Actually in this thing the top two elements of the stack are held in registers. So if you wanted to use the ANSI C compiler instead of FORTH, you might just think of it as a two-register machine (there are actually quite a few more) with a few special stack-manipulation instructions.
How that compares on a practical level, I have no idea. But it's one of the reasons the JVM has a stack-based model and no registers: it makes writing compilers for JVM byte code a lot easier.
In terms of compiler writing, if you wanted to write any optimisations you'd probably want to manipulate 3-address code internally, if only because the data flow based analyses would be made easier (and then you would emit stack code in the last step).
This might be a reason why the JVM designers decided to use a stack machine, but I doubt it's an important one - they'd easily be able to get ahold of a few decent compiler writers if they needed to.
http://research.ihost.com/lcpc06/final/7/7_Paper.pdf
Actually, here's the original for sale, but the abstract doesn't explicitly mention being NP-complete.
http://www.sciencedirect.com/science/article/pii/00960551819...
The biggest disadvantage is their performance because it's hard to make them do operations in parallel.