Bad XMM performance on java.lang.MathBench.sqrtDouble

We have new j.l.Math JMH here: https://github.com/openjdk/jmh-jdk-microbenchmarks/blob/master/micros-jdk11/src/main/java/org/openjdk/bench/java/lang/MathBench.java

We discovered a funny "regression" with plain sqrt.

   vsqrtsd 0x50(%r10),%xmm0,%xmm0

17-b12:
Benchmark Mode Cnt Score Error Units
MathBench.sqrtDouble thrpt 12 94635.738 ± 47.086 ops/ms
StrictMathBench.sqrtDouble thrpt 12 94652.203 ± 10.816 ops/ms

I added another case:

    @Benchmark
    public double sqrtDoubleAfterAdd() {
        return Math.sqrt(double81 + double4Dot1);
    }

So the double add result would be at hand into the sqrt, now we get

  1.22% ││ 0x00007ff2d93e85d7: vmovsd 0x50(%r10),%xmm0
  1.51% ││ 0x00007ff2d93e85dd: vaddsd 0x40(%r10),%xmm0,%xmm0
  1.93% ││ 0x00007ff2d93e85e3: vsqrtsd %xmm0,%xmm0,%xmm0 ;*invokestatic sqrt {reexecute=0 rethrow=0 return_oop=0}

and the performance goes UP!

Benchmark Mode Cnt Score Error Units
MathBench.sqrtDouble thrpt 4 94649.452 ± 40.929 ops/ms
MathBench.sqrtDoubleAfterAdd thrpt 4 210876.199 ± 163.751 ops/ms