RFR: 8170595: Optimize Class.isAnonymousClass

[Claes Redestad]

Hi,

On 2016-12-01 22:25, Claes Redestad wrote:
> On 12/01/2016 10:21 AM, Mandy Chung wrote:
>>> On Dec 1, 2016, at 9:52 AM, Claes Redestad 
>>> <claes.redestad at oracle.com> wrote:
>>>
>>> Hi,
>>>
>>> good suggestion, this tidies up a bit while not affecting score:
>>>
>>> http://cr.openjdk.java.net/~redestad/8170595/webrev.02
>> I like this better.  It may be useful to add a private isTopLevel 
>> Class() for  getSimpleBinaryName to call that will benefit 
>> getSimpleName and isMemberClass?
>
> Good idea, but this seems like an independent optimization we should 
> do as a separate follow-up, no?

Found that both isMemberClass and isLocalClass can be optimized in a 
vein very similar to isAnonymousClass. isTopLevelClass helps
getSimpleBinaryName a bit, and makes sense to implement for 
completeness. After deeper testing it's clear that one of the
costlier operations in this code is getDeclaringClass0, which may spend 
plenty of time (~30ns/op for String.class) in the VM walking the
list of inner classes to find the enclosing class - if any.

A very small improvement can also be attained by restructuring code so 
that we don't instantiate the EnclosingMethodInfo
unless it's needed, but keep the checking to keep semantically identical 
behavior:

http://cr.openjdk.java.net/~redestad/8170595/webrev.03/

This brings significant improvements to some variants:

Baseline:
Benchmark                       Mode  Cnt    Score    Error  Units
Clazz.isAnonymousClass_Anon     avgt   25  211.010 ± 22.974  ns/op
Clazz.isAnonymousClass_Regular  avgt   25  132.198 ± 23.817  ns/op
Clazz.isMemberClass_Anon        avgt   25  336.149 ± 43.215  ns/op
Clazz.isMemberClass_Regular     avgt   25   92.502 ±  9.804  ns/op
Clazz.isLocalClass_Anon         avgt   25  326.158 ± 34.051  ns/op
Clazz.isLocalClass_Regular      avgt   25   32.086 ±  2.750  ns/op

Patch:
Benchmark                       Mode  Cnt    Score    Error  Units
Clazz.isAnonymousClass_Anon     avgt   25  174.526 ± 16.292  ns/op
Clazz.isAnonymousClass_Regular  avgt   25   33.136 ±  2.636  ns/op
Clazz.isMemberClass_Anon        avgt   25  115.029 ±  9.464  ns/op
Clazz.isMemberClass_Regular     avgt   25   85.706 ±  3.984  ns/op
Clazz.isLocalClass_Anon         avgt   25  177.411 ± 22.927  ns/op
Clazz.isLocalClass_Regular      avgt   25   31.940 ±  2.346  ns/op

Since the changes are all similar but still not too expansive, I'm 
leaning towards just keeping this
change in one bug.

Thanks!