Add Constant Propagation pass after MIR passes

[Leo-Besancon]

This PR closes #427

Tasks:

• Implement folding of the following cases
  • BinOp(Const, Const) -> Const (BinOp being Add, Sub, Mul and Exp)
  • Add(x, 0) | Add(0, x) -> x
  • Mul(x, 0) | Mul(0, x) -> 0
  • Mul(x, 1) | Mul(1, x) -> x
  • Exp(0, x) -> 0
  • Exp(x, 0) -> 1 (note: this one is redundant with how we expand Exp node during lowering MIR > AIR, I've only included it here to be exhaustive)
  • Sub(x, 0) -> x
• Adapt lowering to Air (e.g. relax the Sub(x,y) requirement for Enf nodes)
• Clean code

[bobbinth]

Looks good! Thank you! I left a few questions/comments inline.

[bobbinth]

Looks good! Thank you! I left a few small comments inline. After these are addressed, we should be good to merge.

One general question: what's the main reason to do constant propagation in MIR rather than in the AlgebraicGraph? It seems like doing it in the AlgebraicGraph could be a bit more straight-forward?

[Leo-Besancon]

Looks good! Thank you! I left a few small comments inline. After these are addressed, we should be good to merge.

One general question: what's the main reason to do constant propagation in MIR rather than in the AlgebraicGraph? It seems like doing it in the AlgebraicGraph could be a bit more straight-forward?

Good point! Initially I just put it there as it was a regression introduced by MIR (expressions previously handled in AST's ConstantPropagation were no longer handled), but it would have been easier in the Air.
But now I think the main reason to keep it here is to ensure that computed indices (#444) are folded before the end of MIR.
This ensures that we do not have to handle vectors in the AIR (or during lowering of MIR to AIR).

[bobbinth]

Looks good! Thank you!

Good point! Initially I just put it there as it was a regression introduced by MIR (expressions previously handled in AST's ConstantPropagation were no longer handled), but it would have been easier in the Air.
But now I think the main reason to keep it here is to ensure that computed indices (#444) are folded before the end of MIR.
This ensures that we do not have to handle vectors in the AIR (or during lowering of MIR to AIR).

Makes sense! Let's still create an issue to investigate moving constant folding into AIR, but maybe keeping a simpler pass for index folding here (not sure it would make sense - but we can discuss that in the issue).

[Leo-Besancon]

Looks good! Thank you!

Good point! Initially I just put it there as it was a regression introduced by MIR (expressions previously handled in AST's ConstantPropagation were no longer handled), but it would have been easier in the Air.
But now I think the main reason to keep it here is to ensure that computed indices (#444) are folded before the end of MIR.
This ensures that we do not have to handle vectors in the AIR (or during lowering of MIR to AIR).

Makes sense! Let's still create an issue to investigate moving constant folding into AIR, but maybe keeping a simpler pass for index folding here (not sure it would make sense - but we can discuss that in the issue).

Perfect, here is the issue: #446