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

[Florian Hahn via llvm-dev]

On 25/11/2017 18:10, Andrew Trick via llvm-dev wrote:
> 
> 
>> On Nov 24, 2017, at 6:00 PM, 陳韋任 <chenwj.cs97g at g2.nctu.edu.tw 
>> <mailto:chenwj.cs97g at g2.nctu.edu.tw>> wrote:
>>
>>     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.
>>
>>
>> ​Isn't GCNMaxOccupancySchedStrategy [1] already an example on 
>> using GenericScheduler as the base class?
>>
>> [1] http://llvm.org/doxygen/classllvm_1_1GCNMaxOccupancySchedStrategy.html
> 
> Yes. I think GenericScheduler and related logic should continue to 
> evolve to make that approach easier and more clear. The idea should be 
> “take what you need and build on top". Not “take everything and randomly 
> fiddle with dozens of knobs”.
> -Andy
>


Yes, at the moment, re-using functionality from GenericScheduler is 
quite hard.

A quite simple (and naive) way to make the existing infrastructure more 
re-usable would be to make GenericScheduler::tryCandidate virtual or 
allow providing a scheduler that takes a tryCandidate function as 
argument. To tune the heuristics, all people would have to do is to 
provide their target-specific tryCandidate function and re-use the logic 
to do bottom-up/top-down scheduling + the processor simulation in 
SchedBoundary. I think this could cover most use-cases, where people 
just want to tweak the heuristics.

If dynamic dispatch is a concern compile-time wise, I am sure we could 
come up with a templated version or something else to get the same thing 
with static dispatch.

What do you think? I would be quite happy to prototype something along 
those lines, as I think this would be helpful when migrating other 
backends to the MachineScheduler, e.g. for the ARM backend, slight 
tweaks to the heuristics seems beneficial. Having them in a target 
specific scheduler would make things easier.

Cheers,
Florian