design suggestion: llvm multiple runtime contexts
Solution 1:
I think you need a conceptual framework to "hang" your ideas on. Thinking of the various executing bits as commands (perhaps even implementing using the command pattern) will give you a more obvious set of interaction points. That being said; you will need a context for each discrete execution you wish to return to. More than two will require that you create appropriate book keeping. I believe that two is handled essentially for free in boost.
Communication between executing bits is similarly up to you. Creating a state (memento) that is shared across execution contexts is one solution that comes to mind. You may also already have suitable state built into your run time, then no extra layer will be required. As you pointed out globals are not your friend in these interactions.
Versioning and Name resolution are also an issue. Keeping the executing bits separate goes a long way toward solving this problem. Once you resolve the coordination issue this is more a matter of tracking what bits you have already created. This also means that there is no need for recycling, just create new each time and there is no reload. You will also have to manage the end of life of these bits once they have completed executing.
I am proposing one ExecutionEngine
per executing bit. To not do this means a great deal more work attempting to "protect" working code from the effects of code that is wrong. I believe it is possible to do this with a single engine, but it would be significantly more risky.