RFR: 8261731: shallow copy the internal buffer of a scalar-replaced java.lang.String object

[Claes Redestad]

On Wed, 24 Mar 2021 00:06:45 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> @navyxliu I think you want to fix a lot of issues once. You need to split problems you see.
>> 
>> I did few experiments and the one issue which could be fixed first is scalarizing of array created by `Arrays.copyOfRange()` used by `String::substring()`.
>> 
>> I see that current JDK easy scalarize new array and arraycopy():
>>      private static int test4(byte[] src) {
>>          byte[] dst = new byte[6];
>>          System.arraycopy(src, 1, dst, 0, 6);
>>          return dst[2] + dst[5];
>>      }
>> But it did not scalarize:
>>      private static int test5(byte[] src) {
>>          byte[] dst = Arrays.copyOfRange(src, 1, 7);
>>          return dst[2] + dst[5];
>>      }
>> Even so code in `Arrays.copyOfRange()` is the same as in `test4()`:
>> https://github.com/openjdk/jdk/blob/master/src/java.base/share/classes/java/util/Arrays.java#L3817
>> 
>> Arrays does not escaped but it is not eliminated because Loads are not redirected into `dst` array as in `test4()` case.
>> 
>> I think this should be fixed first.
>
> Hmm, actually it was easy solved with `-XX:DisableIntrinsic=_min` :(
> 
> This intrinsic was implemented long ago. the code is used for `arraycopy` optimization. But for some reason it does not work here.
> 
> I think this should be investigated first-first.  It would be interesting to see performance effects if we disable this intrinsic at all.

You seem to have had a run-in with these intrinsics causing issues before: https://bugs.openjdk.java.net/browse/JDK-8039104 - is this a similar issue?

On a trivial micro adding together a few Math.min values the intrinsic seem to have no effect on my x64 system, down to generating the exact same assembly.

-------------

PR: https://git.openjdk.java.net/jdk/pull/2570