In computer science, the funarg problem (function argument problem) refers to the difficulty in implementing first-class functions (functions as first-class objects) in programming language implementations so as to use stack-based memory allocation of the functions.

The difficulty only arises if the body of a nested function refers directly (i.e., not by argument passing) to identifiers defined in the environment in which the function is defined, but not in the environment of the function call. A standard resolution is either to forbid such references or to create closures.

There are two subtly different versions of the funarg problem. The upwards funarg problem arises from returning (or otherwise transmitting "upwards") a function from a function call. The downwards funarg problem arises from passing a function as a parameter to another function call.

When one function A calls another function B during a typical program's execution, the local state of B (including parameters and local variables) must be stored somewhere. In a leaf function, this can be stored in registers, but if B calls another function C which itself uses those registers, they must be preserved somewhere in memory so B can proceed after C returns.

In most compiled programs, this local state is stored on the call stack in a data structure called a stack frame or activation record. This stack frame is pushed, or allocated, before B calls C, and is popped, or deallocated, just before B returns. The upwards funarg problem arises when the return value of B is itself a function F which refers to B's state (say, some local variable V) after B has returned. A may call F, which needs V to still be allocated. Therefore, the stack frame containing B's state variables must not be deallocated when B returns, violating the stack-based function call paradigm.

One solution to the upwards funarg problem is to simply allocate all activation records from the heap instead of the stack and rely on some form of garbage collection or reference counting to deallocate them when they are no longer needed. Managing activation records on the heap has historically been perceived to be less efficient than on the stack (although this is partially contradicted) and has been perceived to impose significant implementation complexity. Most functions in typical programs (less so for programs in functional programming languages) do not create upwards funargs, adding to concerns about potential overhead associated with their implementation. Furthermore, this approach is genuinely difficult in languages that do not support garbage collection.

Some efficiency-minded compilers employ a hybrid approach in which the activation records for a function are allocated from the stack if the compiler is able to deduce, through static program analysis, that the function creates no upwards funargs. Otherwise, the activation records are allocated from the heap.

Another solution is to simply copy the value of the variables into the closure at the time the closure is created. This will cause a different behavior in the case of mutable variables, because the state will no longer be shared between closures. But if it is known that the variables are constant, then this approach will be equivalent. The ML languages take this approach, since variables in those languages are bound to values—i.e. variables cannot be changed. Java also takes this approach with respect to anonymous classes (and lambdas since Java 8), in that it only allows one to refer to variables in the enclosing scope that are effectively final (i.e. constant).