Isn't non-TEST_AND_SET code long dead?

I was bemused to notice this afternoon that the backend does not build
if you have not defined HAVE_TEST_AND_SET; furthermore, this has been
true at least since 6.4. (slock() is compiled anyway, and it calls
TAS(), which will be an undefined symbol.) From the lack of
complaints we can deduce that no one has run Postgres on a
non-TEST_AND_SET platform in quite a while.

Kinda makes me wonder what other bit-rot has set in in the non-TAS
code, and whether we ought not just rip it out rather than try to
"maintain" exceedingly delicate code that's gone untested for years.
bufmgr.c, in particular, has behavior that's nontrivially different
when HAVE_TEST_AND_SET isn't defined --- who wants to promise that
that still works?

regards, tom lane