Take, for example, an 8-bit/byte store. Without BWX the sequence would be something like:
bic a0, #3, t1
and a0, #3, t4
ldl t2, (t1)
insbl a1, t4, t3
mskbl t2, t4, t2
bis t2, t3, t2
stl t2, (t1)
ret zero, (ra)
But with BWX, it becomes: stb a1, 0(a0)
ret zero, (ra)People use and maintain them and they have very little impact outside arch/ nowadays so they're on the happy side of cost/benefit I guess.
I'm not sure if Alphas are even being made anymore, even 15-20 years ago.
Alpha's memory model has problems with providing atomic access to single bytes, which i'd imagine in a kernel is a bit annoying :-)
And then there's just the social aspect, m68k was used in the Amiga/Atari/Mac/QL/x68k, so there is a whole generation of us m68k fans who are willing to keep it alive.
Alpha has it's fans (me included!), but it's not exactly the same. So in a way it's no surprise it's slowly bitrotting away.
That required this smp_read_barrier_depends() through the kernel, but actually in recent years that has basically been subsumed by other concurrent access primitives that all the core kernel must use, so I think alpha is no longer much of a problem outside arch/alpha