[jira] Updated: (HARMONY-4476) [drlvm][jit][opt][abcd] Two-state Inequality Graph for both Lower and Upper problems, to ensure correctness and simplify the code, ability to dump stats

["Egor Pasko (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HARMONY-4476?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Egor Pasko updated HARMONY-4476:
--------------------------------

    Attachment: 0001-Two-state-Inequality-Graph-with-ABCD-Stats-option.txt

the patch attached

also implemented option: -XX:jit.arg.dump_abcd_stats=true that dumps ABCD stats (total vs. eliminated) to "bounds_checks.log" file.

amount of checks removed: is the same (tested on DaCapo hsqldb, the only benchmark I could run without problems)
tests from HARMONY-2141,  HARMONY-2144, HARMONY-2147 pass as previously

the code is more readable, adds all PI instructions first (to both states of the graph) then marks upper- and lower- redundant checks using this graph. The dot file dumped for the graph depicts upper-only edges in blue color and lower-only in red. Both-states checks are default color (black).

could anybody please run 'build test' on windows just to make sure I did not break something? I do not have windows yet :)

if you find it difficult to apply, let me know, I will merge it quickly into the newest version

> [drlvm][jit][opt][abcd] Two-state Inequality Graph for both Lower and Upper problems, to ensure correctness and simplify the code, ability to dump stats
> --------------------------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: HARMONY-4476
>                 URL: https://issues.apache.org/jira/browse/HARMONY-4476
>             Project: Harmony
>          Issue Type: Improvement
>          Components: DRLVM
>         Environment: all
>            Reporter: Egor Pasko
>         Attachments: 0001-Two-state-Inequality-Graph-with-ABCD-Stats-option.txt
>
>
> Creating two Inequality Graphs makes it difficult to make sure the graph looks like the right one. This issue proposes to create two-state InequalityGraph. One state shall represent the upper-bound problem, the other -- lower-bound. Constraints arising from statements like "a=b", "a=b+const" are the same in both states of the graph, whereas constraints arising from statements like "if (a < b)" are different (basically, the edges change direction to the opposite and negate their length)

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