Type-erased peripherals

[FeldrinH]

Currently there seems to be no way to obtain type-erased peripherals. Consider for example that I have Qei<'d, TIM1>, Qei<'d, TIM2> and Qei<'d, TIM3>. I would now like to put them all in an array. However, for this I need them all to have the same type.

It would be nice to have someting like AnyTIM, where I could call TIM1.erase(), TIM2.erase(), etc to turn them all into AnyTIM and then have an array of Qei<'d, AnyTIM>.

PS: I think AnyTIM is better than AnyQei because it's more flexible. If I write code that is generic over T in Qei<'d, T>, then that would work with both concrete TIMx types as well as AnyTIM. With AnyQei there is no easy way to write code that is generic over both Qei and AnyQei. Also, I think AnyTIM is just a nicer abstraction, because conceptually there aren't two different Qei types, there is one Qei type that works for either a specific timer or any timer.

[Dirbaio]

This is a very wanted feature!

It's already been done for some drivers, see this PR for uart, for example: #2836 It works slightly differently to your proposal: instead of adding AnyUart etc to make specifying the uart in a generic param optional, it completely removes the generic param.

The reason AnyUart/AnyTimer etc wouldn't work is we can't type-erase before creating the driver because then we can't check the given pins and DMA channels are valid at compile time when creating the driver.

Timers are a bit tricky since there's so many variations. We'd need a generic param to "keep" information about the capabilities such as Timer<'d, Basic>, Timer<'d, Advanced>, similar to the Blocking/Async mode param in uart.

For QEI it could be just Qei<'d> since you simply can't create it with a non-qei-capable timer.

[FeldrinH]

Would it be possible to erase after the fact, so you create e.g. Timer<'d, TIM1> and then call timer.erase() to get Timer<'d, Advanced>?

Getting rid of the generic parameter alltogether would be great as well. I wonder if it has a performance impact, since the register block address and other metadata presumably needs to be stored in the driver struct?

[Dirbaio]

I did some code size measurements back then and the erased version wasn't bigger #980 (comment)

I didn't measure perf, but my guess is it doesn't make a difference either. In the end it's about loading a pointer from self vs loading a pointer from a constant pool.