StringBuilder version of java.util.regex.Matcher.append* (original) (raw)

Peter Levart peter.levart at gmail.com
Tue Apr 8 12:54:23 UTC 2014

On 04/07/2014 07:00 PM, Xueming Shen wrote:

On 04/04/2014 10:08 AM, Xueming Shen wrote:

On 4/3/14 4:43 PM, Jeremy Manson wrote:

Good catch, thanks.

I think we should probably just go with the (equivalent to the) StringBuffer variant. I'm pretty loathe to modify the StringBuilder directly if we are going to back that change out. Do you want me to generate a new patch? I can/will send out an updated webrev before push. the latest webrev. http://cr.openjdk.java.net/~sherman/8039124/webrev -Sherman

Hi Sherman,

This seems most straight-forward and simple. My proposed variant of appending directly to StringBuffer/StringBuilder (thus to AbstractStringBuilder) had also a performance disadvantage, since code in appendExpandedReplacement() could see at least two different subclasses of AbstractStringBuilder and therefore could not inline calls to AbstractStringBuilder's virtual methods and had to use duo or even megamorphic calls. I measured and it appears that this has more overhead than an additional StringBuilder copy for common-sized replacements...

One thing to note. Matcher.appendReplacement(StringBuffer, ...) has never been specified as or implemented to be an atomic operation from the StringBuffer's perspective. But this could easily be achieved:

795 public Matcher appendReplacement(StringBuffer sb, String replacement) { 796 // If no match, return error 797 if (first < 0) 798 throw new IllegalStateException("No match available"); 799 StringBuilder result = new StringBuilder(); 800 appendExpandedReplacement(replacement, result);

801 // Append the intervening text 802 sb.append(text, lastAppendPosition, first); 803 // Append the match substitution 804 sb.append(result);

805 lastAppendPosition = last; 806 return this; 807 }

...I don't know if this makes uncontended locking exhibit any performance penalties since the "sb" monitor is locked twice (nested). I haven't measured, sorry.

Peter

-Sherman

Jeremy

On Thu, Apr 3, 2014 at 2:27 PM, Xueming Shen <xueming.shen at oracle.com <mailto:xueming.shen at oracle.com>> wrote: On 03/25/2014 02:07 PM, Jeremy Manson wrote: Okay. Thanks, Sherman. Here's an updated version. I've diverged a bit from Peter's version. In this version, appendExpandedReplacement takes a StringBuilder. The implications is that In the StringBuilder case, it saves creating a new StringBuilder object. In the StringBuffer case, it creates a new StringBuilder, but it doesn't have to acquire and release all of those locks. Hi Jeremy, It appears the "optimized" StringBuilder version will cause following test case failure, in which the "xyz" will be copied into the result buffer, even when the replacement string triggers a IAE. // Check nothing has been appended into the output buffer if // the replacement string triggers IllegalArgumentException. Pattern p = Pattern.compile("(abc)"); Matcher m = p.matcher("abcd"); StringBuilder result = new StringBuilder(); try { m.appendReplacement(result, ("xyz$g")); } catch (IllegalArgumentException iae) { if (result.length() != 0) System.err.println(" FAILED"); } We may have to either catch the IAE and reset the sb, or create a new sb, as the StringBuffer does. -Sherman

I also noticed a redundant cast to (int), which I removed. Jeremy On Fri, Mar 21, 2014 at 7:34 PM, Xueming Shen <xueming.shen at oracle.com <mailto:xueming.shen at oracle.com>> wrote: let's add the StringBuilder method(s), if you can provide an updated version, I can run the rest (since it's to add new api, there is an internal CCC process need to go through). -Sherman On 3/21/14 5:18 PM, Jeremy Manson wrote: So, this is all a little opaque to me. How do we make the go/no-go decision on something like this? Everyone who has chimed in seems to think it is a good idea. Jeremy On Thu, Mar 20, 2014 at 10:38 AM, Jeremy Manson <jeremymanson at google.com <mailto:jeremymanson at google.com>> wrote: Sherman, If you had released it then (which you wouldn't have been able to do, because you would have to wait another two years for Java 7), you would have found that it improved performance even with C2. It is only post-escape-analysis that the performance in C2 equalized. Anyway, I think adding the StringBuilder variant and deferring / dealing with the Appendable differently is the right approach, FWIW. Jeremy On Thu, Mar 20, 2014 at 10:25 AM, Xueming Shen <xueming.shen at oracle.com <mailto:xueming.shen at oracle.com>> wrote: 2009? I do have something similar back to 2009 :-) http://cr.openjdk.java.net/~sherman/regexreplace/webrev/ <http://cr.openjdk.java.net/%7Esherman/regexreplace/webrev/> Then the ball was dropped around the discussion of whether or not the IOE should be thrown. But if we are going to/have to have explicit StringBuilder/Buffer pair anyway, then we can keep the Appendable version as private for now and deal with the StringBuilder and Appendable as two separate issues. -Sherman On 03/20/2014 09:52 AM, Jeremy Manson wrote: That's definitely an improvement - I think that when I wrote this (circa 2009), I didn't think about Appendable. I take it my argument convinced someone? :) Jeremy On Thu, Mar 20, 2014 at 1:32 AM, Peter Levart<peter.levart at gmail.com <mailto:peter.levart at gmail.com>>wrote: On 03/19/2014 06:51 PM, Jeremy Manson wrote: I'm told that the diff didn't make it. I've put it in a Google drive folder... https://drive.google.com/file/d/0BGaXa6O4K5LY3Y0aHpranM3aEU/ edit?usp=sharing Jeremy Hi Jeremy, Your factoring-out of expandReplacement() method exposed an opportunity to further optimize the code. Instead of creating intermediate StringBuilder instance for each expandReplacement() call, this method could append directly to resulting StringBuffer/StringBuilder, like in the following: http://cr.openjdk.java.net/~plevart/jdk9-dev/MatcherWithStringBuilder/ <http://cr.openjdk.java.net/%7Eplevart/jdk9-dev/MatcherWithStringBuilder/> webrev.01/ Regards, Peter

On Wed, Mar 19, 2014 at 10:41 AM, Jeremy Manson<jeremymanson at google.com_ _<mailto:jeremymanson at google.com>> wrote: Hi folks, We've had this internally for a while, and I keep meaning to bring it up here. The Matcher class has a few public methods that take StringBuffers, and we've found it useful to add similar versions that take StringBuilders. It has two benefits: - Users don't have to convert from one to the other when they want to use the method in question. The symmetry is nice. - The StringBuilder variants are faster (if lock optimizations don't kick in, which happens in the interpreter and the client compiler). For interpreted / client-compiled code, we saw something like a 25% speedup on String.replaceAll(), which calls into this code. Any interest? Diff attached. Jeremy

More information about the core-libs-dev mailing list