Most Unsafe.get*() access shapes are losing vs. the plain Java accesses

If you run a simple benchmark like this:

    @Benchmark
    public int stream_unsafe_char() {
        int s = 0;
        for (int c = 0; c < size; c++) {
            s += U.getChar(charArr, CHAR_ARR_OFFSET + CHAR_ARR_SCALE * c);
        }
        return s;
    }

...then you will notice that generated code has a stray "movslq":

 0x00007ffad8741180: movslq %r8d,%r9 <--- unnecessary sign extension
 0x00007ffad8741183: movzwl 0x10(%r10,%r9,2)
 0x00007ffad8741189: add %ecx,%edx
 0x00007ffad874118b: inc %r8d
 0x00007ffad874118e: cmp %r11d,%r8d
 0x00007ffad8741191: jl 0x00007ffad8741180

Indeed, the baseline test that does the plain Java access:

    @Benchmark
    public int plain() {
        int s = 0;
        for (int c = 0; c < size; c++) {
            s += charArr[c];
        }
        return s;
    }

...does without sign extension:

 0x00007fc3d0f84ea0: movzwl 0x10(%r11,%r8,2),%ecx
 0x00007fc3d0f84ea6: add %ecx,%edx
 0x00007fc3d0f84ea8: inc %r8d
 0x00007fc3d0f84eab: cmp %r10d,%r8d
 0x00007fc3d0f84eae: jl 0x00007fc3d0f84ea0

This conversion costs some cycles on my 1x4x2 4 GHz Haswell, running with Linux x86_64, 8u40 EA:

Benchmark (size) Mode Cnt Score Error Units
UnsafeMovslq.plain 1000 avgt 5 343.719 ± 0.694 ns/op
UnsafeMovslq.unsafe 1000 avgt 5 362.350 ± 5.302 ns/op