> Often there are two types of special IPRs or memory regions (or ASIs > in SPARC land): one that says "flush the cache line (if any) that > matches this address" and another that says "flush the contents of the > cache line at this physical index into the cache array". So a generic > cache flush would be a loop over the N lines in the cache using the > second type of access, while if you're manually trying to keep a > particular location coherent you might use the first type of access. But how would you flush just the lines for a particular address space? I'm guessing that no machines do that today, but I can imagine machines of the future not having full coherence across the entire chip, and wanting to do some sort of coarse software based coherence for process migration. I could imagine a "flush cache line X if it matches asn Y" instruction.
Nate _______________________________________________ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users