patch: clarify missing template arguments when parsing base specifier

[Nick Lewycky]

On 27 August 2013 00:04, Chandler Carruth <chandlerc at google.com> wrote:

>
> On Mon, Aug 26, 2013 at 11:40 PM, Nick Lewycky <nlewycky at google.com>wrote:
>
>> On 26 August 2013 23:22, Chandler Carruth <chandlerc at google.com> wrote:
>>
>>>
>>> On Mon, Aug 26, 2013 at 11:09 PM, Nick Lewycky <nlewycky at google.com>wrote:
>>>
>>>> On 26 August 2013 22:53, Chandler Carruth <chandlerc at google.com> wrote:
>>>>
>>>>> I think this case has more problems than just verbosity...
>>>>>
>>>>> On Mon, Aug 26, 2013 at 10:38 PM, Nick Lewycky <nlewycky at google.com>wrote:
>>>>>
>>>>>> a.cc:1:56: error: no template named 'Foox'; did you mean 'Foo'?
>>>>>>  template <typename T> class Foo {}; class Bar : public Foox {};
>>>>>> [point at 'Foox' suggest 'Foo']
>>>>>>
>>>>>
>>>>> Why does it assume Foox is a template?
>>>>>
>>>>
>>>> It's already proven that it's not not-a-template.
>>>>
>>>>  a.cc:1:29: note: 'Foo' declared here
>>>>>> template <typename T> class Foo {}; class Bar : public Foox {};
>>>>>> [point at 'Foo']
>>>>>> a.cc:1:56: error: expected template argument list after template-id
>>>>>> template <typename T> class Foo {}; class Bar : public Foox {};
>>>>>> [point at 'Foox']
>>>>>>
>>>>>
>>>>> And given that we then hit this error, why do we even consider the Foo
>>>>> typo correction? Do we prefer that over a "Fooxie" class due to shorter
>>>>> edit distance? That doesn't seem right. I would intuitively expect the lack
>>>>> of "<..." to be a stronger signal than any edit distance, and thus
>>>>> disqualify template-ids from the typo correction candidate set.
>>>>>
>>>>
>>>> No. We only go down this patch after we've done a lookup and typo
>>>> correction on non-templates, and found nothing.
>>>>
>>>
>>> I'm suggesting that a missing header or exceeding the maximum edit
>>> distance threshold seems just as plausible as using a template without
>>> template arguments. I'm not claiming that I have some strong reason to
>>> believe one interpretation or the other to be more likely, only that it
>>> doesn't seem clear-cut in either direction to me.
>>>
>>
>> ... if it can't find a template, then it doesn't mention templates:
>>
>> $ echo 'class Bar : public Foo {};' | llvm/Debug+Asserts/bin/clang -x c++
>> -
>> <stdin>:1:20: error: expected class name
>> class Bar : public Foo {};
>>                    ^
>> 1 error generated.
>>
>> $ echo 'template<typename T> class Fooa; class Foob {}; class Bar :
>> public Foo {};' | llvm/Debug+Asserts/bin/clang -x c++ -
>> <stdin>:1:68: error: unknown class name 'Foo'; did you mean 'Foob'?
>> template<typename T> class Fooa; class Foob {}; class Bar : public Foo {};
>>                                                                    ^~~
>>                                                                    Foob
>> <stdin>:1:40: note: 'Foob' declared here
>> template<typename T> class Fooa; class Foob {}; class Bar : public Foo {};
>>                                        ^
>> 1 error generated.
>>
>> Today, clang emits the exact same diagnostic "expected class name" even
>> when you do have Foo declared as a template. That's the only thing I'm
>> trying to fix, but it has this weird side-effect in that even asking Sema
>> isTemplateName() causes does typo-correction, and when that typo-correction
>> succeeds it issues this "no template" diagnostic.
>>
>
> I really do understand that, and I don't think your patch is wrong. I just
> think the problem in typo correction it exposes is severe enough to warrant
> fixing.
>

Ah, okay! Then I'm glad I included it in the email. I was worried that you
saw my example of what I thought was a bad diagnostic, and interpreted it
as the goal of my patch.

Specifically, I don't think we should be typo-correcting to a template-id
> when we have syntactic indicators that a template-id isn't valid. For
> example, given "foo->baz", we should only correct to a method name "bar" if
> we can also correct to "bar()" because it has a zero-argument overload
> (modulo weird cases where the expression "foo->bar" is valid). There are
> likely other places (maybe even with methods!) where we get this wrong, but
> I think all of them are pretty dubious. Here, I think the lack of "<" is a
> *really* strong signal as there is just no way to use a template-id in that
> context. If all template parameters had default arguments, perhaps 'Foo<>'
> would be a good correction, but that's a more clever step.
>

Good catch. Foo could be a template function (or a template variable in
C++14 I suppose) instead of a template class. That I know how to fix.

Unfortunately, the template typo correction is wired into
LookupTemplateName. Even if I bypass isTemplateName and create a
LookupResult with diagnostics suppressed, we still get the typo
corrections. I do not know the design of clang well enough to know how to
change that.

Nick
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130827/80e7b688/attachment.html>