Update extract-tests.py to use same test directives as rustdoc.

[wting]

Refactored the file quite a bit, I can add unit tests if desired. There's a few changes from the previous version's behavior:

• destination directory will be created if it doesn't exist
• strings and file is written as unicode

I have a few questions, but will ask them in #11362.

[wting]

Updated following brson's comments in #11362.

{ .should_fail } results in //should-fail at the top of the resulting .rs file. { .xfail-test } results in a similar //xfail-test.

[adrientetar]

There is one thing that would be cool (additionally): checking if there are crate-level compiler attributes at the top of the file #[foo()]; (w/ the semicolon) and if so only append extern mod extra; after that attribute.
Right now, if we want to add a crate-level attr to a code example we have to add extern mod extra; directive first so that they're not invalidated.

For example, this is not valid:

extern mod extra;
#[allow(unused_imports)];

I had to add a trailing extern mod extra; in the Conditions Guide to workaround this, and it is hacky (also annoying when you don't know about it).

So that makes me wondering: why not just drop the extern mod extra business and add it manually in the relevant examples? (cc @brson)

[wting]

It looks like compiler attributes can be multiple lines. Should I worry about this case?

[wting]

Updated with support for single line compiler directives. More specifically matches the regex #[(.*)];.

Also removed unnecessary extern mod extra; lines from guide-conditions.md.

[brson]

Don't the docs also use 'notrust'?

[brson]

Thanks. This is awesome work. I'd like to request just a little bit more here before merging. The ultimate intent here is to bring all our testing tools inline to the same terminology. To that end can you remove extract-test.py's support for 'xfail-test' completely, and instead of having it ignore the 'ignore' tag make it turn 'ignore' into '//xfail-test' when emitting the tests? (I know this sounds completely counterintuitive on the surface but in most of our testing tools 'ignore' means 'this is a test that is broken' - the next step is to probably change compiletest to also use 'ignore' and not 'xfail-test').

Additional next steps (that you don't have to do now): modify compiletest to understand 'should-fail'. Go through all the currently disabled doc tests convert those that appropriate to should_fail.

[adrientetar]

This is really cool work @wting, thanks!

[wting]

• add notrust to ignored tags
• remove xfail-test tag support
• ignore tag now generates //xfail-test

[brson]

I'll test this locally tomorrow. I suspect that removing xfail-test parsing will cause all the current snippets marked xfail-test to start failing; they should now be ignore.

[brson]

Indeed, the current .xfail-test tags in the doc/ folder need to be converted to .ignore so that they continue to be ignored. @wting can you do so?

[wting]

Yeah I'll update it by Monday.

[wting]

All .xfail-test tags in doc/*.md and doc/*.md.po files have been converted to .ignore tag.

[wting]

@brson: bump for review.

[alexcrichton]

This appears to need a rebase again, sorry about that? Feel free to ping me on a rebase and I'll r+ this

[wting]

@alexcrichton: Rebased.

[adrientetar]

@wting Have you looked at the log?
It looks like you forgot some .xfail-test → .ignore in some spots in guide-ffi.md.

(I would grep again through source to see if there aren't other places where it got missed.)

[wting]

Thanks for the catch @adridu59. I think I missed some of the xfail-test after rebasing but I've updated the patch accordingly.

[wting]

Earlier I did a fetch and rebase on mozilla/rust, thought it was a legitimate submodule bump. Anyway I've gone ahead and reverted the llvm submodule pointer back to 535989a9.

[alexcrichton]

Could you squash the two commits together?

[wting]

Squashed!