JMH microbenchmarks vm.compiler.signum.* produce problematic scores because fmov is used to transmit data between GPRs and FPRs, the time cost is significant, same as mentioned by JDK-8251525.

Particularly, Signum._1_signumFloatTest performance can be improved much by setting -XX:+UnlockDiagnosticVMOptions -XX:+UseSignumIntrinsic, which generates facgt/ushr/bsl pairs for the unrolling code snippet of the hot for-loop of the test case. Verified on Ampere-1A and found the score of patched became 40x better than base (thrpt, ops/s). This JBS is to turn the option ON by default for all Ampere CPUs.