[RFC] Allow DiffPairs and Pins to be inverted individually

[anuejn]

Currently, DiffPairs and Pins can be either inverted completely or not at all.
This is a problem in corner cases where the hardware inverts some parts of a logical Bundle of ios (for example to ease routing in case of differential pairs; see #386).

I hereby propose to extend the Platform DSL in the following way:

• Add a invert= kwarg to the constructor of DiffPairs
• Allow Tuples of booleans of the width of the pin list to be passed to the invert= kwarg of both DiffPairs and Pins. Each bool in the tuple then controls the inversion of one pin / differential pair.

[whitequark]

Right, this is the obvious solution. The main issue I see with it is that when you have a lot of diffpairs, it can be hard to correlate the booleans in the invert= tuple with the actual diffpairs.

[whitequark]

As an alternative I was wondering if something like DiffPairs("A1 !A2", "B1 !B2") would work. The ! sigils would then have to match in both the P and the N pin lists. Of course, invert=True should still invert all of them, and conversely, it probably makes sense to accept ! for single-ended pins too.

[adamgreig]

I like putting it in the pin list string(s), it's compact but I think also quite clear. Maybe "~A1" would be a better match to the nmigen negation syntax?

[whitequark]

Sure, ~ is fine by me.

[anuejn]

I also really like that idea :)

[whitequark]

Sounds great. I'll bring this up on the next meeting but I think you can start implementing it as it seems unlikely anyone's going to be opposed.

[anuejn]

sounds good :)

[anuejn]

I think we should support similar syntax for the use with connectors (which can be routed through e.g. inverted level shifters or have crossed differential pairs). So I think it would be nice to have syntax elements in connectors that can say "invert this signal if it is a diff pair" or "invert this signal always".

I propose that we use ~ for always invert and ? for invert in differential pairs in connectors.

[whitequark]

I would like to avoid introducing that functionality (and the associated syntactic burden) until someone actually needs it.

[anuejn]

/me needs the latter (inverted differential pairs in connectors)

[whitequark]

Okay. I still think it should be done separately because while this one is fairly straightforward (it's just a way to write a tuple that's less error prone than usual), what you request with connectors requires completely rethinking the way we do connectors.

[anuejn]

Okay I see that.
So I will implement this RFC as-is and go with out of tree hacks for now.
If I really need this feature to become upstream I will write another RFC :)

[whitequark]

I think we can do it upstream; I said earlier that I'd like to avoid introducing it because it sounded theoretical (which is usually bad because features done in abstract tend to not survive collision with reality). If you actually hit that problem I think we can certainly add it. Let's discuss the connector semantics on the next meeting?