[cfe-commits] r125756 - in /cfe/trunk: include/clang/AST/DeclBase.h lib/AST/DeclBase.cpp

[Chris Lattner]

On Feb 18, 2011, at 9:08 AM, Douglas Gregor wrote:

>>> hierarchy far more often in Sema.
>>> 
>>> Anyway, thanks for the prod. 2% speed loss for 3.5% memory improvement isn't worthwhile, IMO, so I'll see about backing out these changes.
>> 
>> Do you know what is causing the performance regression?
> 
> 
> I did some rough testing yesterday, and the main culprits seemed to be Decl::isOutOfLine(), DeclaratorDecl::getInnerLocStart() and TagDecl::getInnerLocStart(). However, there's a *lot* going on on trunk now, so I'm going to do some more fine-grained testing with just the devirtualization changes in the hope that I can salvage this memory optimization.

Great, thanks!  It seems that it should be salvagable, hopefully there is something simple going on :).  In principle, a switch and an indirect call should be ~ the same speed.

-Chris