Hi! We had a curious problem with SPECjvm98 _213_javac and cacao. A newly recompiled cacao ended with 10% preformance degradation. After further investigation we ended with VMSystem.identityHashCode() and it's return value. _213_javac make heavy use of Hashtable's and the objects stored in the hashtable use VMSystem.identityHashCode() as their hash function. Since alignment is normally 8 or 16-bytes, the distribution of the objects seems to be not optimal. We ended with 40000 Object.equals() calls more and a doubled runtime for Hashtable.get() (measured clock cycles), as Object.equals() is used in Hashtable.get(). A shifting of the VMSystem.identityHashCode() pointer value helps (sometimes), but shifting by 3 makes it even worse (what was my guess to make it better). Any thoughts? Should we change Hashtable.hash()? Or return a better value for VMSystem.identityHashCode()? But most runtimes only return the memory pointer... TWISTI
