7.6.1 Completion and Finalization

Discussion: Don't confuse the run-time concept of completion with the compile-time concept of completion defined in 3.11.1. 

Reason: {AI95-00162-01} {AI95-00416-01} Expressions and statements are masters so that objects created by subprogram calls (in aggregates, allocators for anonymous access-to-object types, and so on) are finalized and have their tasks awaited before the expressions or statements are left. Note that expressions like the condition of an if_statement are masters, because they are not enclosed by a simple_statement. Similarly, a function_call which is renamed is a master, as it is not in a simple_statement.

{AI95-00416-01} We have to include function_calls in the contexts that do not cause masters to occur so that expressions contained in a function_call (that is not part of an expression or simple_statement) do not individually become masters. We certainly do not want the parameter expressions of a function_call to be separate masters, as they would then be finalized before the function is called. 

Ramification: {AI95-00416-01} The fact that a function_call is a master does not change the accessibility of the return object denoted by the function_call; that depends on the use of the function_call. The function_call is the master of any short-lived entities (such as aggregates used as parameters of types with task or controlled parts). 

Ramification: As explained in 3.10.2, the set of objects with the same accessibility level as that of the master includes objects declared immediately within the master, objects declared in nested packages, objects created by allocators (if the ultimate ancestor access type is declared in one of those places) and subcomponents of all of these things. If an object was already finalized by Unchecked_Deallocation, then it is not finalized again when the master is left.

Note that any object whose accessibility level is deeper than that of the master would no longer exist; those objects would have been finalized by some inner master. Thus, after leaving a master, the only objects yet to be finalized are those whose accessibility level is less deep than that of the master.

To be honest: Subcomponents of objects due to be finalized are not finalized by the finalization of the master; they are finalized by the finalization of the containing object. 

Reason: We need to finalize subcomponents of objects even if the containing object is not going to get finalized because it was not fully initialized. But if the containing object is finalized, we don't want to require repeated finalization of the subcomponents, as might normally be implied by the recursion in finalization of a master and the recursion in finalization of an object. 

To be honest: Formally, completion and leaving refer to executions of constructs or entities. However, the standard sometimes (informally) refers to the constructs or entities whose executions are being completed. Thus, for example, “the subprogram call or task is complete” really means “the execution of the subprogram call or task is complete.” 

Reason: This allows the finalization of a component with an access discriminant to refer to other components of the enclosing object prior to their being finalized. 

Reason: Note that we talk about the type of the allocator here. There may be access values of a (general) access type pointing at objects created by allocators for some other type; these are not finalized at this point.

The freezing point of the ultimate ancestor access type is chosen because before that point, pool elements cannot be created, and after that point, access values designating (parts of) the pool elements can be created. This is also the point after which the pool object cannot have been declared. We don't want to finalize the pool elements until after anything finalizing objects that contain access values designating them. Nor do we want to finalize pool elements after finalizing the pool object itself. 

Ramification: Finalization of allocated objects is done according to the (ultimate ancestor) allocator type, not according to the storage pool in which they are allocated. Pool finalization might reclaim storage (see 13.11, “Storage Management”), but has nothing (directly) to do with finalization of the pool elements.

Note that finalization is done only for objects that still exist; if an instance of Unchecked_Deallocation has already gotten rid of a given pool element, that pool element will not be finalized when the master is left.

Note that a deferred constant declaration does not create the constant; the full constant declaration creates it. Therefore, the order of finalization depends on where the full constant declaration occurs, not the deferred constant declaration.

An imported object is not created by its declaration. It is neither initialized nor finalized. 

Implementation Note: An implementation has to ensure that the storage for an object is not reclaimed when references to the object are still possible (unless, of course, the user explicitly requests reclamation via an instance of Unchecked_Deallocation). This implies, in general, that objects cannot be deallocated one by one as they are finalized; a subsequent finalization might reference an object that has been finalized, and that object had better be in its (well-defined) finalized state. 

This paragraph was deleted.{AI95-00162-01}

Reason: {AI95-00162-01} This effectively imports all of the special rules for the accessibility level of renames, allocators, and so on, and applies them to determine where objects created in them are finalized. For instance, the master of a rename of a subprogram is that of the renamed subprogram. 

Ramification: It is not a bounded error for Initialize to propagate an exception. If Initialize propagates an exception, then no further calls on Initialize are performed, and those components that have already been initialized (either explicitly or by default) are finalized in the usual way.

{8652/0023} {AI95-00169-01} It also is not a bounded error for an explicit call to Finalize or Adjust to propagate an exception. We do not want implementations to have to treat explicit calls to these routines specially. 

Reason: {8652/0024} {AI95-00193-01} {AI95-00256-01} In the case of assignments that are part of initialization, there is no need to complete all adjustments if one propagates an exception, as the object will immediately be finalized. So long as a subcomponent is not going to be finalized, it need not be adjusted, even if it is initialized as part of an enclosing composite assignment operation for which some adjustments are performed. However, there is no harm in an implementation making additional Adjust calls (as long as any additional components that are adjusted are also finalized), so we allow the implementation flexibility here. On the other hand, for an assignment_statement, it is important that all adjustments be performed, even if one fails, because all controlled subcomponents are going to be finalized. Other kinds of assignment are more like initialization than assignment_statements, so we include them as well in the permission. 

Ramification: {8652/0024} {AI95-00193-01} Even if an Adjust invoked as part of the initialization of a controlled object propagates an exception, objects whose initialization (including any Adjust or Initialize calls) successfully completed will be finalized. The permission above only applies to objects whose Adjust failed. Objects for which Adjust was never even invoked must not be finalized. 

Discussion: {8652/0104} {AI95-00179-01} The standard does not specify if storage is recovered in this case. If storage is not recovered (and the object continues to exist), Finalize may be called on the object again (when the allocator's master is finalized). 

Ramification: For example, upon leaving a block_statement due to a goto_statement, the Program_Error would be raised at the point of the target statement denoted by the label, or else in some more dynamically nested place, but not so nested as to allow an exception_handler that has visibility upon the finalized object to handle it. For example,

The goto_statement will first cause Finalize(Y) to be called. Suppose that Finalize(Y) propagates an exception. Program_Error will be raised after leaving Inner_Block_Statement, but before leaving Main. Thus, handler number 1 cannot handle this Program_Error; it will be handled either by handler number 2 or handler number 3. If it is handled by handler number 2, then Finalize(Z) will be done before executing the handler. If it is handled by handler number 3, then Finalize(Z) and Finalize(X) will both be done before executing the handler. 

Ramification: If, in the above example, the goto_statement were replaced by a raise_statement, then the Program_Error would be handled by handler number 2, and Finalize(Z) would be done before executing the handler. 

Reason: We considered treating this case in the same way as the others, but that would render certain exception_handlers useless. For example, suppose the only exception_handler is one for others in the main subprogram. If some deeply nested call raises an exception, causing some Finalize operation to be called, which then raises an exception, then normal execution “would have continued” at the beginning of the exception_handler. Raising Program_Error at that point would cause that handler's code to be skipped. One would need two nested exception_handlers to be sure of catching such cases!

On the other hand, the exception_handler for a given master should not be allowed to handle exceptions raised during finalization of that master. 

Ramification: This case includes an asynchronous transfer of control. 

To be honest: {Program_Error (raised by failure of run-time check)} This violates the general principle that it is always possible for a bounded error to raise Program_Error (see 1.1.5, “Classification of Errors”). 

Discussion: We considered defining a pragma that would apply to a controlled type that would suppress Finalize operations for library-level objects of the type upon partition termination. This would be useful for types whose finalization actions consist of simply reclaiming global heap storage, when this is already provided automatically by the environment upon program termination. 

Discussion: Or equivalently, a Finalize procedure should be “idempotent”; applying it twice to the same object should be equivalent to applying it once. 

Reason: A user-written Finalize procedure should be idempotent since it can be called explicitly by a client (at least if the type is "visibly" controlled). Also, Finalize is used implicitly as part of the assignment_statement if the type is nonlimited, and an abort is permitted to disrupt an assignment_statement between finalizing the left-hand side and assigning the new value to it (an abort is not permitted to disrupt an assignment operation between copying in the new value and adjusting it). 

Discussion: {AI95-00287-01} Either Initialize or Adjust, but not both, is applied to (almost) every controlled object when it is created: Initialize is done when no initial value is assigned to the object, whereas Adjust is done as part of assigning the initial value. The one exception is the object initialized by an aggregate (both the anonymous object created for an aggregate, or an object initialized by an aggregate that is built-in-place); Initialize is not applied to the aggregate as a whole, nor is the value of the aggregate or object adjusted.

{assignment operation (list of uses)} All of the following use the assignment operation, and thus perform value adjustment: 

The following also use the assignment operation, but adjustment never does anything interesting in these cases: 

This paragraph was deleted.{AI95-00287-01}

Finalization of the parts of a protected object are not done as protected actions. It is possible (in pathological cases) to create tasks during finalization that access these parts in parallel with the finalization itself. This is an erroneous use of shared variables.

Implementation Note: One implementation technique for finalization is to chain the controlled objects together on a per-task list. When leaving a master, the list can be walked up to a marked place. The links needed to implement the list can be declared (privately) in types Controlled and Limited_Controlled, so they will be inherited by all controlled types.

Another implementation technique, which we refer to as the “PC-map” approach essentially implies inserting exception handlers at various places, and finalizing objects based on where the exception was raised.

{PC-map approach to finalization} {program-counter-map approach to finalization} The PC-map approach is for the compiler/linker to create a map of code addresses; when an exception is raised, or abort occurs, the map can be consulted to see where the task was executing, and what finalization needs to be performed. This approach was given in the Ada 83 Rationale as a possible implementation strategy for exception handling — the map is consulted to determine which exception handler applies.

If the PC-map approach is used, the implementation must take care in the case of arrays. The generated code will generally contain a loop to initialize an array. If an exception is raised part way through the array, the components that have been initialized must be finalized, and the others must not be finalized.

It is our intention that both of these implementation methods should be possible. 

Finalization depends on the concepts of completion and leaving, and on the concept of a master. Therefore, we have moved the definitions of these concepts here, from where they used to be in Section 9. These concepts also needed to be generalized somewhat. Task waiting is closely related to user-defined finalization; the rules here refer to the task-waiting rules of Section 9. 

{8652/0021} {AI95-00182-01} Corrigendum: Fixed the wording to say that anonymous objects aren't finalized until the object can't be used anymore.

{8652/0023} {AI95-00169-01} Corrigendum: Added wording to clarify what happens when Adjust or Finalize raises an exception; some cases had been omitted.

{8652/0024} {AI95-00193-01} {AI95-00256-01} Corrigendum: Stated that if Adjust raises an exception during initialization, nothing further is required. This is corrected in Ada 2005 to include all kinds of assignment other than assignment_statements.

{AI95-00162-01} {AI95-00416-01} Revised the definition of master to include expressions and statements, in order to cleanly define what happens for tasks and controlled objects created as part of a subprogram call. Having done that, all of the special wording to cover those cases is eliminated (at least until the Ada comments start rolling in).

{AI95-00280-01} We define finalization of the collection here, so as to be able to conveniently refer to it in other rules (especially in 4.8, “Allocators”).

{AI95-00416-01} Clarified that a coextension is finalized at the same time as the outer object. (This was intended for Ada 95, but since the concept did not have a name, it was overlooked.)