RFR: [6904367]: (coll) IdentityHashMap is resized before exceeding the expected maximum size

[Ivan Gerasimov]

On 08.07.2014 4:44, Martin Buchholz wrote:
>
>
>
> On Mon, Jul 7, 2014 at 9:41 AM, Martin Buchholz <martinrb at google.com 
> <mailto:martinrb at google.com>> wrote:
>
>     I'd like to offer an alternative version of this change.  webrev
>     coming soon.
>
>
> Here's the promised webrev.
> http://cr.openjdk.java.net/~martin/webrevs/openjdk9/IdentityHashMap-capacity/ 
> <http://cr.openjdk.java.net/%7Emartin/webrevs/openjdk9/IdentityHashMap-capacity/>
>
> - Fixes a typo in the javadoc.
> - removes the "threshold" field (WAT, a cache to avoid multiplying by 3?)
> - optimizes 3*x into x + x << 1

My personal preference would be x + x + x, but I thought JIT can do this 
kind of optimizations itself.
Out of curiosity I've created a microbenchmark:

Benchmark                             Mode   Samples        Score Score 
error    Units
o.s.MyBenchmark.testMethod_01_X3      avgt       200 5.900        
0.041    ns/op
o.s.MyBenchmark.testMethod_02_PPP     avgt       200 6.029        
0.035    ns/op
o.s.MyBenchmark.testMethod_03_PSH     avgt       200 5.907        
0.071    ns/op

On my machine 3*x and x + (x<<1) appear to be of the same speed (#1 and 
#3 above).
x + x + x (case #2) happens to be ~2% slower.

Given the optimization doesn't introduce any speedup, wouldn't it be 
better to use 3*x for its clarity?

Sincerely yours,
Ivan


> - adds more test assertions
> - removes the unusual 4/3 slack space for expansion on deserialization.
>
> TODO: the test should be testng-ified, I think.
>
>