avr: Skip No More!

[Patryk27]

This pull request gets rid of all of the #[avr_skip]s introduced in the past.

My heart uC skips, skips, skips, skips, skips, skips a beat an intrinsic

... Olly Murs, frustrated about having to pull someone's intrinsics when working on AVR firmwares (colorized).

Changes here are related to #711, but note that we continue not to provide __u?divmod[hq]i4 - it shouldn't be terribly difficult to implement those (I even have a work-in-progress branch in rustc that exposes the proper calling convention), it's just a different kind of change that will be done in another pull request.

In general, the changes boil down to:

• removing overzealous #[avr_skip] on intrinsics that don't actually have special convention (such as __addsf3),
• implementing proper support for the other intrinsics (such as __unordsf2 or __divmodsi4).

I've tested new code via avr-tester - in particular, there's a test that uses randomized math operations and another one that checks floating point math, and both pass using the implementations provided by compiler-builtins.

Related discussions:

• Problem with floats on atmega32u4 - `undefined reference to '__extendsfdf2'` Rahix/avr-hal#641
• Native support for AVR instrinsics #711

[tgross35]

Why are all the new use super::* imports required?

[Patryk27]

So that the CmpResult is in scope (see the discussion above on intrinsics! generating a new module for each function); we could use fully qualified paths, but that makes the code less readable imo.

[tgross35]

Thanks, that makes more sense. I think I would prefer use crate::float::cmp::CmpResult here, or -> crate::float::cmp::CmpResult in float::cmp since a glob import in the macro sounds like a recipe for future confusion. It would also be fine to reexport CmpResult back up to the top level too so that gets shortened to crate::CmpResult.

[tgross35]

Also did you mean for this PR to have draft in the title, but not actually be marked as a draft? :)

[Patryk27]

Also did you mean for this PR to have draft in the title, but not actually be marked as a draft? :)

Hah, #justgitlabthings

[tgross35]

Also,

(I even have a work-in-progress branch in rustc that exposes the proper calling convention)

This shouldn't be necessary if there are only a couple of intrinsics. Instead, #[naked] + naked_asm! can be used, and the calling convention is just done in the assembly. The intrinsic implementation can be done in asm (probably better for performance), or trampoline to an extern "C" function we define.

[Patryk27]

Ok, it's ready!

I've tested the code on ATmega328P and ATtiny167 (just simavr, since I've recently relocated and my electrical stuff is still in boxes in some warehouse, waiting for final delivery 😭).

[tgross35]

Sorry about the conflicts, this will need a rebase since I moved some things around (git should be able to figure it out). Just that and #791 (comment), then I think this will be ready to go.

[Patryk27]

Okie, ready; I also added a comment to the CmpResult alias that links to LLVM, so that it's clear where the types come from.