Add RUSTC_EMIT option to pass on --emit to crates during bootstrap

[Zoxc]

This makes it a bit more convenient to inspect the generated code of rustc crates. I also have an upstream patch to add demangling support to llvm-reduce which pairs well with this to make the LLVM IR output more readable.

[rustbot]

r? @albertlarsan68

(rustbot has picked a reviewer for you, use r? to override)

[albertlarsan68]

I have a few remarks:

• Have you tested that rustc is built correctly while dumping those representations?
• Is there another way to do this currently/maybe it can be useful to generalize this principle to the other emit values (MIR, HIR, THIR, etc.)
• Can you add documentation in the right places (cc @ozkanonur for guidance)

Note for future/follow-up/now:

• Allow to set exactly which format is to be output, I fear that with all the disk gets filled rapidly (maybe unfounded fear but having 6 copies of the same program is not the best thing to have (currently the code itself, incremental compilation data, and final objects, with this PR add ASM output, LLVM IR and bitcode)).

[Zoxc]

x.py build library seems to work at least. I haven't done any more extensive testing.

The --emit option supports MIR, so that would be easy to add.

I could add the ability to specify the output formats with a syntax like rustc_arena=mir,llvm-ir;rustc_infer=llvm-bc.

[albertlarsan68]

x.py build library seems to work at least. I haven't done any more extensive testing.

Can you test building a stage 2 compiler and using it?
If it is too expensive for you, I can do it.

The --emit option supports MIR, so that would be easy to add.

Yes please!

I could add the ability to specify the output formats with a syntax like rustc_arena=mir,llvm-ir;rustc_infer=llvm-bc.

This would be great!
If you add MIR support, this addition would be very useful.

[onur-ozkan]

Can you add documentation in the right places (cc @ozkanonur for guidance)

The related documentation place for this context points to src/bootstrap/README.md file. Later on we can provide helpful informations in dev-guide, but that's outside of this PR's scope.

[Zoxc]

I've renamed the environmental variable to RUSTC_EMIT. It now uses uses the syntax I proposed and it just passes the format on to rustc so mir is also supported.

The dev guide seems to be the right place to add documentation as src\bootstrap\README.md explicitly says it's not usage documentation.

[albertlarsan68]

Post a PR to the rustc dev guide, and apply my comments if you want.

[albertlarsan68]

As a follow-up, make this configurable to be able to output ATT syntax.

[Zoxc]

I've opened a PR to add documentation.

[albertlarsan68]

It would also be great if you add an entry to the changelog for bootstrap (located in src/bootstrap/CHANGELOG.md IIRC).

[Zoxc]

Done.

[bors]

☔ The latest upstream changes (presumably #109164) made this pull request unmergeable. Please resolve the merge conflicts.

[jyn514]

Why not use RUSTFLAGS in combination with x clean rustc_data_structures (or whatever crate you're interested in)?

[Zoxc]

RUSTFLAGS is not per crate and I don't want to dig through cargo/rustc's output files to find the emitted files.

[albertlarsan68]

It seems like rustc does not respect the output dir everytime. For example, LLVM IR and ASM do still output in the "normal" target directory, for example to build/host/stage1-std/x86_64-unknown-linux-gnu/release/deps/compiler_builtins-acc40b1213caf41b.compiler_builtins.6d7cb61c-cgu.0.rcgu.s for the ASM output for compiler_builtins.

[bors]

☔ The latest upstream changes (presumably #111402) made this pull request unmergeable. Please resolve the merge conflicts.

[bjorn3]

That may be a result of the hacks we use to force compiler-builtins to build with multiple codegen units. --emit asm/llvm-ir/llvm-bc/obj normally forces a single codegen unit for back compat reasons with when we didn't use multiple codegen units for a single crate yet.

[jyn514]

In the meantime, you could use x clean rustc_foo && MAGIC_EXTRA_RUSTFLAGS=--emit=... to avoid invalidating the build cache and only pass the flags to a single crate.

[JohnCSimon]

triage:
@Zoxc can you please fix the merge conflict

FYI: when a PR is ready for review, send a message containing
@rustbot ready to switch to S-waiting-on-review so the PR is in the reviewer's backlog.

[Zoxc]

I think I'll close this for now. It doesn't interact well with multiple CGUs and LTO. It might be better to base it on save-temps or overhaul the IR emitting command line option.