Cleanup Issue DYNAMIC-EXTENT-FUNCTION

Status
passed, jun 89 X3j13
Forum
CLEANUP
Category
ADDITION
References
Scope and Extent Issue DYNAMIC-EXTENT

Problem Description

Proposal DYNAMIC-EXTENT:NEW-DECLARATION, passed at the March 89 meeting, provides a mechanism for declaring that the values of variables have only dynamic (rather than indefinite) extent. It would be useful to have similar functionality to indicate that functional bindings may have only dynamic extent. (For example, this would permit compilers to stack-allocate closures.)

Proposal (EXTEND)

Extend the DYNAMIC-EXTENT declaration to accept arguments that are lists of the form (FUNCTION <name>) where <name> is a function name, as well as symbols.

A (FUNCTION <name>) list appearing in a DYNAMIC-EXTENT declaration is used to declare that the lexically visible functional binding of <name> has dynamic extent. Except for the interpretation of <name> as the name of a function instead of the name of a variable, such a declaration otherwise has semantics that are identical to those already described in proposal DYNAMIC-EXTENT:NEW-DECLARATION.

Rationale

This permits a programmer to offer advice to an implementation about what functions may be stack-allocated for efficiency.

It may be difficult or impossible for a compiler to infer this same information statically.

Current Practice

JonL says that Lucid's compiler can stack-allocate closures, but they have no mechanism for programmers to give the compiler permission to do so.

HPCL-I has an UPWARD-CLOSURES declaration that pervasively affects all closures created within the scope of the declaration.

The Symbolics Genera compiler can often infer when functions can be implemented to have dynamic extent. Also, if a function has a SYS:DOWNWARD-FUNCTION declaration in front of its body, then the function is implemented with dynamic extent regardless of whether the compiler thinks all uses are "downward". (This declaration is rather peculiar because its scope is actually larger than the lambda expression containing the declaration; implementationally, it's the surrounding function definition.)

Cost to Implementors

No cost is forced since implementations are permitted to simply ignore the DYNAMIC-EXTENT declaration.

Cost to Users

None. This change is upward compatible.

There may be some hidden costs to debugging using this declaration (or any feature which permits the user to access dynamic extent objects without the compiler proving that they are appropriate). If the user misdeclares something and returns a pointer into the stack (or stores it in the heap), an undefined situation may result and the integrity of the Lisp storage mechanism may be compromised. Debugging these situations may be tricky, but users who have asked for this feature have indicated a willingness to deal with such costs. Nevertheless, the perils should be clearly documented and casual users should not be encouraged to use this declaration.

Cost of Non-Adoption

Some portable code would be forced to run more slowly (due to GC overhead), or to use non-portable language features.

Benefits

The cost of non-adoption is avoided.

Aesthetics

This declaration allows a fairly low level optimization to work by asking the user to provide only very high level information. The alternatives (sharpsign conditionals, some of which may lead to more bit-picky abstractions) are far less aesthetic.

Discussion

Loosemore supports DYNAMIC-EXTENT-FUNCTION:EXTEND.

This proposal does not attempt to address the issue of specifying dynamic extent for anonymous closures (which is really a special case of the more general problem of specifying dynamic extent for unnamed objects of any type). It's possible, although often awkward, to restructure the program to give the object a name and explicitly identify its extent.

One possible solution to the problem of dynamic extent for anonymous lambdas would be to clarify that a reference to a closed-over variable or function appearing lexically within a FUNCTION form is enough to cause its value to be "saved" when the FUNCTION form is executed, regardless of whether or not that reference is actually executed when the resulting function is called. Then, if all of the closed-over functions and variables referenced within a closure are declared to have dynamic extent, the closure could be assumed to have dynamic extent as well. (More precisely, its maximum extent would be the intersection of the extents of the closed-over functions and variables.)

Edit History