OptimizeStringConcat fails on pre-sized StringBuilder shapes

A sample benchmark:

    @Benchmark
    @CompilerControl(CompilerControl.Mode.DONT_INLINE)
    public String explicitSized() throws Exception {
        String time = String.valueOf(System.nanoTime());
        String thread = Thread.currentThread().getName();
        String m = method;
        return new StringBuilder(63 + time.length() + thread.length() + m.length())
                .append("[")
                .append(time)
                .append("] ")
                .append(thread)
                .append(":")
                .append("Calling an application method \"")
                .append(m)
                .append("\" without fear and prejudice.")
                .toString();
    }

This benchmark regresses severely with Compact Strings. Turning OptimizeStringConcat off does not degrade performance further, which means it is broken with Compact Strings to begin with. The disassembly confirms it. Also see the (broken) workaround patch below:

# JDK 9 Baseline
LogLineBench.explicitSized: 83.267 ± 3.021 ns/op
LogLineBench.explicitSized:·gc.alloc.rate.norm: 384.000 ± 0.001 B/op

# JDK 9 Baseline, -XX:-OptimizeStringConcat
LogLineBench.explicitSized: 190.090 ± 5.893 ns/op
LogLineBench.explicitSized:·gc.alloc.rate.norm: 1296.000 ± 0.001 B/op

# JDK 9 Compact Strings
LogLineBench.explicitSized: 148.536 ± 5.032 ns/op
LogLineBench.explicitSized:·gc.alloc.rate.norm: 704.000 ± 0.001 B/op

# JDK 9 Compact Strings, -XX:-OptimizeStringConcat
LogLineBench.explicitSized: 143.668 ± 7.654 ns/op
LogLineBench.explicitSized:·gc.alloc.rate.norm: 704.000 ± 0.001 B/op