While this issue manifests with G1, I believe this is an optimizer glitch that should be eliminated by optimizers themselves. In some variants, we end up generating:

    0x00007f82e135007a: movsbl 0x310(%r15),%r9d
    0x00007f82e1350082: test %r9d,%r9d
╭ 0x00007f82e1350085: jne 0x00007f82e1350087
 ↘ 0x00007f82e1350087: mov %rax,%r10

...and then also kill %r9.

Reproducer:
 http://cr.openjdk.java.net/~shade/8145612/G1UselessBranch.java
 http://cr.openjdk.java.net/~shade/8145612/benchmarks.jar

perfasm for G1:
 http://cr.openjdk.java.net/~shade/8145612/g1.perfasm

perfasm for Parallel:
 http://cr.openjdk.java.net/~shade/8145612/parallel.perfasm