Re: Problem with NSString

[David Chisnall]

On 30 Jan 2011, at 21:35, Ivan Vučica wrote:

> That is awesome? Perhaps optimization pass could take a look if NSString 
> subclasses exist and optimize if that is not the case as well?

There's already a profile-driven pass that will look at this in libobjc2 svn.  
In fact, it doesn't check if there are subclasses, it checks if the instances 
are always (or usually) the same class (so, even if they are GSSubString or 
something, it will still work).  If they are then it adds a cache for the 
lookup and tries speculative inlining.

If the speculative inlining works, then it's slightly faster tan a C function 
call that is not inlined (of course, any place where speculative inlining will 
work, C static inlining will also work, so C will actually be faster).

>>> This is currently only done in an experimental optimisation pass on my 
>>> machine, but it will probably become one of the standard Objective-C passes 
>>> used by clang for the GNUstep runtime sometime this year.
>> 
>> Great! I am looking forward to this.
> 
> Ditto!


I'll try to finish / commit it at FOSDEM then.

It's not a high priority, because message send overhead is rarely the 
bottleneck for message sends to strings.  I just put it together because I'm a 
bit OCD when it comes to optimising stuff...

David

-- Sent from my IBM 1620