transformations: (riscv_scf) add a pass to fuse perfectly nested loops

[superlopuh]

No description provided.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.80%. Comparing base (5237976) to head (ca3046d).
Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2540      +/-   ##
==========================================
- Coverage   89.81%   89.80%   -0.01%     
==========================================
  Files         351      352       +1     
  Lines       44108    44170      +62     
  Branches     6592     6603      +11     
==========================================
+ Hits        39615    39667      +52     
- Misses       3532     3543      +11     
+ Partials      961      960       -1     

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

[AntonLydike]

Does this pass consider data dependencies between the loops? I think it'd be better to fuse at a higher level (before we get memrefs even) such that we can infer this more easily.

[AntonLydike]

Disregard my uninformed comment

[AntonLydike]

Thanks for this really neat rewrite! This is very interesting! I would really like to see some more comments and documentation on the conditions and calculations (e.g. it has to be a nested for loop where outer_step = len(inner_range), inner_lb == 0, etc...)

On a side note, I am not sure if I would call this loop fusion. For me it implies that two loop bodies are merged into one. Isn't this closer to a canonicalization? It does not change any actual iteration domains, just simplifies IR.

[AntonLydike]

Thanks!