Re: [jira] Commented: (HARMONY-6045) [classlib] [luni] Optimize java.lang.String.toUpperCase(), String.toLowerCase() and String sharing for more performance gains

[Regis]

Hi,

I uploaded a new micro benchmark StringConvertCase.v2.java to HARMONY-6045, 
which also compare String.compareToIgnoreCase.

Following are results on Linux:

* Without any patches:
String.toUpperCase
count per cycle: 1000000
2681 2646 2656 2672 2666
String.toLowerCase
count per cycle: 1000000
1888 1899 1890 1855 1860
String.compareToIgnoreCase
count per cycle: 1000000
4177 3959 3958 3995 3998

* Apply patch HARMONY-6045v2.diff:
String.toUpperCase
count per cycle: 1000000
5736 5699 5690 5647 5714
String.toLowerCase
count per cycle: 1000000
1145 1146 1194 1183 1146
String.compareToIgnoreCase
count per cycle: 1000000
3397 2888 2887 2888 2890

It seems HARMONY-6045v2.diff cause performance degradation  on 
String.toUpperCase. So I reverted String.toUpperCase, changed 
Character.toUpperCase to toUpperCase instead, and some other minor fixes (patch 
HARMONY-6045.v3.diff), I got:

* Apply patch HARMONY-6045.v3.diff:
String.toUpperCase
count per cycle: 1000000
2059 2010 2052 2016 2048
String.toLowerCase
count per cycle: 1000000
1145 1185 1137 1170 1136
String.compareToIgnoreCase
count per cycle: 1000000
1785 1492 1532 1501 1542

String.toUpperCase, String.toLowerCase and String.compareToIgnoreCase all got 
improvement, so I proposal to apply HARMONY-6045.v3.diff first. It seems 
HARMONY-6045v2.diff contains code to deal with specific locale "az", I hope some 
experts in this area may give some advices.

Regis wrote:
I found at String:619, it should be
return codePoint - ('a' - 'A');

and at String:595, it should be
return (char) (ch + ('a' - 'A'));

Is it right? And for private char compareValue, could we return directly 
if it's ASCII lower case, like:
    private char compareValue(char ch) {
        if (ch < 128) {
            if ('A' <= ch && ch <= 'Z') {
                return (char) (ch + ('a' - 'A'));
            }
            return ch;
        }
        return Character.toLowerCase(Character.toUpperCase(ch));
    }

Regis wrote:
Kevin Zhou wrote:
It looks good! :)

Not so good, it cause test failures. seems most of them are failed in 
compareToIgnoreCase and toUpperCase. Would you please take a look at it?


On Mon, Apr 20, 2009 at 1:53 PM, Regis <xu.re...@gmail.com> wrote:

Regis wrote:

Kevin Zhou wrote:

Previously, it is proved by using benchmarks like Specjbb.
Since it's been a long time I don't really remember the accurate
statistics
of how much improvements it brings.

All right, I will try to write some micro benchmarks :)

I have attached a very simple micro benchmark for converting case of 
ASCII
character. The patch seems great:

* before the patch HARMONY-6045v2.diff:
String.toUpperCase
count per cycle: 1000000
2779 2727 2706 2768 2745
String.toLowerCase
count per cycle: 1000000
1929 1934 1940 1963 1911

* after the patch:
String.toUpperCase
count per cycle: 1000000
1380 1350 1355 1359 1423
String.toLowerCase
count per cycle: 1000000
1182 1189 1189 1177 1171

* RI 1.5:
String.toUpperCase
count per cycle: 1000000
3123 3181 3132 3117 3156
String.toLowerCase
count per cycle: 1000000
3375 3328 3394 3336 3390

for String.toUpperCase, average from 2745 to 1373.4, improve almost 
100%
for String.toLowerCase, average from 1935.4 to 1181.6, improve 64%

But I noticed there are tests failed after patch:
org.apache.harmony.luni.tests.java.lang.Character_UnicodeBlockTest
org.apache.harmony.luni.tests.java.lang.String2Test

org.apache.harmony.luni.tests.internal.net.www.protocol.file.FileURLConnectionTest 

org.apache.harmony.luni.tests.java.util.Arrays2Test
org.apache.harmony.luni.tests.java.net.URITest
org.apache.harmony.luni.tests.java.net.URLConnectionTest


--
Best Regards,
Regis.








--
Best Regards,
Regis.