Svadev mailing list

[John Criswell <criswell AT illinois.edu>]

On 1/26/11 9:02 PM, Tarek Chammah wrote:
> Hi SVAers,
>
> I would like to know how fit the code in DSAA's AliasAnalysisGroup
> implementation is, basically the DSAA::alias method at al.  John
> mentioned that it hadn't been used in years and might harbor latent
> bugs.

I spoke with Will and Arushi yesterday.  I'll let them fill you in on 
the details, but basically, the DSA AliasAnalysis Group most likely does 
not work.  The primary problem is that very few LLVM passes use the 
AliasAnalysisGroup interface, and those that do apparently don't follow 
the conventions required for using it correctly.  We therefore have not 
had a good way to test dsa-aa because none of our research code uses it.
> Would you recommend its use compared to directly querying the DSGraph
> nodes?  Does utilizing the thin AA abstraction reduce precision of the
> underlying information?

Yes, you'll need to examine the points-to graphs directly and properly 
interpret the Incomplete, Unknown, and Int2Ptr flags.  Also, you will 
need to do a simple interprocedural analysis to place DSNodes from 
different functions into equivalence classes to properly detect 
aliasing.  Will has written code to do that; I'll email you a link to it 
off-list.

-- John T.

> Thanks,
>
> Tarek
> _______________________________________________
> svadev mailing list
> svadev AT cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/svadev