[LowerComb] Truth table to mux tree lowering

[teqdruid]

Add a truth_table to mux tree lowering to the LowerComb pass.

[sequencer]

FYI TruthTable already has an PLA-based implementation: https://github.com/chipsalliance/chisel/blob/435d5cc43a0b49038ab4615383fc58be224d9845/src/main/scala/chisel3/util/pla.scala#L7

Do you think the truth table should conclude dont care as well?
If so, we may need to consider introducing QMC and espresso to CIRCT.

[teqdruid]

@darthscsi @mikeurbach @seldridge Do you agree with adding a pass to lower this? I'll add the 'casez' lowering later.

[teqdruid]

Do you think the truth table should conclude dont care as well?

We specifically discussed this at a recent ODM and decided against it, at least to start.

[sequencer]

Using mux-based to implement a LUT might be inefficient for ASIC designs, it that possible to be configured to add other implementations which may be designed for ASIC specifically in the future?

[teqdruid]

Using mux-based to implement a LUT might be inefficient for ASIC designs, it that possible to be configured to add other implementations which may be designed for ASIC specifically in the future?

Yes. We'd add a pass option to specify a lowering style -- mux vs casez vs primitive vs .... Casez is planned.

[sequencer]

Sounds great, but if not having dont care bits, would casez version be efficient? Or the dc will get implemented in the future?

[teqdruid]

casez with selected/computed ? bits is required to get us the xprop behavior we want. See the June 7th ODM for that discussion.

[sequencer]

Thanks for information!

[mikeurbach]

Having a lowering with the Comb dialect in general makes sense to me. Just left a few comments. I'm not exactly sure what you need to lower to in order to achieve what you want, but the general implementation looks good.

[teqdruid]

#5433 <- pass boilerplate PR

[teqdruid]

Having a lowering with the Comb dialect in general makes sense to me. Just left a few comments. I'm not exactly sure what you need to lower to in order to achieve what you want, but the general implementation looks good

A tree of trinary operations works as does a casez with ? used in places where an x input doesn't cause the output to change. Or a casex, though there are other problems with casex and most tools issue a warning on all casex statements. That's all we've found so far.

[teqdruid]

Since I've already gotten a bunch of feedback, I'll be relying on post-review comments for further revisions.