[cfe-dev] AST's VTableContext object and the Microsoft C++ ABI

[John McCall]

On Jun 17, 2012, at 2:08 PM, Charles Davis wrote:
> I'm working on improving support for v-tables under the Microsoft C++ ABI, and while poking around in IRGen, I discovered the VTableContext class that lives in the AST library.
> 
> This class is responsible for generally laying out the (single, unless the class needs a construction v-table) v-table for each class that needs one. It is very specific to the Itanium ABI concept of a virtual table, which means it can't be used as-is for the Microsoft ABI. Trouble is, it's used by various pieces of IRGen--which is how I discovered it.
> 
> So, I'm asking for your advice. What should I do about this?
> 
> - Should I make certain methods of VTableContext virtual, and override them for the Microsoft ABI?
> - Should I hide this class behind the CXXABI objects? If I go this route, I'll have to indirect the pieces of IRGen currently using this object through the CGCXXABI object.

This is right way to go.  Having a single implementation with some special cases works for class layout because the *questions* that we want to ask of a class layout remain largely the same across ABIs:  for example, "what is the offset of this field?" or "what is the offset of this base class object?" or "what is the size of this type, laid out as a complete object?"  That's just not true of v-tables, due mostly to the distinction between vf-tables and vb-tables, but also because entire significant concepts (e.g., a VTT or a construction vtable) are inapplicable.  So it makes sense that generic IR generation should defer these things to the ABI object, which knows the right questions to ask.

John.