[PATCH] Fix return type deduction for member templates

[Richard Smith]

On Thu, Jun 13, 2013 at 3:40 PM, Faisalv <faisalv at gmail.com> wrote:
>
>
> On Jun 13, 2013, at 4:30 PM, Richard Smith <richard at metafoo.co.uk> wrote:
>
>> Hi Faisal,
>>
>> On Thu, Jun 13, 2013 at 11:30 AM, Faisal Vali <faisalv at gmail.com> wrote:
>>> While implementing return type deduction for generic lambdas, I stumbled upon
>>> a bug which led to clang mishandling the following code:
>>>
>>> struct Lambda {
>>>  template<class T> auto operator()(T t) {
>>>    return 5;
>>>  }
>>> template<class T> auto foo(T t) { return t; }
>>> };
>>> void test() {
>>>    Lambda L;
>>>    int i = L(3);
>>> }
>>>
>>>
>>> The issue (as i understand it) was that in some contexts only the
>>> function template (and not the specialization) would get passed to
>>> DiagnoseUseOfDecl (which deduces the return type of the function).  So
>>> the specialization would flow through to Codegen with auto as its
>>> return type.
>>>
>>> So, with that in mind, this patch implements a trivial fix.
>>>
>>> What do you think Richard? Is there a better way to address this bug?
>>
>> I was aware of this issue, but got snowed under with other things.
>> This bug is not limited to 'auto' type deduction; the other code in
>> DiagnoseUseOfDecl suffers the same way. For instance:
>>
>> struct Lambda {
>>  template<class T> static __attribute__((unused)) int foo(T) {
>>    return 5;
>>  }
>> };
>> int bar() {
>>    Lambda L;
>>    return L.foo(3);
>> }
>>
>> ... does not warn that the function is marked 'unused' but is used,
>> but if you call it as Lambda::foo(3), it does warn.
>>
>> I think the right fix here is to pass both the declaration found by
>> name lookup and the declaration that is actually used into
>> DiagnoseUseOfDecl.
>>
>
> What if i just call diagnoseuseofdecl again on the used specialization instead of adding a parameter to it?

If this is the only call which has the problem, then that would be OK.
I suspect there are other cases, though, and there are probably cases
which have the opposite bug (passing only the used decl and not the
found decl). As a quick fix, making two calls would work, but it's
probably not what we want in the longer term.