[LLVMdev] Google Summer of Code Idea

Xi Wang xi.wang at gmail.com
Wed Mar 5 00:44:06 PST 2008


On Mon, Mar 3, 2008 at 6:55 PM, Richard Warburton
<richard.warburton at gmail.com> wrote:
> >  No
> >  I've got a BDD based version of andersen's around already (uncommitted).
> >  It's slower in time (by a factor of 8 or so), but very slightly (1-2
> >  megabytes of memory) more memory efficient on some cases. Nowadays,
> >  the bitmap version is actually more memory efficient than the BDD
> >  version, even if you turn off BDD caching, etc.
> >
> >  The thing you have to understand is that these context sensitive BDD
> >  algorithms work okay on java, but fall down pretty badly on C/C++,
> >  because it is a much bigger problem.   Whaley's is just brute force,
> >  and Zhu's is pretty darn complex.
>
> I was interested to see whether Whaley's algorithm really works
> particularly, given its simplicity and the claims made in the papers.

I have implemented a tool based on the Phoenix compiler framework at
MSR, which can dump relations as BDD from C/C++ code and feed them
into John Whaley's bddbddb for further analysis.  As Dan said,
currently the performance is not good enough for C/C++ code (due to
more complex relations, BDD variable orders, etc.).  I am still trying
several optimizations.  John also told that it took quite a lot of
effort to get the Java version to scale.

I am interested to see an implementation based on LLVM.

Xi



More information about the llvm-dev mailing list