[foreign-abi] RFR: 8253823: Investigate ways to make handoff-like operation more explicit [v6]

[Paul Sandoz]

On Wed, 30 Sep 2020 18:11:08 GMT, Maurizio Cimadamore <mcimadamore at openjdk.org> wrote:

>> The API currently doesn't do a good job at differentiating between simple wither-like methods (e.g. withAccessModes)
>> and deep side-effecting operations (e.g. withOwnerThread).
>> This patch explores a way to make handoff more of a first-class concept of the API - you can take apart an existing
>> segment and reconstruct it with a given `HandoffTransform` (a recipe which specifies what properties should be
>> associated with the reconstructed segment). Turns out that `handoff` and `HandoffTransform` is the ideal place where to
>> put several of the *advanced* operations we used to have exposed directly by the memory segment API
>> (`withCleanupAction`, `registerCleaner`). We can also add na overload for `handoff` which takes a native scope and
>> bounds the segment to it.  By reorganizing the API in this way, it is now clear that memory segment feature _two_
>> terminal operations: `close` and `handoff`. We can document this in the API in a more centralized fashion, without
>> having to repeat the same text across seemingly unrelated methods. For clients, it is also clearer that doing an
>> `handoff` is a big reconstruction, which is different from a simple wither.  I've also somewhat simplified the access
>> mode story; there's now a single HANDOFF mode which covers all types of handoff - this was clearly needed, since the
>> access modes we had before were very fine-grained, and it was very painful for client to prevent against *all* kinds of
>> terminal operations (for instance, I realized that the current impl has no access modes check for
>> `withCleanupAction`).  Impl note: for now the HandoffTransform implementation is mutable and not thread-safe - the
>> javadoc calls this out. I'm open to make this more immutable/value oriented if this is desirable.  Note: this PR is
>> against the `foreign-abi` so that it shows how `NativeScope` is affected by the changes. If the changes brought in this
>> patch looks good, it would be better to push the memory access changes in the `memory-access` branch in a separate PR,
>> and then follow up with the remaining `NativeScope` changes here.  Javadoc  link:
>> http://cr.openjdk.java.net/~mcimadamore/panama/segment-rebuilder-javadoc_v2/javadoc/jdk/incubator/foreign/package-summary.html
>
> Maurizio Cimadamore has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Revert spurious test changes

This is indeed simpler (esp. also for implementation of the clean up actions).

I am not entirely convinced with clean up action being only associated with unchecked segment creation. For example the
[BLIS object API](https://github.com/flame/blis/blob/master/docs/BLISObjectAPI.md) may be used to create
(`bli_obj_create`) and free (`bli_obj_free`) the allocated region of memory representing a matrix with elements of a
certain type. The BLIS `obj_t` would be allocated separately as a segment and it may be useful to attach a cleanup
action to that segment to free the region using the native method call.

src/jdk.incubator.foreign/share/classes/jdk/internal/foreign/MemoryScope.java line 165:

> 163:         try {
> 164:             justClose();
> 165:             if (scopeCleanable != null) {

`scopeCleanable` is checked above and if non-`null` throws.

src/jdk.incubator.foreign/share/classes/jdk/internal/foreign/MemoryScope.java line 286:

> 284:     static class ScopeCleanable extends PhantomCleanable<MemoryScope> {
> 285:         final Cleaner cleaner;
> 286:         Runnable cleanupAction;

Mark field and class as `final`

-------------

Marked as reviewed by psandoz (Committer).

PR: https://git.openjdk.java.net/panama-foreign/pull/361