Author: Gareth McCaughan
Date: 15:05:53 05/16/02
Go up one level in this thread
On May 15, 2002 at 19:32:04, Dann Corbit wrote: >>> Actually not. I assumed he was talking about a combination of global >>> optimization together with optimal reordering of the objects to minimize >>> jump distance [at compile time and then at link time]. >>> >>> I have no idea why rewriting the code at runtime would be an advantage. >>> Undoubtably, I simply do not understand the message that I responded to. >> >> Same principle as profile-driven optimization, only more so. >> Track which way jumps go and arrange code accordingly to >> minimize cache misses and pipeline stalls. > > That's profile guided optimization in C++ land. I find it hard to imagine > that doing it on the fly would be faster than profile guided data, but > I often lack imagination. Depends on the application. If the behaviour is consistent within a run but varies greatly between runs, then doing it on the fly can be a win. Or if it varies during the run, but slowly. I don't have any numbers for any of this; I've no idea how often it would be worth doing, or by how much. -- g
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.