Closure Conversion - index out of bounds

[anydsl]

Closure Conversion - index out of bounds

[NeuralCoder3]

Example file links to 404. Do you have an update?

Here is the file from when the issue was opened:
https://github.com/NeuralCoder3/thorin2/blob/84af1a073ab4e1a848def18cce2e6afc7fe567f1/enzyme/brussel.thorin

Here is the file currently: (Nothing big regarding this issue should have changed)
https://github.com/NeuralCoder3/thorin2/blob/84af1a073ab4e1a848def18cce2e6afc7fe567f1/enzyme/brussel.thorin
(the enzyme folder was moved into the eval folder)

[leissa]

Example file links to 404. Do you have an update?

[leissa]

I have to double-check with @f-fries but I think the problem is that the whole clos dialect assumes that the first argument is a mem. Otherwise, things go south as in your example.

[leissa]

This is indeed the problem. I opened a new issue #126 as it really is an enhancement and closing this here for now.

[NeuralCoder3]

This is indeed the problem. I opened a new issue #126 as it really is an enhancement and closing this here for now.

Therefore, one needs a pass that reorders mem to the front and propagates it to functions without mem (higher order arguments and others) in order to generate the code.
Or would it be easier to adapt closure conversion?

[leissa]

I think it makes everyone's life easier if we can just assume that the first argument is mem if present at all. But we should also enhance the closure conversion to also cope with functions without mem. I think it's cool to have this explicitly annotated on the type that a function is pure and it would be a pity if we would enforce impure functions.

[f-fries]

It might still be a good idea to uniformly add mems before/during closure conversion as @NeuralCoder3 has suggested.
Closure conversion actually introduces new side-effects (environment allocation etc).
So some pure functions would get a "pure closure" while others would end up with an "impure closure" with a different type.