Smarter Adjoint Handling in SquinToStimPass (and in general) #468
Description
I ran into this problem debugging the SquinToStimPass in @david-pl 's ongoing PR #436 which uses all squin statements against relevant passes/analyses to ensure they're all accounted for.
I suppose this RFC is a combination of asking what the best way to resolve this is and pushing for something I thought of some time ago but feel wouldn't hurt to have now.
The Adjoint statement in squin has a boolean attribute that determines if it's unitary. By default, it's set to False. However, there has to be some logic to be able to set it to True when the operator inside the adjoint is unitary. The motivating example for this is currently if I have something like:
op = adjoint(sqrt_x)
This will not get properly expressed as SQRT_X_DAG in the rewrite to Stim. The reason is because there is some guard logic I believe @kaihsin implemented where if the adjoint is not unitary, then the squin qubit dialect to stim rewrite immediately stops.
I can see the following solutions to the problem:
- Just make the guard a bit more relaxed and check if what's inside the adjoint is a SqrtX, SqrtY, or SqrtZ to allow things to go through
- Have a rewrite that looks at instances of such adjoints and fixes up the boolean to reflect the unitarity
This brings me to another idea that I don't think is too much work. This is more directly related to what @david-pl has in his test where he has adjoint(x) (x being Pauli X). There's currently no adjoint canonicalization pass but I think it would quite trivial to have some rewrite that can see see such things and realize the adjoint can be removed altogether or a chain of adjoints can be collapsed.
Please feel free to let me know if I should take the simplest route here to make the SquinToStimPass happy or if it's worth killing two birds with one stone 🪨 🐦 🐦