Completely prebuilt riscv rust toolchain

[dvc94ch]

LLD support for RISCV just landed upstream! llvm-mirror/lld@9e54d15 riscvarchive/riscv-lld#1

• Wait for rust's llvm/lld to be updated
• Test lld for the riscv target
• Enable it in the target spec by default
• Ship gdb --enable-targets=all with rustup One step setup #43

[danc86]

FYI I have been testing lld already (from the rust-lld fork) and we have a bunch of fixes we need to do in riscv-rt... I will post patches soon, was struggling to get past that stuff with #[naked] first :-)

[bradjc]

I can help test when this lands in rust nightly.

[danc86]

It looks like the lld patch landed in master. Do we know if they will backport it to the 7.x branch?

If they do, do we know if/when rust will rebase its lld 7.x tree?

[dvc94ch]

@danc86 I don't think anyone is going to backport the patch for us. We can submit a PR against rust-lld since a backport might be accepted.

[japaric]

Instead of backporting the patches you could update the lld fork to the latest upstream version (if it doesn't require updating the llvm fork).

[danc86]

Yeah it might be a bit tricky. I noticed that the upstream release_70 branch of lld has accumulated a bunch of commits that depend on new enum values and header changes in the release_70 branch of llvm. So they would have to updated in sync with each other I think.

I am not familiar enough with how llvm and lld do their branches and releases though...

[dvc94ch]

thanks to @danc86 for his work on lld support! rust-lang/lld#1

[danc86]

Ah, posting rust-lang/lld#1 was easy :-) all the hard work went into rust-embedded/riscv-rt#10 instead 😅

Alex just merged the lld patch as well, so we should be able to update the riscv32imac-unknown-none-elf target to use rust-lld next!

[danc86]

PR to update lld in rust-lang/rust was: rust-lang/rust#53492

It was merged several days ago but there seems to be no new build ready to be rustup'ed just yet.

Once rustup can give you a nightly with working lld I will post another PR to update the riscv32imac-unknown-none-elf target to use it by default.

[dvc94ch]

@danc86 have you tried the c extension recently? This rust-embedded/riscv-rt#8 (comment) likely fixed the issues I was having. I haven't tried it since I'm currently to lazy to build rustc from src 😄

[danc86]

You should be able to test +c by just tweaking compiler flags in .cargo/config, right? Or with a custom target spec. I don't think you need to rebuild all of rust.

I haven't tried it myself, will give it a go over the weekend.

I thought we needed the .push(norelax) thingy when setting the global pointer, don't we? Or does that not matter since rust-lld can't actually do the relaxations right now anyway? :-)

[danc86]

I also want to go back to gdb and figure out why its stack traces are broken... I've got a laundry list of stuff to investigate when I have some free time over the weekend. :-)

[dvc94ch]

You should be able to test +c by just tweaking compiler flags in .cargo/config, right? Or with a custom target spec. I don't think you need to rebuild all of rust.

Libcore won't be built with the +c flag then. But I guess it's not required for testing purposes...

I thought we needed the .push(norelax) thingy when setting the global pointer, don't we? Or does that not matter since rust-lld can't actually do the relaxations right now anyway? :-)

yes that is what I was thinking...

[danc86]

Oh right. I have been using cargo xbuild to rebuild core whenever I fiddle with the target. It seems to work.

[japaric]

@danc86 @dvc94ch how did the LLD experiments go? People were asking about this (switching the RISCV targets to LLD) on reddit.

[dvc94ch]

Just started testing it today, can't say yet. Building a small example soc that provides embedded hal drivers in a plug and play fashion. So you select the components and get a hal that is for the soc.

[dvc94ch]

Opened a PR rust-lang/rust#53822

[therealprof]

Meanwhile there are a variety of different toolchains available out-of-the-box: https://doc.rust-lang.org/nightly/rustc/platform-support.html

So I think this issue can be considered done.