r194002 - Try to correct a mistyped "-" or ">" to "->" for some C++ cases.

[Kaelyn Uhrain]

On Mon, Nov 11, 2013 at 4:13 PM, Richard Smith <richard at metafoo.co.uk>wrote:

> On Thu, Nov 7, 2013 at 2:15 PM, Kaelyn Uhrain <rikka at google.com> wrote:
>
>> Richard,
>>
>> Thanks for the review feedback!
>>
>> Here's an alternate version that better handles cases (such as the
>> example you gave) where the code is actually valid without the "-" or ">"
>> being changed to "->", by only trying the correction if an error diagnostic
>> would have been emitted. It is similar to what I had mentioned doing,
>> except that if the correction fails, ParseRHSOfBinaryExpr is called a
>> second time to emit the original diagnostic messages. I ended up doing it
>> this way because I encountered a problem with my previous idea: how to get
>> the token stream in the right state if the correction fails and the
>> original errors should be emitted. Preprocessor::EnableBacktrackAtThisPos()
>> and friends work for resetting the token stream after the call to
>> ParseRHSOfBinaryExpr but don't quite work for setting the token stream back
>> to the state after the (first) call to ParseRHSOfBinaryExpr after trying
>> the correction and having it fail (emitting the stored diagnostics from the
>> call to ParseRHSOfBinaryExpr to avoid having to call it a second time only
>> works if the token stream can also be restored to the correct state).
>>
>
> This looks a lot cleaner, thanks. I have a handful of somewhat minor
> concerns:
>
>
> DelayingDiagnosticConsumer doesn't appropriately handle
> DiagnosticConsumer::IncludeInDiagnosticCounts. "clang-check -fixit" is
> supposed to return 0 if it fixes all errors; it looks like it won't if it
> fixes this error, for this reason. This is a bit tricky: you don't know
> whether to include the diagnostic in diagnostic counts until you know
> whether you're emitting it, and you don't know that until you know whether
> you can recover.
>

In which circumstances is the diagnostic count not being handled correctly?
DelayingDiagnosticConsumer simply passes all the diagnostics it captured on
to a regular DiagnosticConsumer in the event those diagnostics should be
emitted; unless I'm missing something it is just acting as a buffer/queue
in front of another DiagnosticConsumer, and the latter should be handling
the diagnostic counts. (But this and all of the other points you brought up
are why I wanted and appreciate the code review! All this mucking around
with the parser and diagnostics is a bit outside familiar territory.)


>
> You're suppressing diagnostics that come from outside the immediate
> context of the error (for instance, if the RHS triggers a template
> instantiation, you shouldn't suppress diagnostics from there, because they
> won't be emitted if the code later uses the same template specialization
> again). That means that you can get a --fixit run that fixes all errors,
> and yet rebuilding the code still produces errors; that's not really
> supposed to happen.
>

Yeah, the case of there being errors in the initial code path and the
correction causing errors and sticking with the original set of errors
leaves a lot to be desired. I'd prefer to be able to use a second
DelayingDiagnosticConsumer to catch the errors from trying the correction
and on failure just emit the original diagnostics and set the parser to the
correct state, but couldn't figure out a way to achieve that with the
current tentative parsing/backtracking facilities. It would still leave the
issue of cleaning up template instantiations triggered by whichever path
wasn't taken that weren't triggered by the path that was taken. And in
those cases, the (suppressed) diagnostics from the instantiations would be
bogus anyway as the resulting state is that the instantiations shouldn't
have ever occurred.


> I'm also a little concerned about the cost of turning on tentative
> parsing/backtracking here. Enabling tentative parsing (allegedly!) has a
> nontrivial cost, but I suppose we won't see pointers to class types on the
> LHS of '-' or '>' very frequently.
>

Yeah, both this version and my original patch were based on the assumption
that pointers to classes wouldn't be common on the LHS of '-' or '>' and so
incurring a bit of overhead to be able to catch and diagnose a mistyped
"->" wouldn't be a huge compile-time issue.


>
>
> I don't really have good suggestions for the above -- this is a really
> tricky issue to recover from, and we're not really set up for tentatively
> performing arbitrary parsing actions.
>
>
> Tiny things:
>  * You should use getAs<RecordType>, not getAsStructureType() -- this
> should work for classes and unions, too.
>

Not sure about unions, but getAsStructureType() seems to work just fine for
classes. I will change it though... I always err toward using
getAsStructureType instead of getAs<RecordType> whenever I look at the API
docs as they don't indicate the former is deprecated and I always feel like
it may have some extra logic that is needed but not provided by getAs<>. (I
also tend forget that getAs is a template function as the doxygen docs
render such things poorly; e.g. in
http://clang.llvm.org/doxygen/classclang_1_1Type.html the
"template<typename T >" is grey instead of black and in the far left of the
column with return types separated by a large amount of whitespace from the
"const T* getAs()" and actually on the line above the return type... and
the 1-line blurb for getAs in the "Public Member Functions" table is rather
misleading: "This will check for a TypedefType by removing any existing
sugar until it reaches a TypedefType or a non-sugared type."... but I
digress.)

 * Use TentativeParsingAction rather than using PP.*Backtrack* directly --
> this will fix a minor bug where you get the source location of the '->'
> token wrong.
>

Ah ok, I'll look into that. I found the PP.*Backtrack* functions long
before TentativeParsingAction when trying to figure out how to get the
parser or lexer to backtrack.

 * Can you add an RAII object for registering and unregistering the
> DelayingDiagnosticConsumer?
>

Sure thing.

Thanks again for reviewing these patches!

Cheers,
Kaelyn
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20131111/8de95be4/attachment.html>