core: Implement MemrefLayoutAttr base class and use in parser.

[PapyChacal]

This is a barebone mimic of the upstream MemrefLayoutAttrInterface, as it does not implement any of the methods; only mark attributes as acceptable by the parser. This solve the parser problem though, and allow the same extensibility on accepted layouts.

The uniplemented methods were not implemented on the existing layout attributes anyway, so I'm in favor of just merging, moving on, and think about implementing them in an interface style later.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.86%. Comparing base (82a6ad6) to head (9c6e004).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2718   +/-   ##
=======================================
  Coverage   89.86%   89.86%           
=======================================
  Files         367      367           
  Lines       47465    47468    +3     
  Branches     7234     7234           
=======================================
+ Hits        42654    42659    +5     
+ Misses       3698     3697    -1     
+ Partials     1113     1112    -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jorendumoulin]

I have tested this on the snax side, and everything works beautifully. Thanks a lot for your help on this!

Somewhat unrelated: the changes mix the use of isinstance and isa. When should one use which?

[superlopuh]

isinstance is a built-in Python thing, and I would say is the recommended thing to use when it's sufficient. isa is an xDSL function, that can handle more things, notably Python generics and PyRDL constraints.

[math-fehr]

Thanks for implementing this!
I do remember not implementing it properly, and I do really HATE what MLIR is doing here, but that's life.
It is one of the rare case in MLIR where parsing depends on an interface, which means here that if you want to use a memory layout that is unregistered, then the parsing will fail even though we may want it to work (Just saying this so you know in case you encounter this).

[PapyChacal]

Thanks for implementing this! I do remember not implementing it properly, and I do really HATE what MLIR is doing here, but that's life. It is one of the rare case in MLIR where parsing depends on an interface, which means here that if you want to use a memory layout that is unregistered, then the parsing will fail even though we may want it to work (Just saying this so you know in case you encounter this).

Good point; but how do you know if your unregistered attribute is a memory space or layout then?