AVR support

[japaric]

Parent issue: #2

Next steps:

• Wait until AVR support lands in upstream LLVM
• Update rust-lang LLVM fork (Upgrade to LLVM 4.0 rust-lang/rust#37609), implement cabi stuff, etc.
• Experiment with custom targets out of tree (to get the specification right)
• Get libcore compiling without modifications
• Upstream all LLVM patches required for the AVR fork (Upstream all of the patches for llvm-8.0svn support avr-rust/rust-legacy-fork#116)
• Write documentation about writing/compiling/flashing/debugging Rust
  programs for this target
• Add the AVR target to the compiler (as a built-in target) (tracking issue RFC: Merging the avr-rust fork upstream rust-lang/rust#44052)
• 🎉

See also avr-rust/rust for out of tree work towards AVR support in Rust.

[dylanmckay]

Here's an email I sent out to the llvm-dev list very recently regarding the status of the AVR backend.

tl;dr there's really only 3 files needed before the backend will be 80% usable.

[japaric]

Hey, @dylanmckay. I'm wondering what's the current status. I saw that there's only one "PR" (differential revision) in the review queue and that it has already been accepted but has not landed yet.

After that lands, will LLVM be ready to compile Rust programs for AVR? Will you (and @shepmaster) be sending a PR updating Rust's LLVM next?

Happy to help with anything to land this in rust-lang/rust. 😄 (I should finish landing my PR that moves rust-lang/rust to the WIP Rust port of compiler-rt intrinsics; I suspect that it may help here)

[shepmaster]

a PR updating Rust's LLVM next?

Eventually, yeah, except the LLVM 4.0 changes are more annoying than any previous LLVM changes I've had to make.

moves rust-lang/rust to the WIP Rust port of compiler-rt intrinsics

I think it will help. Especially because one "bug" we have at the moment is that I've forgotten to publish the WIP branch of compiler-rt so people are getting build failures.

[japaric]

Eventually, yeah

I'm really looking forward to LLVM 4.0 in rust-lang/rust so if you make it happen you'll have my gratitude.

the WIP branch of compiler-rt

Did you have to implement new intrinsics? I only see build system related changes in your fork.

[shepmaster]

I'm really looking forward to LLVM 4.0

What benefits do you expect to see from it?

implement new intrinsics

I think we might have to, but haven't yet.

[japaric]

What benefits do you expect to see from it?

I'm working on embedding lld, the LLVM linker, into rustc and having LLVM 4.0 lets me use the latest lld release which has better linker script support. (Linker script support in lld 3.9 is minimal) I'm hoping that someday I'll be able to build Rust programs for Cortex M devices without having to install arm-none-eabi-gcc first.

[dylanmckay]

I saw that there's only one "PR" (differential revision) in the review queue and that it has already been accepted but has not landed yet.

I'm about to commit that one now :)

After that lands, will LLVM be ready to compile Rust programs for AVR? Will you (and @shepmaster) be sending a PR updating Rust's LLVM next?

It should be able to compile very trivial instructions. There are two source files that will fix that.

• AVRBranchSelector - The branch relaxation pass (which will be rewritten to use the generic relaxation pass before it is merged)
• AVRExpandPseudoInsts - This is required for basically any 16-bit operation (hint: almost everything). I've put this one off until last because we can compile really basic examples without it. On top of that is that it is essentially a 2000 line hack full of code duplication.

[raphaelcohn]

@japaric Big shout out for lld embedded into rustc. It's not an easy thing to do - I know I'd struggle - but a rust toolchain that's entirely independent of GNU binutils and gcc is an enormous win for rust.

[japaric]

@raphaelcohn Check rust-lang/rust#36120. It's pretty much already working on Linux and it can link Cortex-M programs without arm-none-eabi-gcc. It hasn't landed because I couldn't get lld to build properly on Windows 😞

a rust toolchain that's entirely independent of GNU binutils and gcc is an enormous win for rust.

Indeed! If all goes well we'll also be able to (cross) compile statically linked MUSL binaries without needing glibc installed on the system as well. Ahh, the dream of the "zero installation steps" cross compilation setup may soon come true.

[dylanmckay]

• The machine code test suite is now fully passing and currently in review (D26387)
• About half of the code generation tests are fully passing and are also currently in review (D26388)