[cfe-commits] r61746 - in /cfe/trunk: include/clang/AST/ include/clang/Basic/ include/clang/Parse/ lib/AST/ lib/Parse/ lib/Sema/ test/SemaCXX/

[Douglas Gregor]

On Jan 5, 2009, at 8:49 PM, Chris Lattner wrote:

> On Jan 5, 2009, at 8:10 PM, Douglas Gregor wrote:
>>> Doing this comparison with types seems awkward, can you just do
>>> something like:
>>>
>>> if (const RecordType *RT = ArgType->getAsRecordType())
>>> if (RT->getDecl() == this) ...
>>>
>>> instead?  likewise in hasConstCopyAssignment.
>>
>> That won't work, since there can be multiple RecordDecls for a  
>> given record type.
>
> Is there a similar mechanism that will work?

Use Context.getTypeDeclType to get at the RecordType corresponding to  
the RecordDecl, then compare the canonical types :)

We don't currently have an easier way to map from a RecordDecl to the  
"canonical" RecordDecl known to the RecordType.

>
>>>> +  if (getLangOptions().CPlusPlus) {
>>>>  CheckExtraCXXDefaultArguments(D);
>>>> +    if (!T->isPODType())
>>>> +      cast<CXXRecordDecl>(Record)->setPOD(false);
>>>
>>> I see that you basically clear the 'ispod' bit whenever something is
>>> parsed that prevents a class from being a POD.  Would it be better  
>>> to
>>> make it a predicate that walks the member list to determine podness?
>>> The advantage of doing this is that it would put all the "podness"
>>> checks together into one method instead of scattered throughout  
>>> sema.
>>> What do you think?
>>
>> Most declarations of a class type will need to perform the check  
>> for a POD type, which is recursive (it's only a POD if all of its  
>> non-static data members are PODs, and so on), so it might be  
>> expensive to compute this property on the fly. Personally, I prefer  
>> the way Sebastian's doing it: there will be quite a few other class  
>> properties like this, too.
>
> What if the query is lazy but cached?  Many types are defined in  
> headers that are never used, I can't believe that we'd end up  
> querying isPod for all of them...

We could certainly make the query lazy but cached, but why? The most  
expensive check needed to maintain this flag is "if (!T->isPODType())".

>
>>> Would it be possible to synthesize these lazily
>>> like G++ does?
>>
>>
>> Yes, but let's implement the semantics of these operations first,  
>> so that we have known, working regression tests when we go and  
>> optimize away these implicitly-declared special member functions.
>
> Ok, that is fair enough.  Would it make sense to add a README file  
> or something for notes on C++ optimization opportunities?


That's a good idea.

	- Doug