[cfe-commits] r60900 - in /cfe/trunk: include/clang/Parse/Ownership.h

[Sebastian Redl]

Sebastian Redl wrote:
> Fariborz Jahanian wrote:
>   
>> I am having the same concerns as Steve. We are adding enormous level
>> of abstractions without looking at its consequences down the road.  
>> Aren't we
>> concerned about potential performance implications for one (smart  
>> pointers don't
>> get smart for  cheap :).
>>     
> They're not free, but they are pretty cheap. However, Doug messed up my 
> planned performance test by doing his name lookup rework at the same 
> time as I did the replacements. I still plan to do some serious testing 
> soon; however, I expect the largest impact to come from transforming the 
> Action interface to smart pointers, and I have only started with this.
>   
So I did some rough performance testing. My test method was thus:

time Release/bin/clang -fsyntax-only insn-attrtab.i

where insn-attrtab.i is a preprocessed version of GCC's insn-attrtab.c, 
a generated 3.3MB behemoth consisting mostly of a huge switch statement.

I went through the SVN log and tested every version directly before and 
after every smart pointer-related commit of mine. I'm happy to say that 
over the entire length of the tested area (-r 60760:60983) the average 
user time needed to process the file has not changed significantly. A 
rough average (read: me running the command 5-10 times and taking a 
number that appeared to be in the middle of the observed times) went 
from 0.50 for 60760 to 0.52 for 60983, with a variation of +-0.03. I 
conclude that the changes done so far had no significant impact on 
performance.

I'll keep testing.

Sebastian