Update to nightly-2021-09-11

[ghost]

There are quite a few changes in rustc I had to address, most notably black_box is now a backend-driven intrinsic. No more hacks for llvm_asm!

I haven't yet looked at the codegen tests: no spare time, sorry. Hopefully, this evening.

[antoyo]

Thanks!

[ghost]

(There are not enough hours in the day)

I have good news, bad news, and bad news (it's midnight and I'm a little bit unimaginative at midnights; I'll catch up another day).

Good news: I've been able to verify that the clobber_api is indeed correctly translated to out("reg")_ operands, transparently to the backend (I compiled a simple separate file). Also, turns out that not every register's name as given by rustc is a valid GCC name for the clobber list. I'm sending a separate PR for it.

Bad news: running codegen tests in requires being able to save intermediate artifacts to a file (judging by this). I haven't find a way to save GIMPLE to a separate file, or anywhere but stderr. (I checked your fork as well). Related to #81.

Bad news: clobber_api("C") clobbers k[1-7] AVX-512 registers, which the docs of course do not mention. I'm starting to love open-source world, and it's very weird, kiss-kill kind of love. Unfortunately, GCC does not allow to clobber these registers unless the -mavx512f option is passed, and I see gccjit::context does not support testing for added options. Another patch to libgccjit...

[bjorn3]

Bad news: clobber_api("C") clobbers k[1-7] AVX-512 registers

I think it should be changed to only do so when AVX-512 is actually enabled for the current function I think. It makes sense that GCC doesn't allow it when AVX-512 isn't enabled. I am slightly suprised that LLVM does allow it. cc @Amanieu.

which the docs of course do not mention.

The actual code for this is https://github.com/rust-lang/rust/blob/3ed6c1d23fd40f4367259a531465e809eb00ec27/compiler/rustc_target/src/asm/mod.rs#L802 which does indeed have k1-k7. Guess the docs need to be updated.

[Amanieu]

Yes, LLVM simply ignores any invalid clobbers. So for GCC you would just have to gate the clobbers based on what target features are enabled.

[Amanieu]

Note that this only applies to clobbers, LLVM still requires outputs (even dummy ones) to be valid registers for the current target.

This is the code that handles this in the LLVM codegen backend: https://github.com/rust-lang/rust/blob/master/compiler/rustc_codegen_llvm/src/asm.rs#L175

[ghost]

I see gccjit::context does not support testing for added options. Another patch to libgccjit...

I think this was wrong. We can just check features provided by the frontend as suggested by @Amanieu .

[antoyo]

@Commeownist: Is it okay for you if I fix the remaining reviews in order to merge this PR? I'd like to merge it soon to update the PR in rustc to move things forward.

[ghost]

@antoyo Cool.

[ghost]

@antoyo I've resolved reviews (except I'd like to leave the TODO comment) and updated to nightly-2021-09-17. I think you can just merge it as is.

I'll leave codegen tests for another day.

[antoyo]

There's an error in the CI which seems unrelated: it's about the bootstrap.

[ghost]

@antoyo Are you sure that rust-toolchain file is respected? These lines from the log make me somewhat concerned:

downloading https://static.rust-lang.org/dist/2021-09-08/rustfmt-nightly-x86_64-unknown-linux-gnu.tar.xz
...
extracting /home/runner/work/rustc_codegen_gcc/rustc_codegen_gcc/rust/build/cache/2021-09-08/rustfmt-nightly-x86_64-unknown-linux-gnu.tar.xz

[ghost]

I suspect it's related to those wonderful git checkout <hash> commands all over test.sh.

[antoyo]

Does this work locally? Perhaps the CI is having issue…

[bjorn3]

@antoyo Are you sure that rust-toolchain file is respected? These lines from the log make me somewhat concerned:

downloading https://static.rust-lang.org/dist/2021-09-08/rustfmt-nightly-x86_64-unknown-linux-gnu.tar.xz
...
extracting /home/runner/work/rustc_codegen_gcc/rustc_codegen_gcc/rust/build/cache/2021-09-08/rustfmt-nightly-x86_64-unknown-linux-gnu.tar.xz

Rustc uses a nightly rustfmt instead of the version associated with the bootstrap compiler. I don't quite remember why though.

[bjorn3]

Bootstrap failed due to a new warning about an unused field that doesn't exist on beta. You could add deny-warnings = false to the [rust] section of config.toml:

rustc_codegen_gcc/test.sh

Line 150 in 8ec7976

[build]

[ghost]

@antoyo Ready

[antoyo]

Thanks!