> To mitigate this we can use locks around the statements that update the counters. However, these locks are expensive and we’ve been hesitant to place locks around all the counters just to make sure they are accurate. Performance over accurate statistics, right?
Does this mean the Varnish authors have intentional data races all over the codebase? Can't the compiler do whatever it wants in these cases (because it's undefined behavior)? Can't this completely blow up?
http://software.intel.com/en-us/blogs/2013/01/06/benign-data...
Edit: I'm not an expert on this stuff so I'm probably wrong -- maybe someone can tell me what I'm missing.