Cache: Fix the LRU policy for classic memory hierarchy
[gem5.git] / src / mem / cache / tags / lru.cc
index 81d82c231a5f40c03ab03b4a96cb8e55c5abf853..c73f557b9dcef054bda0ab213a21ffda81ebf3ab 100644 (file)
 
 #include <string>
 
-#include "mem/cache/base.hh"
 #include "base/intmath.hh"
+#include "debug/CacheRepl.hh"
+#include "mem/cache/tags/cacheset.hh"
 #include "mem/cache/tags/lru.hh"
+#include "mem/cache/base.hh"
 #include "sim/core.hh"
 
 using namespace std;
 
-LRUBlk*
-CacheSet::findBlk(Addr tag) const
-{
-    for (int i = 0; i < assoc; ++i) {
-        if (blks[i]->tag == tag && blks[i]->isValid()) {
-            return blks[i];
-        }
-    }
-    return 0;
-}
-
-
-void
-CacheSet::moveToHead(LRUBlk *blk)
-{
-    // nothing to do if blk is already head
-    if (blks[0] == blk)
-        return;
-
-    // write 'next' block into blks[i], moving up from MRU toward LRU
-    // until we overwrite the block we moved to head.
-
-    // start by setting up to write 'blk' into blks[0]
-    int i = 0;
-    LRUBlk *next = blk;
-
-    do {
-        assert(i < assoc);
-        // swap blks[i] and next
-        LRUBlk *tmp = blks[i];
-        blks[i] = next;
-        next = tmp;
-        ++i;
-    } while (next != blk);
-}
-
-
 // create and initialize a LRU/MRU cache structure
 LRU::LRU(unsigned _numSets, unsigned _blkSize, unsigned _assoc,
          unsigned _hit_latency)
@@ -108,20 +73,21 @@ LRU::LRU(unsigned _numSets, unsigned _blkSize, unsigned _assoc,
     warmupBound = numSets * assoc;
 
     sets = new CacheSet[numSets];
-    blks = new LRUBlk[numSets * assoc];
+    blks = new BlkType[numSets * assoc];
     // allocate data storage in one big chunk
-    dataBlks = new uint8_t[numSets*assoc*blkSize];
+    numBlocks = numSets * assoc;
+    dataBlks = new uint8_t[numBlocks * blkSize];
 
     unsigned blkIndex = 0;       // index into blks array
     for (unsigned i = 0; i < numSets; ++i) {
         sets[i].assoc = assoc;
 
-        sets[i].blks = new LRUBlk*[assoc];
+        sets[i].blks = new BlkType*[assoc];
 
         // link in the data blocks
         for (unsigned j = 0; j < assoc; ++j) {
             // locate next cache block
-            LRUBlk *blk = &blks[blkIndex];
+            BlkType *blk = &blks[blkIndex];
             blk->data = &dataBlks[blkSize*blkIndex];
             ++blkIndex;
 
@@ -149,21 +115,21 @@ LRU::~LRU()
     delete [] sets;
 }
 
-LRUBlk*
-LRU::accessBlock(Addr addr, int &lat, int context_src)
+LRU::BlkType*
+LRU::accessBlock(Addr addr, int &lat, int master_id)
 {
     Addr tag = extractTag(addr);
     unsigned set = extractSet(addr);
-    LRUBlk *blk = sets[set].findBlk(tag);
+    BlkType *blk = sets[set].findBlk(tag);
     lat = hitLatency;
     if (blk != NULL) {
         // move this block to head of the MRU list
         sets[set].moveToHead(blk);
         DPRINTF(CacheRepl, "set %x: moving blk %x to MRU\n",
                 set, regenerateBlkAddr(tag, set));
-        if (blk->whenReady > curTick
-            && blk->whenReady - curTick > hitLatency) {
-            lat = blk->whenReady - curTick;
+        if (blk->whenReady > curTick()
+            && blk->whenReady - curTick() > hitLatency) {
+            lat = blk->whenReady - curTick();
         }
         blk->refCount += 1;
     }
@@ -172,27 +138,23 @@ LRU::accessBlock(Addr addr, int &lat, int context_src)
 }
 
 
-LRUBlk*
+LRU::BlkType*
 LRU::findBlock(Addr addr) const
 {
     Addr tag = extractTag(addr);
     unsigned set = extractSet(addr);
-    LRUBlk *blk = sets[set].findBlk(tag);
+    BlkType *blk = sets[set].findBlk(tag);
     return blk;
 }
 
-LRUBlk*
+LRU::BlkType*
 LRU::findVictim(Addr addr, PacketList &writebacks)
 {
     unsigned set = extractSet(addr);
     // grab a replacement candidate
-    LRUBlk *blk = sets[set].blks[assoc-1];
-    if (blk->isValid()) {
-        replacements[0]++;
-        totalRefs += blk->refCount;
-        ++sampledRefs;
-        blk->refCount = 0;
+    BlkType *blk = sets[set].blks[assoc-1];
 
+    if (blk->isValid()) {
         DPRINTF(CacheRepl, "set %x: selecting blk %x for replacement\n",
                 set, regenerateBlkAddr(blk->tag, set));
     }
@@ -200,32 +162,69 @@ LRU::findVictim(Addr addr, PacketList &writebacks)
 }
 
 void
-LRU::insertBlock(Addr addr, LRU::BlkType *blk, int context_src)
+LRU::insertBlock(Addr addr, BlkType *blk, int master_id)
 {
     if (!blk->isTouched) {
         tagsInUse++;
         blk->isTouched = true;
         if (!warmedUp && tagsInUse.value() >= warmupBound) {
             warmedUp = true;
-            warmupCycle = curTick;
+            warmupCycle = curTick();
         }
     }
 
+    // If we're replacing a block that was previously valid update
+    // stats for it. This can't be done in findBlock() because a
+    // found block might not actually be replaced there if the
+    // coherence protocol says it can't be.
+    if (blk->isValid()) {
+        replacements[0]++;
+        totalRefs += blk->refCount;
+        ++sampledRefs;
+        blk->refCount = 0;
+
+        // deal with evicted block
+        assert(blk->srcMasterId < cache->system->maxMasters());
+        occupancies[blk->srcMasterId]--;
+    }
+
     // Set tag for new block.  Caller is responsible for setting status.
     blk->tag = extractTag(addr);
 
+    // deal with what we are bringing in
+    assert(master_id < cache->system->maxMasters());
+    occupancies[master_id]++;
+    blk->srcMasterId = master_id;
+
     unsigned set = extractSet(addr);
     sets[set].moveToHead(blk);
 }
 
 void
-LRU::invalidateBlk(LRU::BlkType *blk)
+LRU::invalidateBlk(BlkType *blk)
 {
     if (blk) {
+        if (blk->isValid()) {
+            tagsInUse--;
+            assert(blk->srcMasterId < cache->system->maxMasters());
+            occupancies[blk->srcMasterId]--;
+            blk->srcMasterId = Request::invldMasterId;
+        }
         blk->status = 0;
         blk->isTouched = false;
         blk->clearLoadLocks();
-        tagsInUse--;
+
+        // should be evicted before valid blocks
+        unsigned set = blk->set;
+        sets[set].moveToTail(blk);
+    }
+}
+
+void
+LRU::clearLocks()
+{
+    for (int i = 0; i < numBlocks; i++){
+        blks[i].clearLoadLocks();
     }
 }