With compiler 1.60.0:

$ rustc --version rustc 1.60.0 (7737e0b5c 2022-04-04)

The contents of teh .stderr files is slightly different so the .stderr files needed updating using TRYBUILD=overwrite cargo test.

Without this change I was getting errors like the following, with the change the tests pass.

  test tests/ui/has-arguments.rs ... mismatch

  EXPECTED:
  ┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈
  error: this attribute takes no arguments

    = help: use `#[contracts]`

   --> $DIR/has-arguments.rs:3:1
    |
  3 | #[contracts(a, b)]
    | ^^^^^^^^^^^^^^^^^^
    |
    = note: this error originates in the attribute macro `contracts` (in Nightly builds, run with -Z macro-backtrace for more info)
  ┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈

  ACTUAL OUTPUT:
  ┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈
  error: this attribute takes no arguments

           = help: use `#[contracts]`

   --> tests/ui/has-arguments.rs:3:1
    |
  3 | #[contracts(a, b)]
    | ^^^^^^^^^^^^^^^^^^
    |
    = note: this error originates in the attribute macro `contracts` (in Nightly builds, run with -Z macro-backtrace for more info)
  ┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈┈
  note: If the actual output is the correct output you can bless it by rerunning
        your test with the environment variable TRYBUILD=overwrite

  test tests/ui/has-expr-argument.rs ... mismatch

I spent a little time looking at why when I run TRYBUILD=overwrite carto test the *.stderr files do not have $DIR/xxx in the output. My conclusion is the compiler use to include the $DIR/xxx output, but now outputs the actual path.

You can see that in this change on trybuild the stderr test files were updated and they do not have the $DIR/xxx:

wink@3900x 22-05-18T00:16:29.992Z:~/prgs/rust/clones/dtolnay-trybuild (master)
$ git diff 6cbc5a2843a66dfe934479cf2cf31afbe4b4dbb3^ 6cbc5a2843a66dfe934479cf2cf31afbe4b4dbb3
diff --git a/tests/ui/compile-fail-2.stderr b/tests/ui/compile-fail-2.stderr
index a59ae39..daa33d4 100644
--- a/tests/ui/compile-fail-2.stderr
+++ b/tests/ui/compile-fail-2.stderr
@@ -1,5 +1,5 @@
 error: ERROR
- --> $DIR/compile-fail-2.rs:1:1
+ --> tests/ui/compile-fail-2.rs:1:1
   |
 1 | compile_error!("ERROR");
   | ^^^^^^^^^^^^^^^^^^^^^^^^
diff --git a/tests/ui/compile-fail-3.stderr b/tests/ui/compile-fail-3.stderr
index 27619cf..7fc2291 100644
--- a/tests/ui/compile-fail-3.stderr
+++ b/tests/ui/compile-fail-3.stderr
@@ -1,5 +1,5 @@
 error[E0277]: `*mut _` cannot be shared between threads safely
-   --> $DIR/compile-fail-3.rs:7:5
+   --> tests/ui/compile-fail-3.rs:7:5
     |
 7   |     thread::spawn(|| {
     |     ^^^^^^^^^^^^^ `*mut _` cannot be shared between threads safely
wink@3900x 22-05-18T00:16:44.332Z:~/prgs/rust/clones/dtolnay-trybuild (master)

So my suggestion is that in the future we don't manually add the $DIR/ and instead use the generated files as is. If you'd like I can create another PR with the those unmodified files.

First off, thanks for writing this excellent blog post on structuring and testing procedural macros! I really like the way it's broken down in to parsing, analysis, intermediate representation, and code generation. I am currently attempting to implement a procedural macro of my own following this structure but I've run into an issue that is probably more specific to proc-macro-error than it is your guide, but I thought it would be worth raising as an issue here anyway before I start trying to dig into that codebase.

The issue is with writing test cases to validate what happens when the parsing stage encounters invalid syntax. I imagine that tests for the analysis stage could be similarly affected, but I haven't made it that far yet in my own macro implementation :upside_down_face:. (I also wasn't intentionally trying to test invalid syntax, I just haven't made it far enough in my parse implementation for the syntax i'm targeting to be properly parsed.)

So the problem is that the parse function calls proc-macro-error macros like abort and abort_call_site which results in a panic in tests that validate error cases since according to its docs:

This attribute MUST be present on the top level of your macro (the function annotated with any of #[proc_macro], #[proc_macro_derive], #[proc_macro_attribute]).

To illustrate with an example, here is a contrived test that I created in a clone of this repo:

zsh/3 3540 [148]  (git)-[main]-% git diff
diff --git a/src/parse.rs b/src/parse.rs
index 5572a87..709b66b 100644
--- a/src/parse.rs
+++ b/src/parse.rs
@@ -51,4 +51,18 @@ mod tests {
             ),
         );
     }
+
+    #[test]
+    fn invalid_syntax() {
+        parse(
+            quote!(ident),
+            quote!(
+                #[inline]
+                #[precondition(x <>> == 0)]
+                fn even_to_odd(x: u32) -> u32 {
+                    x + 1
+                }
+            ),
+        );
+    }
 }

(basically just the existing parse::valid_syntax test with a simple ident argument added in the first positional passed to parse)

The docs for proc_macro_error indicate an argument that can be passed to the called allow_not_macro (eg #[proc_macro_error(allow_not_macro)], but that comes with the caveat:

Pay attention: the function this attribute is applied to must return proc_macro::TokenStream.

This is problematic since the parse example in your tutorial relies on proc_macro2::TokenStream.

Anyway, like I mentioned earlier I don't think this is necessarily a problem with this tutorial but thought it would be useful to mention in case any other bright-eyed wannabe Rustaceans find themselves dealing with a similar situation. I'll probably dig into the proc-macro-error crate after submitting this issue to see if it's possible for #[proc_macro_error(allow_not_macro)] to accept proc_macro2::TokenStream as well as proc_macro::TokenStream.

Thanks again for the great tutorial!