Mutation of input arguments

[baggepinnen]

Code like this https://github.com/JuliaDiff/FiniteDifferences.jl/blob/c7a4624092d90b6d49d5701967982226231d4478/src/grad.jl#L18

mutates the input argument, but before returning makes sure the input argument is restored to its original value. However, if the user-provided function errors, the input argument is left in a mutated state, which can cause bugs that are very hard to detect.

[willtebbutt]

Good point. Do you think a try ... finally ... end block, in which the restoration of the original value happens in the finally section, would solve the problem?

[baggepinnen]

It would probably work as long as the catch block has access to the index at which the failure occurred. try/catch is expensive in some situations, but I guess that the function being differentiated is typically much more expensive so perhaps it's ont much of a problem.

[oxinabox]

Yeah, try-finally is the correct pattern for this, we should be doing that.

(Possibly in the future we should use the effects to check if a function is side-effect free, then we can avoid the try-finally, and the defensive copy)