Fail gracefully when Node Table is full

[ssoelvsten]

The Adiar BDD package will inherently always struggle to deal with small decision diagrams in comparison to the classical depth-first implementations. Hence, combining a depth-first implementation with Adiar would create something that is the best of both worlds.

How I imagine this could be achieved is as follows:

• If Sylvan runs out of memory: Run Adiar's algorithms on the BDDs from Sylvan (wrapping the node table into something that works with Adiar's algorithms)
• If the output from Adiar is very small: During the Reduction Phase, create the nodes in Sylvan's unique node table instead in one of Adiar's files.

For this to work, I'd need Sylvan to not exit the entire program, if it runs out of memory. Instead, it needs to fail gracefully and merely returns a null pointer. Would that be a lot of work?

[trolando]

I could at least try to return sylvan_memout as a magic return value, not sure if I would also need to change all algorithms to check for these values though. Crashing out is so convenient! But I see that not every user may want this as the outcome.

[ssoelvsten]

The motivation (use an externally developed BDD package in tandem with Adiar) has also been made obsolete; I am intending to (re)implement depth-first algorithms inside of Adiar.

I'll leave this open, as CUDD/CAL does a "graceful" failure like that. It might be useful for other cases.

[trolando]

Maybe I'll keep this open for a Sylvan 2.1 or later version then. I think explicitly adding an "error state" as a result of an operation will introduce a lot of overhead and potential for errors. We currently have sylvan_invalid but that should never be given to an operation, as it will just cause a crash. If we want to properly handle such error states, then a LOT must be changed, much like how CUDD has to handle special error states like when automated reordering has been done and the operation must be restarted.