[llvm-commits] [llvm] r84038 - in /llvm/trunk: lib/Analysis/BasicAliasAnalysis.cpp test/Analysis/BasicAA/phi-aa.ll

[Chris Lattner]

On Oct 13, 2009, at 6:31 PM, Evan Cheng wrote:
>> Did you look at the compile time impact of this on something like  
>> GVN on 403.gcc?
>
> I am not seeing any compile time impact.

Ok great!

>>
>> Because you only have one 'VisitedPHIs' set, you are conflating  
>> PHIs walked on the LHS with phis walked on the RHS.  I guess it  
>> probably doesn't matter because if the same phi shows up on both  
>> sides, they definitely alias.  This is worth mentioning in a  
>> comment somewhere.
>
> The analysis stops asap it hits a phi source which is itself a phi.  
> So the worst possible case if both LHS and RHS are phis. In that  
> case, its m x n.

Yep, makes sense, please add a comment somewhere to this effect.

Thanks for working on this, improving AA is a great thing to do.

-Chris