… without interlocked pipeline stages was an early weird RISC bet not unlike register windows, branch delay slots or tagged integers. They abandoned it in R4000.
One (probably the only) nifty thing the branch delay slot allowed was in the user-space threads implementation, you didn't have to burn a register for the jump to another thread, you'd do the branch and then clobber the register with the last bit of the jumped-to thread's context. Kinda fiendishly clever.
I'll have to look up the nifty trick. I have a soft spot for delay slots.