[cfe-commits] [PATCH] Allow typedef with unnecessary "typename" when ms-extensions are enabled.

[Will Wilson]

>>> > You should consume the 'typename' token, and try to recover as if the
>>> > 'typename' keyword were not present, following the logic present later
>>> > on in
>>> > that function. As a special case, if that later logic were to find that
>>> > the
>>> > tokens after the 'typename' keyword can't be annotated as a type name,
>>> > you
>>> > should produce a hard error (even with MS extensions enabled).
>>> >
>>> > (In particular, for 'typename identifier', it will be the identifier you
>>> > annotate, not the 'typename' keyword.)
>>>
>>> Sorry for the delay! I've attached an updated patch (with test cases)
>>> that seems to do the job. I've also had to update two other tests to
>>> allow for the different code paths (and thus different errors)
>>> followed due to the attempt at recovering from the error.
>>>
>>> All tests pass locally. It'd be great if some other people could
>>> verify it as well.
>>
>>
>> A few things: This is still only producing a warning if 'typename' is
>> followed by something which isn't a type name in MS mode. That needs to be
>> fixed to produce an error in that case. Also, it would be great to apply
>> typo correction in this case, and to avoid duplicating the code from the
>> second half of the function.
>
> Thanks for giving it the once over.
>
> Fair point on the error in MS mode, I'll rework it.
>
> I agree the code duplication in TryAnnotateTypeOrScopeToken is less
> than optimal (I'll try the recursive approach you've suggested below).
> The typo-correction sounds promising - if I get a change though I'll
> investigate further (with the sound of deadlines rushing by ;).
>
>> How about just recursively calling TryAnnotateTypeOrScopeToken after
>> consuming the invald 'typename' token, then emitting the error diagnostic if
>> it fails or we're not in MS mode, and the warning diagnostic otherwise?
>
> That actually looks like a promising approach... I'll give it a shot
> tomorrow if time allows.

Time did allow. The patch is attached using the recursion approach and
(on the whole) looks a great deal better. Only
Parser::ParseCastExpression() needed updating to ensure it didn't
simple assume it got an annotated token back from
TryAnnotateTypeOrScopeToken(). All tests pass locally.

Cheers,
Will.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: unecessary_typename_msext.diff
Type: application/octet-stream
Size: 3837 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20120508/9f82958a/attachment.obj>