lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Uwe Schindler (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (SOLR-13533) Code Cleanup - Performance
Date Mon, 10 Jun 2019 11:05:00 GMT

    [ https://issues.apache.org/jira/browse/SOLR-13533?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16859925#comment-16859925
] 

Uwe Schindler commented on SOLR-13533:
--------------------------------------

Hi,
I am fine with this patch, except for tests. They should not be "optimized" (although it's
no longer a required optimization, see below).

{quote}
So, something like:

StringBuilder sb = new StringBuilder();
sb.append("Hello, " + "World!");
 

Which defeats the very purpose of using a StringBuilder. And also it causes the internal construction
of extra StringBuilder objects by the JVM. And the JVM is very conservative with that optimization.
It does no interpretation, no recognizing that there is already a StringBuilder present and
adding it to that one.

It is true that for simple concatentation, StringBuilder isn't required. But when it is already
being used, then using append for everything is the best option.
{quote}

This is not quite true anymore. Lucene + Solr master is on Java 11 already, so the class files
are compiled for Java 11 and have target Java 11. This means, simple string concats no longer
use a StringBuilder internally. In fact, if you concat a lot of stuff with static components
like fragments and numbers, it's now better to use string concats instead of StringBuilders.
StringBuilders should only be used for loops or if/then/else constructs.

Starting from Java 9, the "+" operator no longer uses StringBuilder and uses instead the new
feature "indyfied string concat" (which is BTW also used by Elasticsearch in its painless
scripting engine). What happens is that it collects all static parts of the concat in to a
static constant pool entry not even passed to the method call, but part of the invokedynamic
signature and the remaining dynamic parts are pushed to stack as is. This allows to highly
optimize the string concats, which are only one single method call solely containing the dynamic
parts to a prepared MethodHandle that contains all static parts.

So we should also review our string concats with StringBuilder and if they don't use if/then/else
and loops, replace them by "+".

> Code Cleanup - Performance
> --------------------------
>
>                 Key: SOLR-13533
>                 URL: https://issues.apache.org/jira/browse/SOLR-13533
>             Project: Solr
>          Issue Type: Improvement
>      Security Level: Public(Default Security Level. Issues are Public) 
>            Reporter: Koen De Groote
>            Priority: Trivial
>              Labels: performance
>
> EDIT: Again, to clarify, please don't bother yourself with this ticket on company time,
on personal time you could be working on something that makes you money or improves the product
for your feature personally.
>  
> This entire ticket is an afterthough. A look back at the code base that most people don't
have the time for.
> ---------
>  
> Code cleanup as suggested by static analysis tools. Will be done in my spare time.
> If someone reviews this, please also do not take up actual time from your work to do
that. I do not wish to take away from your working hours.
>  
> These are simple, trivial things, that were probably overlooked or not even considered(which
isn't an accusation or something negative). But also stuff that the Java compiler/JIT won't
optimize on its own.
>  
> That's what static analysis tool are good for: picking stuff like that up.
>  
> I'm talking about Intellij's static code analysis. Facebook's "Infer" for Java. Google's
"errorprone", etc...
> These are the kinds of things that, frankly, for the people actually working on real
features, are very time consuming, not even part of the feature, and have a very low chance
of actually turning up a real performance issue.
> So I'm opting to have a look at the results of these tools and implementing the sensible
stuff and if something bigger pops up I'll make a separate ticket for those things individually.
>  
> Creating this ticket so I can name a branch after it.
>  
> The only questions I have are: since the code base is so large, do I apply each subject
to all parts of it? Or only core? How do I split it up?
> Do I make multiple PRs with this one ticket? Or do I make multiple tickets and give each
their own PR?



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org


Mime
View raw message