JDK-6461827 : Performance: java.lang.Class.isAssignable
  • Type: Bug
  • Component: hotspot
  • Sub-Component: compiler
  • Affected Version: 5.0u6
  • Priority: P3
  • Status: Resolved
  • Resolution: Fixed
  • OS: solaris_10
  • CPU: sparc
  • Submitted: 2006-08-19
  • Updated: 2011-03-24
  • Resolved: 2007-03-29
The Version table provides details related to the release that this issue/RFE will be addressed.

Unresolved : Release in which this issue/RFE will be addressed.
Resolved: Release in which this issue/RFE has been resolved.
Fixed : Release in which this issue/RFE has been fixed. The release containing this fix may be available for download as an Early Access Release or a General Availability Release.

To download the current JDK release, click here.
Other JDK 6 JDK 7 Other
5.0u12,hs10Fixed 6u2 b01Fixed 7Fixed hs10Fixed
Related Reports
Relates :  
Relates :  
Relates :  
The performance of java.lang.Class.isAssignable api in SUN's jdk1.5.0_06 build 05 lags behind the peformance of the same in JRockit's jdk1.5.0_06. 

The JRockit api takes 50% (or half) of the time that SUN's jdk1.5.0_06 takes, after the "warm-up". This was measured using a simple micro-benchmark that repeatedly called isAssignableFrom. 


The measurement was done on Solaris 10 on SPARC.

Right now, I am filing this bug so that it's on the radar. I will be digging deeper for more details.

SUGGESTED FIX Here is the #ifdef removal which makes the previous fixes work: --- /export/jrose/ws/tiger-6461827/webrevs/6461827c/src/share/vm/opto/graphKit.cpp- Mon Jan 29 13:15:14 2007 +++ graphKit.cpp Fri Jan 26 23:26:21 2007 @@ -1865,17 +1865,13 @@ r_ok_subtype->add_req( _gvn.transform( new (1) IfTrueNode ( iff3 ) ) ); set_control( _gvn.transform( new (1) IfFalseNode( iff3 ) ) ); // Now do a linear scan of the secondary super-klass array. Again, no real // performance impact (too rare) but it's gotta be done. -#ifdef IA64 Node *psc = new (3) PartialSubtypeCheckNode( subklass, superklass ); - psc->set_req(0, control()); + psc->set_req(0, control()); // only call it when we must Node *idx = _gvn.transform( psc ); -#else - Node *idx = _gvn.transform( new (3) PartialSubtypeCheckNode( subklass, superklass ) ); -#endif Node *cmp4 = _gvn.transform( new (3) CmpINode( idx, _gvn.intcon(0) ) ); Node *bol4 = _gvn.transform( new (2) BoolNode( cmp4, BoolTest::ne ) ); IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); r_not_subtype->add_req( _gvn.transform( new (1) IfTrueNode ( iff4 ) ) );

SUGGESTED FIX Final fix webrevs for 1.6 and 1.5: http://prt-web.sfbay.sun.com/net/prt-archiver.sfbay/data/archived_workspaces/1.6/update2/baseline/2007/20070104202509.jrose.mustang-6461827/workspace/webrevs/webrev-2007.01.04/index.html http://analemma.sfbay.sun.com/net/prt-archiver.sfbay/data/archived_workspaces/1.5/tiger_update12_baseline/2006/20061218164652.jrose.tiger-6461827/workspace/webrevs/webrev-2006.12.18/index.html

EVALUATION It's a badly designed benchmark, so the numbers must not be taken quantitatively; they provide an order-of-magnitude indication of code quality only. As is the case with many microbenchmarks, the warm-up phase OSRs, and the timing phase runs non-OSR code. In this case, the timing numbers are disturbed by the non-OSR compilation that occurs. A quick fix to this (though probably not a stable one) is as follows: --- Main.java~ 2006-08-14 16:06:57.000000000 -0700 +++ Main.java 2007-01-16 16:50:44.033302000 -0800 @@ -103,9 +103,11 @@ private void run(){ runIsAssignableFrom("isAssignable warm-up time"); + runIsAssignableFrom("isAssignable warm-up time #2"); runIsAssignableFrom("isAssignable run time"); runSetField("setField warm-up time"); + runSetField("setField warm-up time #2"); runSetField("setField run time"); } When creating a benchmark, always try it out with -XX:+PrintCompilation and recode until there are no compilations occurring during the timing phase. Also run -Xbatch, of course. The #2 warm-up calls run into the -Xbatch execution barrier and wait for the non-OSR compilation of each test to finish before starting the timing run. Also, result of the call to isAssignableFrom is not tested or otherwise used. That potentially allows more optimization (or cheating!) than the benchmark attempts to test. In this case, the open-coded call to isAssignableFrom can disappear completely. If this happens, the loop containing it also disappears, except for a single iteration to perform null checks on the operands. So any time reported has nothing to do with the speed of isAssignableFrom.

SUGGESTED FIX This should give a simple speed improvement on XML or Beans benchmarks. It makes all sorts of reflection operations go much faster, by open-coding the JNI call which performs reflective type checking. The call can be completed in less than ten memory references if open-coded. There is test version here, if you want to try it: /net/prt-web.sfbay/prt-workspaces/20070102171145.jrose.mustang-6461827/workspace (temporary) /net/foundation.sfbay/export/jrose/builds/mustang-6461827/prt-build (not so temporary)

EVALUATION This is a point fix which can be backported from Dolphin. I am targeting it for 5.0u12 (on the sub-CR) and 6u2. See suggested fix for a webrev. It is co-packaged with another point fix in the same code, for 6297094.

SUGGESTED FIX Here is a suggested point-fix for 1.5: http://javaweb.sfbay/~jrose/webrev/tiger-6461827/ Fixes in later releases are almost identical.

EVALUATION I have the bug reproduced and it is not pretty, we need to consider intrinsifying the IsAssignableFrom method, needs further evaluation to backport fix from 7.0.Commiting to mustang update, as this fix is not appropriate for 6 RC.