compiler: Improve IndexDerivative lowering

[FabioLuporini]

Tiny enhancement to avoid useless stupid temporary variables in the generated code

[FabioLuporini]

"Fabio, have you thought about just performing a round of cse() instead"

"Yes I have, unfortunately it's not that easy because CSE is a Cluster-local pass, which doesn't work well here"

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (c888cee) 86.76% compared to head (df4e53d) 86.76%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2288   +/-   ##
=======================================
  Coverage   86.76%   86.76%           
=======================================
  Files         229      229           
  Lines       42884    42893    +9     
  Branches     7951     7953    +2     
=======================================
+ Hits        37207    37216    +9     
  Misses       5002     5002           
  Partials      675      675           

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

[georgebisbas]

"Fabio, have you thought about just performing a round of cse() instead"

"Yes I have, unfortunately it's not that easy because CSE is a Cluster-local pass, which doesn't work well here"

Why does it not work well? Not sure I understand TBH

[FabioLuporini]

Why does it not work well? Not sure I understand TBH

lower_index_derivative might produce:

Cluster([
  r1 = 0
  r1 += <lowered index derivative>
  r0 = r1
])

CSE would be able to see that r0 is unnecessary here and safely removable. But r0 is needed in later Clusters, so it would produce wrong code ultimately

Note that this is just a simplified example

You would need to understand lower_index_derivatives fully to appreciate all the different flavors of this issue

[georgebisbas]

Why does it not work well? Not sure I understand TBH

lower_index_derivative might produce:

Cluster([
  r1 = 0
  r1 += <lowered index derivative>
  r0 = r1
])

CSE would be able to see that r0 is unnecessary here and safely removable. But r0 is needed in later Clusters, so it would produce wrong code ultimately

Note that this is just a simplified example

You would need to understand lower_index_derivatives fully to appreciate all the different flavors of this issue

Thanks