Regression in error message when a custom derive panics

[dtolnay]

Between nightly-2018-01-26 and nightly-2018-01-27 the error message for a custom derive panic got worse. This is the range 9fd7da9...bacb5c5. Mentioning @Zoxc because #47634 sounds relevant.

src/main.rs

#[macro_use]
extern crate repro;

#[derive(Panic)]
struct S;

fn main() {}

src/lib.rs

extern crate proc_macro;
use proc_macro::TokenStream;

#[proc_macro_derive(Panic)]
pub fn f(_input: TokenStream) -> TokenStream {
    panic!("no can do");
}

Before:

error: proc-macro derive panicked
 --> src/main.rs:4:10
  |
4 | #[derive(Panic)]
  |          ^^^^^
  |
  = help: message: no can do

After:

thread 'rustc' panicked at 'no can do', src/lib.rs:57:5
note: Run with `RUST_BACKTRACE=1` for a backtrace.
error: proc-macro derive panicked
 --> src/main.rs:4:10
  |
4 | #[derive(Panic)]
  |          ^^^^^
  |
  = help: message: no can do

[Zoxc]

It is indeed caused by #47634 where rustc no longer swallows stderr. This does seem desirable to me however and helps track down the bug in the plugin. How would you get panic backtraces on stable?

[dtolnay]

It is not a bug in the plugin -- panicking is the intended way for custom derives to report errors to the user.

#[macro_use]
extern crate serde_derive;

#[derive(Serialize)]
struct S {
    #[serde(broken)]
    f: u8,
}

  |
4 | #[derive(Serialize)]
  |          ^^^^^^^^^
  |
  = help: message: unknown serde field attribute `broken`

Showing the line number within the custom derive at which the error was detected is unnecessary and unhelpful.

[Zoxc]

I see that the RFC uses panics as a temporary way to report errors. We do still want to report the backtrace if desired though.

cc @rust-lang/compiler

[nikomatsakis]

panicking is the intended way for custom derives to report errors to the user.

Longer term, this does seem like a suboptimal way to report errors, since it doesn't allow for more than one error at a time very gracefully, right? Still, if that's the expected way, we have to support it.

We do still want to report the backtrace if desired though.

At the limit, we can do a -Z option... but in any case can't we catch the panic or something to prevent it from being printed to stderr?

cc @rust-lang/libs -- how configurable is the "print to stderr on panic" code here?

[Zoxc]

We can install a panic handler in rustc which does nothing when we are in compiler plugins.

[nikomatsakis]

@Zoxc

We can install a panic handler in rustc which does nothing when we are in compiler plugins.

Seems like a decent idea. Would this also suppress RUST_BACKTRACE?

[Zoxc]

Would this also suppress RUST_BACKTRACE?

It would suppress the note: Run with `RUST_BACKTRACE=1` for a backtrace. line, if that is what you meant.

[alexcrichton]

@nikomatsakis yes set_hook can be used to customize what happens when a panic happens.

[nikomatsakis]

@Zoxc I meant, if one does set RUST_BACKTRACE=1, would you still get a backtrace from inside the plugin. I thought that was what you wanted to enable?

[Zoxc]

@nikomatsakis We could do that, although it is likely users will have RUST_BACKTRACE set when invoking the compiler because they want backtraces of the executables it builds and aren't interested in a backtrace for a compiler plugin error.

[nikomatsakis]

@Zoxc I suppose that could happen with cargo run -- but I am a bit confused. Didn't you write this (emphasis mine):

It is indeed caused by #47634 where rustc no longer swallows stderr. This does seem desirable to me however and helps track down the bug in the plugin. How would you get panic backtraces on stable?

Isn't this arguing that we should show some sort of backtrace from inside the plugin?