step1: Try to compile simple class:

import java.lang.reflect.Method;

public class test {
    private static final int ITERATIONS = 100000;

    public static void main(String[] args) {
        new test().test();
    }

    private void test() {
        for (int i = 0; i < ITERATIONS; i++) {
            testMethod();
        }
    }

    private void testMethod() {
        Helper helper = new Helper("" + System.currentTimeMillis());
        if (helper.string.length() < 1) {
            throw new AssertionError("impossible");
        }
    }

    private class Helper {
        public String string;

        public Helper(String s) {
            this.string = s;
        }
    }
}

step2:
run this class with following vm options:
1) -XX:+TieredCompilation -XX:+PrintCompilation -XX:+PrintEscapeAnalysis -XX:+Verbose
2) -XX:-TieredCompilation -XX:+PrintCompilation -XX:+PrintEscapeAnalysis -XX:+Verbose
3) -Xcomp -XX:+TieredCompilation -XX:+PrintCompilation -XX:+PrintEscapeAnalysis -XX:+Verbose
4) -Xcomp -XX:-TieredCompilation -XX:+PrintCompilation -XX:+PrintEscapeAnalysis -XX:+Verbose

A console log for testMethod escape analysis will have different results for first 3 launch and another for launch number 4.

A first string is the most interesting one:
for vmoptions set 1-2-3:
JavaObject NoEscape(NoEscape) [ 1171F 993F [ 37 42 ]] 25 Allocate === 5 6 7 8 1 ( 23 21 22 1 10 1 ) [[ 26 27 28 35 36 37 ]] ... test::testMethod ... bci:0

and for vmoptions set 4:
JavaObject ArgEscape(GlobalEscape) [ 192F [ 37 42 ]] 25 Allocate === 5 6 7 8 1 ( 23 21 22 1 10 1 ) [[ 26 27 28 35 36 37 ]] ... test::testMethod ... bci:0


So, a helper object is considered to be escaped and can be stack allocated for first 3 cases and considered to be ArgEscape(GlobalEscape) for 4th case, which seems to be wrong