transformations: Translate memref to dsd

[n-io]

Translate memref to csl dsd ops.

Todo: Finalise how global buffers are stored.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.89%. Comparing base (a2ed14d) to head (9811518).
Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3092      +/-   ##
==========================================
- Coverage   89.91%   89.89%   -0.02%     
==========================================
  Files         419      421       +2     
  Lines       53070    53252     +182     
  Branches     8226     8256      +30     
==========================================
+ Hits        47718    47871     +153     
- Misses       4023     4042      +19     
- Partials     1329     1339      +10     

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

[dk949]

🤔

[dk949]

What happens if we hit this? What is to be done?

[n-io]

basically the same as for static, working on this now. If the value is (directly) from an arith const, we can check if we need to generate the offset/stride/size dsd instruction. Mostly, I suppose, we have to act blindly as one does with dynamic values. However, there's nothing in our pipeline that actually generates dynamic values here, so it's purely optical and not actually necessary.

[n-io]

I came to think, if we have a dynamic value, let's assume it is not as de-facto static as coming from an arith.constant (in this case, we should find or write a pass to promote arith.constant to static values). It's probably simpler and cleaner to not analyse dynamic values here and simply generate dsd modify ops for each dynamic value, without analysis.

[dk949]

I think it all makes sense