Type inference regression

[SimonSapin]

Servo compiles with rustc 71c06a5 2016-12-18, but not with 439c312 2016-12-21. Diff: 71c06a5...439c312

The style crate fails to build with:

error[E0282]: unable to infer enough type information about `E`
  --> /home/simon/projects/servo/components/style/font_face.rs:65:9
   |
65 |         try!(self.family.to_css(dest));
   |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ cannot infer type for `E`
   |
   = note: type annotations or generic parameter binding required
   = note: this error originates in a macro outside of the current crate

The to_css method being called here is:

impl ToCss for FontFamily {
    fn to_css<W>(&self, dest: &mut W) -> fmt::Result where W: fmt::Write {
        // ...
    }
}

Its return type is std::fmt::Result which is type Result = std::result::Result<(), std::fmt::Error>;. This call is in another impl of the same method of the same trait:

impl ToCss for FontFaceRule {
    fn to_css<W>(&self, dest: &mut W) -> fmt::Result where W: fmt::Write {
        try!(dest.write_str("@font-face { font-family: "));
        try!(self.family.to_css(dest));
        try!(dest.write_str(";"));

        // ...
    }
}

Adding an explicit type annotation works around this error, but only to uncover another very similar one, then another. The third one was in code generated by serde’s derive(Serialize), so I gave up that approach.

        let r: fmt::Result = self.family.to_css(dest);
        try!(r);

So far I haven’t managed to reduce a test case smaller than "most of Servo", sorry.

I’m suspicious of 6ea1fbb (which is part of #38099) because it’s commit message mentions inference. I’m not building a compiler #38099 to confirm.

[SimonSapin]

CC @GuillaumeGomez, @nikomatsakis

[SimonSapin]

Also fails in 92d4600, so #38099 is not to blame. New diff: 71c06a5...92d4600

I don’t have another guess, and doing a full bisect would take a very long time…

[eddyb]

I can't find the cause in that range, at least nothing changed in the compiler AFAICT.
Can you try 3038f30 (the commit before @alexcrichton's larger rollup)?

[SimonSapin]

Expanding the try! macro:

diff --git a/components/style/font_face.rs b/components/style/font_face.rs
index 70f904c17a..a3d0ca3882 100644
--- a/components/style/font_face.rs
+++ b/components/style/font_face.rs
@@ -62,7 +62,10 @@ impl ToCss for FontFaceRule {
     // Serialization of FontFaceRule is not specced.
     fn to_css<W>(&self, dest: &mut W) -> fmt::Result where W: fmt::Write {
         try!(dest.write_str("@font-face { font-family: "));
-        try!(self.family.to_css(dest));
+        match self.family.to_css(dest) {
+            Err(err) => return Err(err),
+            Ok(_) => {}
+        };
         try!(dest.write_str(";"));
 
         if self.sources.len() > 0 {

(Even simplifying it, this is not using From) gives:

error[E0282]: unable to infer enough type information about `T`
  --> /home/simon/projects/servo2/components/style/font_face.rs:66:13
   |
66 |             Err(err) => return Err(err),
   |             ^^^^^^^^ cannot infer type for `T`
   |
   = note: type annotations or generic parameter binding required

error: aborting due to previous error

[eddyb]

@SimonSapin Alright, that matches what I've seen from @nox.
What about a couple more steps of bisecting? Or maybe compiling this crate with the "servo" feature off
(which, AFAICT, turns off all the macros 1.1 - they might be responsible).

[SimonSapin]

Swapping the order of the two match arms:

error[E0282]: unable to infer enough type information about `T`
  --> /home/simon/projects/servo2/components/style/font_face.rs:66:13
   |
66 |             Ok(_) => {}
   |             ^^^^^ cannot infer type for `T`
   |
   = note: type annotations or generic parameter binding required

error: aborting due to previous error

Changing the Ok(_) pattern to Ok(()):

error[E0282]: unable to infer enough type information about `E`
  --> /home/simon/projects/servo2/components/style/font_face.rs:66:13
   |
66 |             Ok(()) => {}
   |             ^^^^^^ cannot infer type for `E`
   |
   = note: type annotations or generic parameter binding required

error: aborting due to previous error

Hacking ./mach build-geckolib to not use the stable version of the compiler (so effectively disabling the servo feature and enabling the gecko one) also gives the same error.

[SimonSapin]

New range, based on @eddyb-guided manual bisect: 58d58c2 good, 551cb06 bad. Diff: 58d58c2...551cb06, which leaves only #38171 in-between. CC @jseyfried, @nrc.

Also found more clues. More context for the code where the error happens:

use computed_values::font_family::FontFamily;

pub struct FontFaceRule {
    pub family: FontFamily,
    // ...
}

impl ToCss for FontFaceRule {
    fn to_css<W>(&self, dest: &mut W) -> fmt::Result where W: fmt::Write {
        // ...
        try!(self.family.to_css(dest));
        // ..
    }
}

computed is a module generated by a macro, that contains pub use re-exports of other modules:

macro_rules! reexport_computed_values {
    ( $( $name: ident )+ ) => {
        pub mod computed_values {
            $(
                pub use properties::longhands::$name::computed_value as $name;
            )+
        }
    }
}

// Invoke reexport_computed_values! with a list of idents generated by Python code
longhand_properties_idents!(reexport_computed_values); 

If I add a similar pub use not in a macro:

pub mod computed_values_no_macro {
    pub use properties::longhands::font_family::computed_value as font_family;
}

or if I change the original code to use the real path without going through a pub use, this seems to work around this issue completely.

… only to uncover another one:

error: internal compiler error: cat_expr Errd
     --> /home/simon/projects/servo/target/debug/build/style-9107aeac634bb1f9/out/properties.rs:75381:40
      |
75381 |                 align_items::center => align_self::center,
      |                                        ^^^^^^^^^^^^^^^^^^

note: the compiler unexpectedly panicked. this is a bug.

note: we would appreciate a bug report: https://github.com/rust-lang/rust/blob/master/CONTRIBUTING.md#bug-reports

note: run with `RUST_BACKTRACE=1` for a backtrace

thread 'rustc' panicked at 'Box<Any>', /home/simon/projects/rust/src/librustc_errors/lib.rs:382
stack backtrace:
   1:     0x7f2b65e50e53 - std::sys::imp::backtrace::tracing::imp::write::h34c2fe4536cff926
   2:     0x7f2b65e5f63d - std::panicking::default_hook::{{closure}}::he19e50a2201d99c4
   3:     0x7f2b65e5f1db - std::panicking::default_hook::h954c9f5a176f6e88
   4:     0x7f2b65e5fac8 - std::panicking::rust_panic_with_hook::h12277e75a2b70114
   5:     0x7f2b5f5f46b7 - std::panicking::begin_panic::hc146ae5a0122e959
   6:     0x7f2b5f609d9a - rustc_errors::Handler::abort_if_errors::h714fd9b9b0557ee7
   7:     0x7f2b645c6f1f - rustc_passes::consts::check_crate::h2923cde18866be7c
   8:     0x7f2b661f4e49 - rustc_driver::driver::phase_3_run_analysis_passes::{{closure}}::h1fd32502995d307a
   9:     0x7f2b661f163e - rustc_driver::driver::phase_3_run_analysis_passes::hfd6790d5cb183751
  10:     0x7f2b661d6091 - rustc_driver::driver::compile_input::h0e69e9ade6aaa4c7
  11:     0x7f2b6621f441 - rustc_driver::run_compiler::h72646ffc571b7660
  12:     0x7f2b6613454b - std::panicking::try::do_call::hbf38fb785c7d412c
  13:     0x7f2b65e68b66 - __rust_maybe_catch_panic
  14:     0x7f2b66155aa1 - <F as alloc::boxed::FnBox<A>>::call_box::h55b6e5fdb729b74d
  15:     0x7f2b65e5e4d0 - std::sys::imp::thread::Thread::new::thread_start::h0578282da79025df
  16:     0x7f2b5ed9e453 - start_thread
  17:     0x7f2b65b257de - __GI___clone
  18:                0x0 - <unknown>

[SimonSapin]

Found another ICE, I don’t know if it’s related: #38556.

[SimonSapin]

Found a smaller test case for the inference issue:

fn main() {}

pub mod a {
    pub struct FontFamily;

    impl FontFamily {
        pub fn to_css(&self) -> ::std::fmt::Result {
            Ok(())
        }
    }
}

macro_rules! reexport {
    () => {
        mod b {
            pub use a;
        }
    }
}

reexport!();

pub mod c {
    use b::a::FontFamily;

    struct FontFace {
        family: FontFamily
    }

    impl FontFace {
        pub fn to_css(&self) -> ::std::fmt::Result {
            try!(self.family.to_css());
            Ok(())
        }
    }
}

[eddyb]

@SimonSapin That "other ICE" is actually a direct observation of the bug, whereas this inference issue is some weird side-effect of the way we handle errors that don't stop compilation.