Chuck Silvers <chuq@chuq.com> writes:
> ah, that's what you're seeing.  yes, we will create virtual aliases
> in the pipe code now that the hardware will detect, but those aren't
> a problem because all the mappings are read-only.  so if there's no
> L2 cache to notice the aliases then we'll actually have aliased cache
> lines, but those will always have the same data.

OK, I see.

for the record:

1) NEW_PIPE was broken on port-mips with VACs and has no hardware
   machinery to detect virtual alias (data corruption).
2) the cause of the problem is DIRTY virtual alias.
3) cache flush was added to pmap.c(rev. 1.142).
4) pmap.c(1.142) did eliminate DIRTY virtual alias, but leaves CLEAN
   virtual alias.
5) hardware may detect virtual alias, but it is CLEAN and there should
   be no data corruption.
6) NEW_PIPE is now OK.

Thank you for the patience.

--- Takao Shinohara