compile-fail doctest sometimes erroneously claims code is compilable when it isn't

[zackmdavis]

(example found during hour n of chasing down doctest failures for error codes long forgotten)

$ cat repro.rs 
#[inline()]
pub fn something() {}

fn main() {}
$ rustc repro.rs 
error[E0534]: expected one argument
 --> repro.rs:1:1
  |
1 | #[inline()]
  | ^^^^^^^^^^^

error: aborting due to previous error

$ cat repro.md 
```compile_fail
this should fail to compile, because this text is not valid Rust code
```

```compile_fail
#[inline()]
pub fn something() {}

fn main() {}
```
$ rustdoc --test repro.md 

running 2 tests
test repro.md - (line 1) ... ok
test repro.md - (line 6) ... FAILED

failures:

---- repro.md - (line 6) stdout ----
    	thread 'rustc' panicked at 'test compiled while it wasn't supposed to', /checkout/src/librustdoc/test.rs:263:12
note: Run with `RUST_BACKTRACE=1` for a backtrace.


failures:
    repro.md - (line 6)

test result: FAILED. 1 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out

The doctest output claims that the code in a compile_fail block "compiled while it wasn't supposed to," but we saw just a moment before that that code does not, in fact, compile.

$ rustc --version
rustc 1.21.0-nightly (ed16b0a1d 2017-08-05)

[kennytm]

The problem is that the content of an inline attribute is checked during trans, the step which compile_fail will skip.

Duplicate of #40792.

[Mark-Simulacrum]

Closing in favor of #40792.