[bootstrap] Don't reverse the order of dylib search path entries

[jnkel]

The add_dylib_path helper function prepends paths to the beginning of the dynamic linker search path, but it reverses their order while doing so. This is surprising and undocumented, and seems to be unexpected by several callers of this function.

Particularly, rustc_lib_paths appends the ci-llvm path to the list it returns; reversing the order puts ci-llvm at a higher priority than the compiler's lib directory. On my system, this currently causes ./x test to fail (when building the unstable book), because the stage0 compiler is run using CI LLVM instead of stage0 LLVM (which are currently different because stage0 is on LLVM 22 while main is on LLVM 23).

Might be worth a try build as this change could potentially cause issues if there is somewhere we depend on this ordering reversal. I checked all the call-sites (and ran ./x test locally) and I don't think anyone intentionally relied on the ordering being reversed. I did find one snippet that concerned me (from #144303, cc @Kobzol):

rust/src/bootstrap/src/core/build_steps/test.rs

Lines 437 to 448 in f7d782a

	// The `cargo` command configured above has dylib dir path set to the `build_compiler`'s
	// libdir. That causes issues in cargo test, because the programs that cargo compiles are
	// incorrectly picking that libdir, even though they should be picking the
	// `tested_compiler`'s libdir. We thus have to override the precedence here.
	let mut existing_dylib_paths = cargo
	.get_envs()
	.find(|(k, _)| *k == OsStr::new(dylib_path_var()))
	.and_then(|(_, v)| v)
	.map(|value| split_paths(value).collect::<Vec<PathBuf>>())
	.unwrap_or_default();
	existing_dylib_paths.insert(0, builder.rustc_libdir(tested_compiler));
	add_dylib_path(existing_dylib_paths, &mut cargo);

The comment states we're inserting builder.rustc_libdir(tested_compiler) at the highest priority in the search path, but because add_dylib_path reversed the ordering, it's actually inserted at the lowest priority. I don't know how to reproduce the issue this was supposed to fix, so I can't be sure that this PR doesn't cause a regression.

Follow-up to #161335. r? @jieyouxu

[Kobzol]

The code in add_dylib is 11 years old and comes from bootstrap's first Rust commit. I agree that what you describe makes more sense.

Also, this is such a mess.. the ci-llvm directory really shouldn't be added in this function, that's super ad-hoc.

[jieyouxu]

(I'll look at this on Friday)

[jieyouxu]

Thanks. I did some history digging and could not find why this needed to be reversed. This order looks more correct to me.

@bors r+ rollup

View changes since this review

[rust-bors]

📌 Commit 67e8a19 has been approved by jieyouxu

It is now in the queue for this repository.