Figure out why `x.py test --stage 0 rust-analyzer` doesn't work

[fasterthanlime]

(This issue only makes sense after #99603 lands)

@jyn514 is convinced it should if the proper sysroot is used to compile both the proc-macro-srv and proc-macro-test-impl crates.

There's one nested invocation of rustc in proc-macro-srv's build script, but it uses $RUSTC (all it does is grab the trimmed output of rustc -v).

There's one nested invocation of cargo in proc-macro-test's build script (to build proc-macro-test-impl in a temp dir), but it uses $CARGO.

Here's the output of ./x.py test --stage 0 -vv rust-analyzer --test-args like_macro_clone_literals: https://gist.github.com/fasterthanlime/3de46b62906d0762c8c3ac0e295892a4

[jyn514]

fix here:

diff --git a/src/bootstrap/tool.rs b/src/bootstrap/tool.rs
index 7c37621d417..0f20b74bcb4 100644
--- a/src/bootstrap/tool.rs
+++ b/src/bootstrap/tool.rs
@@ -250,6 +250,15 @@ pub fn prepare_tool_cargo(
         }
     }
 
+    if path.ends_with("rust-analyzer") {
+        // RA wants to load proc macros at runtime for one of its tests. But by default, proc macros
+        // are built with the beta proc_macro, not the in-tree proc_macro. Explicitly override this
+        // in the rustc shim.
+        // This is ok because rust-analyzer is `Mode::ToolStd`, so proc_macro is always built before.
+        cargo.rustflag("--sysroot");
+        cargo.rustflag(builder.sysroot(compiler).as_os_str().to_str().unwrap());
+    }
+
     // clippy tests need to know about the stage sysroot. Set them consistently while building to
     // avoid rebuilding when running tests.
     cargo.env("SYSROOT", builder.sysroot(compiler));

will try to write this up more thoroughly soon

[jyn514]

Oh fun, this isn't possible to test unless the proc-macro ABI changed in the last release 🤦

[onur-ozkan]

This no longer seems to be a problem.

Tested on HEAD: f736079

x test --stage 0 rust-analyzer:

running 13 tests
.............

test result: ok. 13 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 1.27ms

        finished in 9.789 seconds

[jyn514]

you need to combine it with a breaking change to the proc macro lib format in the same release cycle.

maybe that's rare enough this isn't worth worrying about?

[onur-ozkan]

you need to combine it with a breaking change to the proc macro lib format in the same release cycle.

Sounds similar to #119172 (comment) problem (when public api changes on compiler crates, ra fails to build).

maybe that's rare enough this isn't worth worrying about?

I think so

[jyn514]

Sounds similar to #119172 (comment) problem (when public api changes on compiler crates, ra fails to build).

kinda, but in this case a sync will only fix it for stage 1, not for stage 0