Jeremy Kun

SIGH ok there's another problem. Now that the lowering includes non-secret ops, the plaintext lowering fails when combining secret-forget-secrets with arith-to-mod-arith (the cleartext packing operations get converted to mod arith...

OK so I resolved the issue as per our discussions, and ended up disabling some of the debug tests involving floats. Let's get this in and I can fix that...

This paper also seems to have a lot to say about polynomial evaluations of low multiplicative depth. Pinging https://github.com/google/heir/issues/1565 to link the issue here.

I read this paper on the flight to HES, I think the maxsat approach will be too slow to run at compile time, but I did talk a lot with...

After discussing this in office hours, I think we came to the following agreement: 1. We should prefer using linalg ops when possible, because they are simpler for the noise...

I think we will ultimately do a lot better by just handling polynomial approximation and evaluation more generally. My rough plan is to add a polynomial.evaluate op, which can then...

In other words, while I wouldn't object to a patch for this, we may as well focus our efforts on the longer term plan.

Cf. @AlexanderViand if the upstream changes are accepted, then I think this will be a much simpler way to support `mod_arith.constant`.

Branch where I started doing this before checkpointing as a TODO: https://github.com/j2kun/heir/tree/collapse-shape-op-alignment-prop

I have seen this when the full xcode is not installed.