Hacker News Reader

Interesting, relevant paper mentioned in response to the linked SO question:

MAO - an Extensible Micro-Architectural Optimizer - http://static.googleusercontent.com/external_content/untrust...

There are figures on page 6 that answer the OP's question directly. And there's also this money quote: "Building an accurate model for modern processors is virtually impossible."

A good way to test that would be to insert NOP's of the right length in place of the MOV to see if its an alignment or predictor line sharing issue.

Very interesting indeed. In particular: "we ﬁnd that most performance results stay ﬂat, in particular for the Nop Killer pass (NOPKILL), which leads to the conclusion that most alignment directives are not helping"

Reply on news.ycombinator.com