[cfe-commits] r60993 - in /cfe/trunk: include/clang/AST/DeclObjC.h lib/AST/DeclObjC.cpp lib/Sema/SemaDecl.cpp lib/Sema/SemaDeclObjC.cpp

[Fariborz Jahanian]

On Dec 17, 2008, at 10:49 AM, steve naroff wrote:

>
> On Dec 17, 2008, at 12:08 PM, Fariborz Jahanian wrote:
>
>>
>> On Dec 17, 2008, at 6:34 AM, steve naroff wrote:
>>
>>>
>>> On Dec 15, 2008, at 1:02 PM, Douglas Gregor wrote:
>>>>>>
>>>>>> Will we need to have the RecordDecl layout for most or all
>>>>>> Objective-C interfaces? If most programs only need the layout  
>>>>>> for a
>>>>>> small number of interfaces, it might be beneficial to compute the
>>>>>> layouts lazily (as we do with the layout of RecordDecls).
>>>>> Lay out is needed when a field is referenced. I will look at this
>>>>> later when I am done with code gen.
>>>>
>>>> Okay!
>>>>
>>>
>>> Just to build on this issue...why does ObjCInterfaceDecl need a  
>>> 'RecordForDecl' slot at all?
>>>
>>> From my perspective, this should be computed lazily (as you  
>>> suggest) and stored separately (using a map that is populated  
>>> during code generation).
>>
>> Storage lay out is done lazily. Build of the record decl can also  
>> be done lazily. Saving a slot ONTOH, may not be worth the extra  
>> complexity, but certainly
>> can be done as well.
>>
>
> Hey Fariborz,
>
> I don't think that storing the slot separately adds a lot of  
> complexity. Here are two arguments for storing RecordForDecl  
> separately:
>
> #1: It is more modular. For example, I just fixed a regression that  
> should have been isolated to code generation. Unfortunately, the  
> regression showed up with -fsyntax-only (because the slot was being  
> mismanaged within the AST).
>
> #2: It scales better. Even though this is just 4 bytes, imagine if  
> every clang client added overhead to the base AST's! Here is an  
> example from the Rewriter...
>
>     llvm::DenseMap<ObjCMethodDecl*, std::string> MethodInternalNames;
>
> I could have added a 'MethodInternalNames' slot, however decided the  
> map was a much better solution.
>
> Summary: We need to avoid dumping stuff into our AST's. The AST's  
> should model the language element, not model application specific  
> operations on the language element.
>
> Agreed?

Yes. I was planing on doing this as part of the lazily building the  
record. I put the  FIXME right on top of the slot declaration for the  
record.

- Fariborz

>
> snaroff
>
>> - Fariborz
>>
>>>
>>>
>>> snaroff
>>>
>>>> 	- Doug
>>>> _______________________________________________
>>>> cfe-commits mailing list
>>>> cfe-commits at cs.uiuc.edu
>>>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>>>
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20081217/68d4be57/attachment.html>