Can you give some insight on how the J1 fits in all this? The presentation explaining what drove its invention suggests it actually was faster than the register-based alternative, although I'm sure that is also partially because it only focused on the instructions needed for its task.
From how I understand the presentation, they ran into trouble because of code size. Basically stack code is much denser because you don't need arguments for registers and, given the 16kbytes -> 6 kbytes is more than just half from going 32->16bit, appearantly a bit more efficient for implementing their program too I guess.
The size was a concern because of the limited size (16kbyes) of bram modules on an fpga.
I can't say why it performs so much better though, maybe just because it's simpler & 16 bit so it's smaller and can be clocked faster on the fpga? That is just a guess though.
I'm not sure if it is fair to compare the J1 to the BOOM, the latter does a lot more I'd think. That being said, it shows that for such use cases a stack machine can be better suited and much less complex to design, maintain and work with.
[0] https://excamera.com/sphinx/article-j1a-swapforth.html
[1] http://www.forth.org/svfig/kk/11-2010-Bowman.pdf