Object file in /tmp produces non-deterministic builds on Windows for crate-type=bin

[danakj]

When building a --crate-type=bin target with rustc, an object file is produced in /tmp/<randomname> and that absolute path is then passed to the linker.

Windows puts paths to all object files in the PDB file, so this results in non-deterministic Windows builds (see #88982 (comment)). To be deterministic all paths must be relative paths, or fixed paths.

To resolve this, we could:

1. Grow a command line to specify the path to the object file that rustc should write, and then pass to the linker.
2. Grow a command line to specify the temp root (replacing /tmp or similar when specified), allowing it to be specified relative to the CWD. Then ensure -Zremap-cwd-prefix gets applied to it in order to produce a relative path.
3. Other??

(FWIW I am cross-compiling Linux to Windows here which is why you see /tmp in the path :) But same things occur building on Windows.)

[danakj]

cc: @michaelwoerister

[bjorn3]

One option would be to move the object file(s) that currently end up in /tmp to be next to the output like we already do for codegen units.

[danakj]

Would you have any pointers to the code that does this for codegen units?

[bjorn3]

rust/compiler/rustc_codegen_ssa/src/lib.rs

Line 82 in 4bd4e2e

let object = emit_obj.then(|| outputs.temp_path(OutputType::Object, Some(&self.name)));

and the code that emits the object file in /tmp is

rust/compiler/rustc_codegen_ssa/src/back/link.rs

Line 1919 in 4bd4e2e

let path = tmpdir.join("symbols.o");

I believe.

[bjorn3]

@rustbot label +A-reproducibility +O-windows-msvc

(@danakj if you find other reproducability issues feel free to use this command to add the right label. most labels can be set by everyone.)

[michaelwoerister]

Yes, that should be fixable the way @bjorn3 suggests. I can't think of a reason why an object file would need to be in /tmp.

I'll assign the issue to me. @danakj, let me know if you plan to look into a fix.

[bjorn3]

cc #71714 which has the same cause.

[zmodem]

@michaelwoerister have you had a chance to look into this? Otherwise I might give it a try, though this is unfamiliar territory for me.