Vector::rearrange(Shuffle) allocations

[Vladimir Ivanov]

Good catch, Tomasz! That's a bug in the implementation of rearrange: 
default implementation (lambda) captures the vector box and it keeps it 
from being eliminated.

The following patch should fix the problem:
 
http://cr.openjdk.java.net/~vlivanov/panama/vector/rearrange/webrev.00/
Best regards,
Vladimir Ivanov

On 14/01/2019 09:56, Tomasz Kowalczewski wrote:
> I have a question about vector rearranges. My recent observation is
> that using IntVector.rearrange causes a lot of allocations (in a JMH
> microbenchmark). Is this expected? I wrote a few simple (but not
> trivial) algorithms and they all did not box vectors into objects and
> all tests were able to run using EpsilonGC. The intrinsic works and
> vpermd instruction is emitted.
> 
> There are 64 bytes allocated per benchmark iteration in the following
> code (the addAll is used so that vector does not escape the method):
> 
> @Benchmark
> public int rearrange() {
>      IntVector input = SPECIES_I256.fromArray(inputArray, 0);
>      IntVector rearrangedVector = input.rearrange(SHUFFLE_1);
> 
>      return rearrangedVector.addAll();
> }
> 
> There is no allocation in following code:
> 
> @Benchmark
> public int noRearrange() {
>      IntVector input = SPECIES_I256.fromArray(inputArray, 0);
> 
>      return input.addAll();
> }
> 
> Output from jmap -histo for first case:
> 
>   num     #instances         #bytes  class name (module)
> -------------------------------------------------------
>     1:       4122313      198969024  [I (java.base@)
>     2:       4119794       65916704  jdk.incubator.vector.Int256Vector
> (jdk.incubator.vector@)
>     3:          8913         808512  [B (java.base@)
>     4:          1982         222408  java.lang.Class (java.base@)
>     5:          8638         207312  java.lang.String (java.base@)
> 
> Any hints will be much appreciated.
>