Re: Basic questions about Gorm

[David Chisnall]

On 9 Apr 2011, at 10:43, Nicola Pero wrote:

>> - Unique protocols (two protocols returned by @protocol()
>> may give different sets of methods - this requires a little
>> bit of help from the compiler, not sure if GCC 4.6 does it)
> 
> Do you have an example/testcase of what you mean by this ?


The runtime only sees protocols that are attached to classes.  If you declare a 
protocol in a compilation unit and then reference it with @protocol(), the 
compiler will generate a reference to it, but the runtime won't see it.  If you 
DON'T declare the protocol, but still reference it with @protocol(), then the 
runtime still won't see it, and the compiler will generate a stub protocol that 
only has the name filled in.

If the runtime sees the protocol, then it can add it to a table and update all 
other references to the protocol to use the same metadata (required / optional 
methods / properties).  The GNUstep runtime does this, and I think the GCC 4.6 
runtime does too.  

To ensure that the runtime sees all protocols, I added a really ugly hack, 
where every protocol that is referenced but not attached to a class or category 
is implicitly added as a category on a class called 
__ObjC_Protocol_Holder_Ugly_Hack.  This ensures that the runtime sees all 
protocols.  The Apple solution to this is to make every protocol a public 
symbol with link-once ODR linkage (in definitions, external linkage in 
references) so you get a linker error if you reference a protocol that is 
declared, or declare it in two different ways.  I'll probably do that at some 
point with the new ABI, but the ugly hack needs to stay for a bit for 
compatibility.

This is documented in README in libobjc2 svn.

David

-- Sent from my IBM 1620