We found it is much easier to handle exceptions without risking stack leaks using a 3-address register machine. Code generation is also simplified in a register machine compared to a stack machine (at least in the case of a VM designed specifically to run SQL statement).

The VM used by SQLite is different from the VM used by things like Javascript or Python in that SQL is not a general-purpose programming language. Most of the opcodes in the SQLite VM are heavy-weight concepts such as "insert a key/value pair into a b-tree" or "decode the N-th column of a table row". These opcodes are implemented with hundreds or thousands of lines of C code. The overhead of instruction dispatch is insignificant in comparison. And so there isn't much of a performance advantage one way or the other between a stack machine and a register machine in SQLite. The choice of a register machine is for maintainability, reliability, and testability.

I do not know, but if I had to guess, it is for performance.

I am no expert in this field, but AFAIK http://static.usenix.org/events/vee05/full_papers/p153-yunhe... (discussed on http://news.ycombinator.com/item?id=704576) has not really been refuted.

I base that on the fact that both Dalvik (http://davidehringer.com/software/android/The_Dalvik_Virtual...) and Lua 5.0 (http://www.lua.org/doc/jucs05.pdf) use register-based VMs.