RFC: Change C++ ABI so that interface dispatch does not depend on reflection.

[David Daney]

Tom Tromey wrote:
>>>>>>"David" == David Daney <ddaney@avtrex.com> writes:
> 
> 
> David> The first case is simple to handle.  If a method is the <clinit>
> David> method just emit the reflection data.
> 
> Cute.
> 
> You could also just have a pointer to this function in the Class
> object.  That way, for a reflection-less install, you could even
> remove the reflection-related fields from Class.  This would require
> hacks to natClass.cc and probably other places, though.
> 

Since the String table (for the <clinit> name) is shared the only real 
overhead is a single entry in the method reflection table (12 bytes on a 
32 bit architecture)  If the class object had a pointer this would be 4 
bytes.  So that would be only 8 bytes of saving, but the class 
initialization time would be faster as you would not have to do a lookup 
in the method table.


> David> Build interface dispatch tables at compile/link time.
> David> Change runtime to use new tables instead of building its own at class
> David> init time.
> 
> Sounds reasonable to me.  I don't think you even need to change the
> runtime very much... it looks to me like the runtime linker already
> skips this step if the 'idt' field is set.
> 
> Building the idt table at compile time, for the C++ ABI, seems like a
> general improvement.  I don't think we need this to be conditional on
> -fno-reflection.

As I said in the original message: For an object like libgcj where 
usually a majority of the classes would remain uninitialized, there is a 
size penalty that *could* adversely effect cache performance.

I don't know if there is a win to reduce class initialization overhead 
at the expense of a larger class data size.

David Daney