StringBuilder buffer allocation support in compiler

[Remi Forax]

On 06/10/2014 12:57 AM, John Rose wrote:
> On Jun 9, 2014, at 3:39 PM, Fredrik Öhrström <oehrstroem at gmail.com 
> <mailto:oehrstroem at gmail.com>> wrote:
>
>> I do not know how Hotspot optimizes string concatenations, but 
>> JRockit did some fun things as well, like replacing StringBuilder (at 
>> runtime) with an internal JRockit specific StringMaker, that merely 
>> stored references to the constant strings in an array.
>
> But could JRockit EA-away the array, unroll the loops, and turn 
> everything into custom code?
>
> Come to think of it, JRockit could.  Nice trick.
>
> Hmm, I wonder if we have any other types that set up with cumulative 
> method-call chains but need to boil down to custom code.  There was 
> this thing called "Streams", wasn't there?  :-)
>
> My bottom line: The JVM needs to get a better handle on these builder 
> patterns.
>
>> Thus the .append(x) did very little work, all the concatenation work 
>> then happened in toString() where it first scanned the appended 
>> strings length and could calculate the final length of the 
>> concatenated string. Thus only a single allocation was necessary and 
>> there was no superfluous copying.
>
> On Jun 6, 2014, at 9:25 AM, Robert Field <Robert.Field at Oracle.com 
> <mailto:Robert.Field at Oracle.com>> wrote:
>
>> Notes from the peanut gallery:  If this is an important case to 
>> optimize, then a set of constructors which take N Strings seems it 
>> would give you much more.
>
> For built-in stuff like string formatting expressions, lambda 
> expressions, and (eventually) object constructor expressions 
> ("enhanced literals"), a very good trick, IMO, is coding the form of 
> the result into an invokedynamic, and letting a system metafactory 
> figure out the optimal implementation.  As you know, Robert, with a MF 
> you can pretend you have an infinite family of object constructors, 
> each with exactly the right signature required at the point of use. 
>  This would work fine for string formatting expressions.  It would be 
> about as compact (or slightly more so) than the current StringBuilder 
> code shapes, but would allow the JVM the option to generate exactly 
> customized code for each combination of constant and variable 
> formatting components.
>
> — John

and the real question is how to let API developers to specify their own 
meta-factory (a meta protocol ?)

Rémi


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/compiler-dev/attachments/20140610/4160f44b/attachment.html>