[cfe-commits] r58866 - in /cfe/trunk: include/clang/AST/DeclBase.h include/clang/Basic/DiagnosticKinds.def lib/Parse/ParseDecl.cpp lib/Sema/SemaDecl.cpp test/SemaCXX/destructor.cpp

[dgregor at apple.com]

On Thu, November 13, 2008 1:50 am, Argiris Kirtzidis wrote:
> Chris Lattner wrote:
>> On Nov 11, 2008, at 12:24 PM, Argiris Kirtzidis wrote:
>>>> One possible fix would be to give the the Declarator a dummy name
>>>> (e.g., "<destructor>", "<conversion function">) or make use of the
>>>> flags that say that this isn't a normal named declarator, then make
>>>> NamedDecl::getName() virtual and override it for destructors,
>>>> conversion operators, etc., to form the name on-demand. Since
>>>> getName() isn't invoked unless we're emitting a
>>>> diagnostic---definitely not in the hot path---it's probably a good
>>>> tradeoff. We'll have to be a little careful throughough, since
>>>> getIdentifier() and getName() will have different strings associated
>>>> with them.
>>>>
>>>
>>> Another benefit of using different IdentifierInfos for "~C" and "C"
>>> is that the unique IdentifierInfo for the destructor allows it to be
>>> cached by the IdentifierResolver.
>>
>> I don't think that follows.  There is nothing that would prevent using
>> something like our "Selector" class (which handles similar [in spirit]
>> things for ObjC).  Selectors wrap identifier infos, are uniqued and
>> are pointer sized by-value objects.
>
> It's not only about uniqueness, IdentifierInfo also has a place to store
> a Decl (FETokenInfo).
>
> A class that wraps IdentifierInfo will work but it needs to be more
> complex than a simple "smart pointer"; 3 bits won't be enough for the
> 'operator' stuff and it will also have to be able to store a Decl.

I suggest that we pause this conversion for a day or so. I'm working on
Chris's suggestion of a Selector-like class for declaration names. Once
I've made it work the way I think it should, I'll post the patch on
cfe-dev with some explanation and we can pick up the discussion there.

  - Doug