Loader Cache improvements

[Dirk Mueller]

On Monday 12 January 2004 08:16, David Hyatt wrote:

> In case you're curious, I implemented the LRU-SP algorithm in the
> loader after reading this paper:

Thanks. Looking at it, it doesn't mention the uncachable hack, so I'm fairly 
happy :-) It should also be obvious that this LRU-SP performs much better in 
*ALL* cases than the uncachable hack. 

However, now that you asked me to look at the algorithm ;-) I notice this 
problem: 

static LRUList* getLRUListFor(CachedObject* o)
{
    int accessCount = o->accessCount();
    int queueIndex;
    if (accessCount == 0) {
        queueIndex = 0;
   } else {
        int sizeLog = FastLog2(o->size());
        queueIndex = sizeLog/o->accessCount() - 1;
        if (queueIndex < 0)
            queueIndex = 0;
        if (queueIndex >= MAX_LRU_LISTS)
            queueIndex = MAX_LRU_LISTS-1;
    }
   return &m_LRULists[queueIndex];
}

Technically, shouldn't accessCount == 0 case result in queueIndex= 
MAX_LRU_LISTS-1 ?

Or are you trying to avoid the "object is deleted between requestImage() and 
its first ref() " case ?

Hmm, thats probably it. though we handle this so far by never putting a 
deref() between the requestXXX and the first ref(). As we only flush on 
deref(), that should be fine. 

hmm, I should add an assert there and see if it is ever triggered..


Dirk