I've just inherited (ahem) a QNX realtime project which uses a void*/downcasting/case statement mechanism to handle messaging. I'd prefer to switch to an abstract base class with pure virtual functions instead but I'm wondering if the original solution was done like that for speed reasons? it looks a lot like it was written originally in C and got moved at some point to C++, so I'm guessing that could be the reason behind it.
Any thoughts on this are appreciated. I don't want to make the code nice, safe and neat and then have it fail for performance reasons during testing.
I doubt that performance will be a concern. If there are sufficient disparate values in the switch/case your compiler may not even optimize it into a jump table, setting up the possibility that the virtual dispatch could be faster than the switch.
If a pure virtual interface makes sense design-wise I would definitely go that way (prototype and profile it if you're really concerned).