While investigating the performance of HashMap.get(k) I noticed that in cases where the static type of k is available C2 isn't taking advantage of the information in making it's inline decisions for get.  The problem is that Phis which are introduced for the join of a diamond aren't being eliminated aggressively during parsing when one of the paths is proved dead.  Additionally unneeded loop phis are hiding the type k in the search loop, impairing the inlining of equals.  Post parse inlining which remove the need for aggressive elimination of phis but for now we need a better strategy for simplification of phis.
Posted Date : 2006-02-10 17:21:33.0

N/A

Some simple changes to parse collapse phis for dead codes paths.  A better strategy is to rely on RPO for parse order and when you reach a block with unparsed predecessors, assume they are dead unless your block is a backward branch target.

Eliminating phis for loops requires some analysis though it can be very cheap if we are looking for the simple case.
Posted Date : 2006-02-10 18:21:07.0

http://hg.openjdk.java.net/jdk7/hotspot-comp/hotspot/rev/194b8e3a2fc4
Posted Date : 2008-09-17 22:10:07.0