabling features of the generic scheduling infrastructure. Those are intended to make it easy to see which parts can be controlled independently so backends can pick and choose scheduling functionality.

My concern is that it becomes very difficult to modify the generic scheduler once it becomes laden with machine specific thresholds. We
want the generic scheduler to be unencumbered with details so that it's easy to redesign/refactor as functionality evolves.

For heuristics that result from careful tuning, it's best for a backend to define its own. Other backends can easily copy that approach, making their own adjustments/additions without the need to reevaluate another platform's workloads.

Once heuristics have been thoroughly tuned for a microarchitecture, I think it should have its own scheduling strategy. I don't think you want other scheduling changes to affect that target until someone can reevaluate your benchmarks.

Of course, you want to duplicate as little of the generic scheduling logic as you can. So I think the challenge is how to expose the
generic scheduler's functionality as a base class or composition of utilities so that defining your strategy doesn't require too much
copy-paste.

-Andy