harmony-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pavel Ozhdikhin (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HARMONY-2084) [drlvm][jit][opt] Jitrino.OPT inliner can cause native SOE exception for some large methods
Date Wed, 15 Aug 2007 10:55:30 GMT

    [ https://issues.apache.org/jira/browse/HARMONY-2084?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12519924
] 

Pavel Ozhdikhin commented on HARMONY-2084:
------------------------------------------

We have a number of entities that should not grow extremely - instructions, operands, nodes,
bytes etc - all those may cause overflow of some optimization. When we encounter such case
we should limit our HIR from infinite grow. This issue is a particluar case of node limitation
in inlining. If another pass needs to use node limit we may consider moving it to HIR IRManager.
Current value of 1300 is got from experimental runs of several bencmarks - almost all compiled
methods do not exceed this threshold. We may want playing with this bound to check if it can
positively affect performance.


> [drlvm][jit][opt] Jitrino.OPT inliner can cause native SOE exception for some large methods
> -------------------------------------------------------------------------------------------
>
>                 Key: HARMONY-2084
>                 URL: https://issues.apache.org/jira/browse/HARMONY-2084
>             Project: Harmony
>          Issue Type: Bug
>          Components: DRLVM
>            Reporter: Mikhail Fursov
>            Assignee: Mikhail Fursov
>            Priority: Minor
>         Attachments: 2084.patch, Test.java
>
>
> The inliner in Jitrino.OPT compiler stops inlining if bytecode size of the current method
is > some threshold.
> Sometimes (for synchronized methods) the bytecode size is small, but the size of the
resulted IR is very large. As the result Inliner can produce very large linear IR graphs that
 cause SOE in recursive algorithms like DFS
> The proposal is to add one more threshold to inliner: stop inlining if the number of
nodes in IR is greater then some const.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message