harmony-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Egor Pasko (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HARMONY-2170) [drlvm][jit] Suggestion of new Range Check elimination optimization.
Date Tue, 27 Nov 2007 10:32:43 GMT

    [ https://issues.apache.org/jira/browse/HARMONY-2170?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12545784
] 

Egor Pasko commented on HARMONY-2170:
-------------------------------------

Evgueni,

1. why didn't you eliminate FastArrayFilling at all? does it give any boost over your loop
with eliminated bounds checks?
2. why do you call it dynamic ABCD? the transformation you describe is a standard "loop versioning"
(known paper)
3. looking at your patch, you are actually eliminating the bounds checks. Why? this _is_ the
job of classic_abcd and it should make it better. Just invoke it after peeling.

The (3) is especially hurting me because I see no point in implementing yet another BoundsCheck
elimination while you can use more general classic_abcd for that. If it does not work in your
case, just let me know the example. I will fire up debugging it.

BTW, sorry for not replying very soon. The JIRA notification has slipped somewhere between
the cracks :)

> [drlvm][jit] Suggestion of new Range Check elimination optimization.
> --------------------------------------------------------------------
>
>                 Key: HARMONY-2170
>                 URL: https://issues.apache.org/jira/browse/HARMONY-2170
>             Project: Harmony
>          Issue Type: Improvement
>          Components: DRLVM
>         Environment: All.
>            Reporter: Sergey Kuksenko
>            Assignee: Mikhail Fursov
>         Attachments: abcde2.patch, dabce.patch, dabce.patch
>
>
> Currently [drlvm][jit] contains ABCD (range check elimination) algorithm.
> ABCD algorithm works pretty well if iterations over an array is perfomed with array.length
access.
> However, there are a set of code patterns where ABCD algorithm won't work.
> For example, java.util classes ArrayList, Vector, HashMap, etc... very often contains
a field where a real upper bound is stored (and usually this upped bound is less then array.legth).
So, iteration over such arrays usually is performed by the following way:
>  for (int i = 0; i < real_length; i++) {
>     array[i]....
> }
> Where "real_length" is a field which contains real upper bound.
> Very often it is too expensive or impossible for JIT to determine value range for this
field and in such ways (when cycle limit is value obtained outside method) ABCD range check
elimination can't work.
> Suggestion:
> In case of cycle like:
> int i=0;
> while(i < real_length) {
>    checkRange(array, i);  // introduce explicit range check operation
>    ...array[i]...
>    i++;
> }
> Split this cycle to:
> int i=0;
> while(i < min(real_length, array.length) ) {
>    ...array[i]...         // NO range check here
>    i++;
> }
> while(i<real_length) {
>    checkRange(array, i);
>    ...array[i]...
>    i++;
> }
> After such splitting range check is useless at the first cycle and may be removed.

-- 
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