Proposal: Allow illegal reflective access by default in JDK 9

[mark.reinhold at oracle.com]

Over time, as we've gotten closer and closer to the JDK 9 GA date, more
and more developers have begun paying attention to the actual changes
in this release.  The strong encapsulation of JDK-internal APIs has, in
particular, triggered many worried expressions of concern that code that
works on JDK 8 today will not work on JDK 9 tomorrow, yet no advance
warning of this change was given in JDK 8.

To help the entire ecosystem migrate to the modular Java platform at a
more relaxed pace I hereby propose to allow illegal reflective access
from code on the class path by default in JDK 9, and to disallow it in
a future release.

In short, the existing "big kill switch" of the `--permit-illegal-access`
option [1] will become the default behavior of the JDK 9 run-time system,
though without as many warnings.  The current behavior of JDK 9, in which
illegal reflective-access operations from code on the class path are not
permitted, will become the default in a future release.  Nothing will
change at compile time.

In detail, the recently-introduced `--permit-illegal-access` option will
be replaced by a more-general option, `--illegal-access`.  This option
will take a single keyword parameter, as follows:

  `--illegal-access=permit`

    This will be the default mode for JDK 9.  It opens every package in
    every explicit module to code in all unnamed modules, i.e., code on
    the class path, just as `--permit-illegal-access` does today.

    The first illegal reflective-access operation causes a warning to be
    issued, as with `--permit-illegal-access`, but no warnings are issued
    after that point.  This single warning will describe how to enable
    further warnings.

  `--illegal-access=warn`

    This causes a warning message to be issued for each illegal
    reflective-access operation.  This is equivalent to the current
    `--permit-illegal-access` option.

  `--illegal-access=debug`

    This causes both a warning message and a stack trace to be shown
    for each illegal reflective-access operation.  This is equivalent
    to combining today's `--permit-illegal-access` option with
    `-Dsun.reflect.debugModuleAccessChecks`.

  `--illegal-access=deny`

    This disables all illegal reflective-access operations except for
    those enabled by other command-line options, such as `--add-opens`.
    This will become the default mode in a future release.

Notes:

  - The proposed default mode enables the run-time system to issue a
    warning message, possibly at some time long after startup, without
    having been explicitly requested to do so.  This may be a surprise
    in production environments, since it's extremely unusual for the
    run-time system to issue any warning messages at all.  If the default
    mode permits illegal reflective access, however, then it's essential
    to make that known so that people aren't surprised when this is no
    longer the default mode in a future release.

  - Warning messages in any mode can be avoided, as before, by the
    judicious use of the `--add-exports` and `--add-opens` options.

  - This proposal will, if adopted, require adjustments to JEP 260,
    "Encapsulate Most Internal APIs" [2].  APIs that are internal to the
    JDK will still be strongly encapsulated from the standpoint of code
    in modules, whether those modules are automatic or explicit, but they
    will not appear to be encapsulated at run time from the standpoint of
    code on the class path.

  - When `deny` becomes the default mode then I expect `permit` to remain
    supported for at least one release, so that developers can continue
    to migrate their code.  The `permit`, `warn`, and `debug` modes will,
    over time, be removed, as will the `--illegal-access` option itself.
    (For launch-script compatibility the unsupported modes will most
    likely just be ignored, after issuing a warning to that effect.)

  - This change will not magically solve every JDK 9 adoption problem.
    The concrete types of the built-in class loaders are still different,
    `rt.jar` is still gone, the layout of a system image is still not the
    same, and the version string still has a new format.

Comments?

- Mark


[1] http://mail.openjdk.java.net/pipermail/jigsaw-dev/2017-March/011763.html
[2] http://openjdk.java.net/jeps/260