Hi,
I am trying to implement some cache replacement/management in M5.
I have problem with one of them which is quite simple and involves little
modification.
It's called dynamic exclusion policy. Based on some history of the accesses,
I want to prevent some accesses to replace the corresponding block.
I just need direct mapped cache. So, let's say, normally after a miss on B,
B is fetched from next level cache (or memory) and the replaces A.
Assume that I know that I don't want B to replace A.
After a miss is satisfied from the memory, using 'handleFill' data would be
replaced.
***************************************************************************
} else if (bus_pkt->isRead() ||
bus_pkt->cmd == MemCmd::UpgradeResp) {
// we're updating cache state to allow us to
// satisfy the upstream request from the cache
blk = handleFill(bus_pkt, blk, writebacks);
satisfyCpuSideRequest(pkt, blk);
}
***************************************************************************
And here is the relevant part in handle fill:
***************************************************************************
if (blk == NULL) {
// better have read new data...
assert(pkt->hasData());
// need to do a replacement
blk = allocateBlock(addr, writebacks);
if (blk == NULL) {
// No replaceable block... just use temporary storage to
// complete the current request and then get rid of it
assert(!tempBlock->isValid());
blk = tempBlock;
tempBlock->set = tags->extractSet(addr);
tempBlock->tag = tags->extractTag(addr);
DPRINTF(Cache, "using temp block for %x\n", addr);
} else {
int id = pkt->req->hasContextId() ? pkt->req->contextId() : -1;
tags->insertBlock(pkt->getAddr(), blk, id);
}
***************************************************************************
Since it has been a miss, the first blk==NULL is satisfied. I thought that,
whenever I want to prevent B from replacing A I can simply return a NULL
pointer in allocateBlock.
Is this gonna work? For example, assume that findVictim always returns NULL
meaning that no replacement is carried out (Cache lines/blocks are filled
only when they are empty/invalid).
Shouldn't all requests become satisfied from memory and simulation run
correctly? I just modified findVictim to always return NULL to test this
characteristic. Neither this way, nor when I
selectively return NULL for findVictim (and hence for allocateBlock) it
works. I expected that using tmpBlock the requests are satisfied in these
cases.
If my expectation is incorrect, anybody knows how can I prevent a based on
the address of the new request and the current resident of the block?
Thank you,
Navid.
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
