Copying to list.

On Wed, Dec 08, 2010 at 10:14:43PM +0000, Peter Collingbourne wrote:
> On Wed, Dec 08, 2010 at 10:30:13PM +0100, Mauro Baluda wrote:
> > On Wed, Dec 8, 2010 at 9:20 PM, Peter Collingbourne <peter at pcc.me.uk> 
> > wrote:
> > > Given a constraint  solver with an FP theory this would be easy
> > 
> > Wouldn't it be possible to use a software implementation of FP
> > arithmetics as a front end to STP?
> > http://www.jhauser.us/arithmetic/SoftFloat.html
> > I would like to experiment in this direction
> 
> Hi Mauro,
> 
> We have considered using software FP in the past.  However, we
> decided against it as we see a couple of pitfalls associated with
> this technique:
> 
> 1. Software FP implementations (by necessity) use a significant amount
>    of branching.  So with a symbolic FP argument and a relatively
>    trivial number of operations you would start getting state space
>    explosions rather quickly.  You could try solving this problem by
>    replacing branches with 'select' instructions: phi-node folding
>    is one way to do this -- this is discussed in my talk -- but this
>    technique is still limited.
> 
> 2. Even if you solved this problem then you would still end up with very
>    large expression trees which would be taxing not only on RAM but on
>    the constraint solver.  Specialised FP solvers can take advantage
>    of knowledge of FP semantics for optimisation purposes.
> 
> However you are welcome to explore this solution using KLEE-FP and
> I would be interested to hear about any progress made.
> 
> Thanks,
> -- 
> Peter

-- 
Peter

Reply via email to