Thanks folks. In the meantime (as part of inspecting that regression that Benjamin Winterberg pointed out)I realized that there often seems to be a significant performance difference between loading an operand as Object and converting to double, and loading it as double (when we're allowed to reorder converts and loads). That is, an "i < size" (where size is in scope and i is double) can be emitted as:
dload 5 // i aload 2 // scope invokedynamic dyn:getProp|getElem|getMethod:size(Object;)D // size dcmpg instead of dload 5 aload 2 invokedynamic dyn:getProp|getElem|getMethod:size(Object;)Object invokestatic JSType.toNumber(Object) dcmpg The first one is actually faster, especially when size is already a number due to dual fields (it was just unprovable statically). For this, I uploaded a small differential patch between what you already reviewed and the new modification. Please review this as well <http://cr.openjdk.java.net/~attila/8035712/webrev.01/> Thanks, Attila. On Mar 4, 2015, at 5:17 PM, Marcus Lagergren <[email protected]> wrote: > +1 if you have done the octane regression analysis. > >> On 04 Mar 2015, at 11:34, Attila Szegedi <[email protected]> wrote: >> >> Please review JDK-8035712 at >> <http://cr.openjdk.java.net/~attila/8035712/webrev.00> for >> <https://bugs.openjdk.java.net/browse/JDK-8035712> >> >> The JIRA issue has notes on the implementation: >> <https://bugs.openjdk.java.net/browse/JDK-8035712?focusedCommentId=13615184&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13615184> >> >> Thanks, >> Attila. >
