Due to JEP 260 (Encapsulate Most Internal APIs), aspect weaving on
Java 16 now requires '--add-opens java.base/java.lang=ALL-UNNAMED' on
the command line. Otherwise there will be illegal access exceptions for
some internal API calls AspectJ needs, most prominently when trying to
define classes in other packages or modules.

This had to be done on several levels:

  - Maven Surefire: running tests in a JVM directly forked by Surefire.
    In order to make this backwards compatible, I added two profiles
    with JDK-level-dependent auto-activation, one 8-15 and one 16+. In
    the latter a property containing the JVM parameter is defined, in
    the former it is empty, i.e. the JVM is started without the
    parameter. In Java 8 the parameter did not even exist, in Java 9+ we
    could use it, but we need to test how users use AspectJ.

  - RunSpec: Whenever an XML test is declared to use '<run>', we need to
    determine the current JVM version and again dynamically add the
    parameter when forking the target JVM.

  - AntSpec: Whenever an XML test is declared to use '<ant>', we need to
    determine the current JVM version dynamically add two properties
    usable from within Ant scripts: 'aj.addOpensKey' and
    'aj.addOpensValue'. Unfortunately, Ant needs to use two '<argLine>'
    parameters, because the two parts of the option are separated by a
    space character.

  - Ant scripts: When triggered by an AntSpec, each Ant target using LTW
    needs to manually set
      <jvmarg value="${aj.addOpensKey}"/>
      <jvmarg value="${aj.addOpensValue}"/>
    for each '<java>' task. It was quite tedious to find all(?) of them.

TODO: In the AspectJ 1.9.7 release notes we need to document that this
parameter is now needed for LTW.

Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>

Replace directory separator '/' and surrogate path separator ',' by
platform-specific separators File.separatorChar and
File.pathSeparatorChar, respectively. Also make sure that replacement
occurs during write access, not read access. This was handled
differently in both sibling classes.

I am not sure if that helps to fix any Linux CI tests, but it is worth a
try.

Signed-off-by: Alexander Kriegisch <Alexander@Kriegisch.name>

Signed-off-by: Lars Grefer <eclipse@larsgrefer.de>

junit test suite based on an ajcTests.xml format test specification.