harmony-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "George Timoshenko (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HARMONY-5602) [drlvm][jit][opt][performance] nullcheck for a field of boolean type
Date Fri, 01 Aug 2008 09:44:33 GMT

    [ https://issues.apache.org/jira/browse/HARMONY-5602?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12618969#action_12618969

George Timoshenko commented on HARMONY-5602:

I got a question regarding the issue. Here is my answer.

>I think that you mean to save a register for t13, and finally generate IA-32 instruction
like CMP [memory address], src2, right?
Yes, you are absolutely right.

>However, I am still confused with what HIR to generate for the two lines of bytecode.
Would you please explain your thoughts more on that?
You mean: you do not know what to generate in HIR instead of the lines I have qouted?

    I27:ldflda [t1.java/lang/ThreadGroup::destroyed] -) t12:ref:bool
    I28:ldind.unc:b [t12] ((t10,t11)) -) t13:int32
    I29:ldci4 #0 -) t14:int32
    I30:if ceq.i4 t13, t14 goto L4

The point is these lines are correct. They are made by translator. Translator does not know
if t13 will be used other than just that comparison or not.
when it generates GETFIELD.

My idea is:

find some place in HLO path where use-def information is available (memopt for example) or
create a new HLO optimization which calculates use-defs for itself.
Then we need to identify all OP_LdInd instructions with mem_item_size < dst_size and the
only usage of dst which mast be a comparison (or any other instruction that allows memory
The significant condition: this only usage must be able to reduce its size to the mem_item_size.
Finally we can reduce:

replace LdInd dst operand (t13 in our example) with the one of mem_item_size size. Do the
same for use instruction.
And then make sure that the propagation finally happens and codegen does not generate a MOV
from memory.

Actually the optimization can be done in the codegenerator.
(Find MOVZX with dst that is used only once in the comparison with immediate which fits into
Then remove the MOVZX and replace the original CMP with the smaller one using MOVZX's src
(memory) operand )
I am not sure where the optimization will be more suitable.

> [drlvm][jit][opt][performance] nullcheck for a field of boolean type
> --------------------------------------------------------------------
>                 Key: HARMONY-5602
>                 URL: https://issues.apache.org/jira/browse/HARMONY-5602
>             Project: Harmony
>          Issue Type: Improvement
>          Components: DRLVM
>            Reporter: George Timoshenko
>            Assignee: Mikhail Fursov
>            Priority: Minor
> The problem is reproducible at least for java_lang_ThreadGroup.remove(Ljava_lang_Thread;)V
> It contains the following byte codes:
> BYTECODE: GETFIELD        bc-offset=7 (java/lang/ThreadGroup::destroyed)
> BYTECODE: IFEQ        bc-offset=10
> HIR for these two bytecodes is:
> I27:ldflda    [t1.java/lang/ThreadGroup::destroyed] -) t12:ref:bool
> I28:ldind.unc:b   [t12] ((t10,t11)) -) t13:int32
> I29:ldci4     #0 -) t14:int32
> I30:if ceq.i4  t13, t14 goto L4
> As you can see bool value is extended to 32-bit integer type type. (The reason is: 3.3.4
of JVM spec.http://java.sun.com/docs/books/jvms/second_edition/html/Overview.doc.html#22909)
> But in the case there are no any other usage of t13. So the comparison can be performed
without loading the field from memory to a register.
> (there is no copy propagation because of MOVZX not just a MOV for loading from memory)
> The same problem may persist for all integer fields of less than 32 bits size.

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

View raw message