[llvm-commits] [llvm] r46224 - /llvm/trunk/include/llvm/ADT/ImmutableSet.h

[Ted Kremenek]

On Jan 21, 2008, at 2:36 PM, Chris Lattner wrote:

> On Jan 21, 2008, at 2:33 PM, Ted Kremenek wrote:
>> URL: http://llvm.org/viewvc/llvm-project?rev=46224&view=rev
>> Log:
>> Replaced (FoldingSet) profiling of ImutAVLTree with a hashing based
>> scheme. The
>> problem was that we previously hashed based on the pointers of the
>> left and
>> right children, but this is bogus: we can easily have different
>> trees that
>> represent the same set. Now we use a hashing based scheme that
>> compares the
>> *contents* of the trees, but not without having to do a full scan of
>> a tree. The
>> only caveat is that with hashing is that we may have collisions,
>> which result in
>> two different trees being falsely labeled as equivalent. If this
>> becomes a
>> problem, we can add extra data to the profile to hopefully resolve
>> most
>> collisions.
>
> I don't think this works Ted: collisions *will* cause the tree to be
> corrupt, and you can't guarantee that they won't happen.  I don't
> think this is valid for all clients of the data structure.

I agree that it doesn't work for all clients.  I'll come up with  
something more sound that works for all clients, while providing  
something like the hashing scheme that is much quicker but doesn't  
provide the same guarantees.