There seems to be an issue with the Integer.bitCount(int) intrinsic on AArch64. Somehow, the input argument into the bitCount intrinsic ends up being cut which is propagated to the caller. Maybe some register corruption.

To reproduce:

$ java -Xbatch -XX:CompileOnly=Reduced::test Reduced.java
 
Output:
Exception in thread "main" java.lang.RuntimeException: Wrong result: 263219110
at Reduced.main(Reduced.java:10)

No exception with -Xint or on x64.

====================
Original report:

The attached Test.java file prints different output when compiled vs. when interpreted, on AArch64 only.

$ java -Xint Test.java
2166766936661037687
-5936201357435948857
-8914451713432026907
-3053072614657754545
-5400070531972085357
3626172417624192429
3415121120042614434
2717387767676077548
-2066234294734040154
4482272172956477350

$ java -Xbatch Test.java
2166766936661037687
-5936201357435948857
-8914451713432026907
-3053072614657754545
-5400070531972085357
3626172417624192429
3415121120042614434
2717387767676077548
933614502
263219110

The last two lines differ. When running with -Xcomp, all lines of the output differ. The output on AMD64 is always the same as for the -Xint run, so I think this must be an AArch64-specific miscompilation.

Observed with a 25-ea+15-LTS-1670 build.