incremental compilation + proc macros: undefined symbol error

[rukai]

error: /home/rubic/Projects/Crates/syn/examples/heapsize2/example/target/debug/deps/libheapsize_derive-3ef99e9f83ff68d2.so: undefined symbol: __rustc_derive_registrar__59da0d2c3725c53af805cb0071a3b093_24
 --> src/main.rs:2:1
  |
2 | extern crate heapsize_derive;
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Disabling incremental compilation or deleting the build files fixes the issue.

$ rustc --version --verbose
rustc 1.25.0-nightly (b5392f545 2018-01-08)
binary: rustc
commit-hash: b5392f54503fdaf04df4b9578510b2baa944f4af
commit-date: 2018-01-08
host: x86_64-unknown-linux-gnu
release: 1.25.0-nightly
LLVM version: 4.0

Steps to reproduce

1. git clone https://github.com/dtolnay/syn
2. cd syn/examples/heapsize2/example/src/ && CARGO_INCREMENTAL=1 cargo run (build succeeds as expected)
3. at syn/examples/heapsize2/heapsize_derive/src/lib.rs:20 insert a line "foo()"
4. cd syn/examples/heapsize2/example/src/ && CARGO_INCREMENTAL=1 cargo run (build fails as expected)
5. at syn/examples/heapsize2/heapsize_derive/src/lib.rs:16 insert a line "fn foo() { }"
6. cd syn/examples/heapsize2/example/src/ && CARGO_INCREMENTAL=1 cargo run (build fails with error above)

In case the changes to source code above weren't clear, refer to the commits 1 & 2 https://github.com/rukai/syn/commits/bug

[michaelwoerister]

Thanks for the bug report, @rukai!

[michaelwoerister]

Let's see if #47181 fixes this...

[dtolnay]

@michaelwoerister #47181 did not fix this. I am still seeing the same thing on rustc 1.25.0-nightly (e6072a7 2018-01-13) which contains that PR.

[michaelwoerister]

Thanks a lot for verifying.

Nominating for priority assignment (I'd say P-High).

[michaelwoerister]

A couple of questions for people knowing about proc-macros (@dtolnay, @jseyfried, @alexcrichton, maybe?):

• Is there ever going to be more than one macro_derive_registrar function per crate?
• Is the symbol naming scheme fixed somewhere or can we change it?

[alexcrichton]

@michaelwoerister there should only be one per crate which is generated by the compiler, so we should have complete control over it.

[alexcrichton]

Syntactically it's manufactured here

[michaelwoerister]

Ok, so generate_derive_registrar_symbol() doesn't actually have to factor the registrar's DefIndex into the symbol name in order to make it unique in the crate graph. Removing the DefIndex should fix this bug.

[michaelwoerister]

I know what's going on here now: The registrar function is generated in its own module, which means that it is not recompiled during incremental compilation unless its contents change. The symbol for the registrar contains the raw DefIndex though, so placing anything with a DefIndex before the generated registrar will change the DefIndex of the registrar and thus also change the symbol we expect it to have. But the object file has been cached, so the actual symbol name does not change. This leads to a mismatch between the actual symbol name and the symbol name that we write into crate metadata.

I think the best solution is to just remove the DefIndex from the symbol name since it is redundant anyway. The same probably goes for regular plugins?