z3.prove is much faster than solver and check

Question

I'm using latest z3 master code from Codeplex, tagged as v4.3.1.

I want a function like prove that has a useful return value and does not print. So, I wrote what seemed obvious:

def prove2(claim):
    s = Solver()
    s.add(Not(claim))
    if s.check() == unsat:
        return True, []
    return False, s.model()

However, this code runs dramatically slower than the default prove function.

The code for prove (slimmed) in src/api/python/z3.py is:

def prove(claim, **keywords):
    s = Solver()
    s.set(**keywords)
    s.add(Not(claim))
    if keywords.get('show', False):
        print s
    r = s.check()
    if r == unsat:
        print "proved"
    elif r == unknown:
        print "failed to prove"
        print s.model()
    else:
        print "counterexample"
        print s.model()

When I add s.set() to my code, it is fast and finds the same counterexample.

What is going on here?

Does that empty call to s.set() somehow clear some option that is bad in general?
.. bad for my particular test?
Something else?

I tried to find out what the default solver options were, but str(s) repr(s), s.__dict__, and google didn't really help.

Any advice is appreciated!

That's an interesting observation. There's currently no convenient way to get the solver options off of a Solver object. Could you perhaps to see what's in **keywords at runtime? — Christoph Wintersteiger, Aug 01 '14 at 15:56
Hi @ChristophWintersteiger, since I'm the one calling `prove(claim)`, I know that `**keywords` is the empty dictionary. — Dan Halperin, Aug 01 '14 at 19:29
Interesting. What exactly do you understand as "dramatically slower"? Do you have a small benchmark that triggers this behaviour? — Christoph Wintersteiger, Aug 03 '14 at 14:28
@ChristophWintersteiger I had an example 9 days ago ;) but I just uncommented the empty call to `set()` and ran my program again and it seemed fine. Maybe I hit a bad RNG state in the one case, I guess... ? — Dan Halperin, Aug 04 '14 at 06:48

score 1 · Accepted Answer · answered Jan 02 '15 at 18:09

1

Best guess is that the default options did poorly for my particular case, perhaps because of a random number difference or some other non-deterministic internal state.

answered Jan 02 '15 at 18:09

Dan Halperin

2,207
1
18
25

z3.prove is much faster than solver and check

1 Answers1