Code review: 7012540 (java.util.Objects.nonNull() incorrectly named)

[mark.reinhold at oracle.com]

Executive summary: requireNonNull is the preferred name.

> Date: Tue, 25 Jan 2011 18:33:47 -0500
> From: brian.goetz at oracle.com

> mark.reinhold at oracle.com wrote:
>> I think requireNonNull(x) is confusing.
> 
> Remember there's two versions of someModifierNonNull being discussed; the one
> currently in Objects is the precondition-enforcing variety, not the
> postcondition-ensuring variety.  Are we talking about the same thing?

Yes, i.e., the former.

>> For those familiar with the "requires/ensures/modifies" triad of verbs as
>> a compact way of identifying the preconditions, postconditions, and side
>> effects of a method or procedure in a comment, a specification, or a more
>> formal design-by-contract framework, "requires" is just wrong.
>> 
>> When analyzing the invocation of foo in your example, the non-nullity of
>> s and t are preconditions of foo and therefore postconditions of the
>> check method.  Naming the check method "requireNonNull" suggests that
>> the check method itself has a precondition that its argument be non-null,
>> when in fact it's the check method's postcondition which ensures that
>> property.
>> 
>> Since postconditions are labeled "ensures" in the "r/e/m" triad, this
>> method should be named "ensureNonNull".
> 
> Right, there's precedent for "ensureXxx" methods to actually change the state
> of things to ensure the postcondition, such as ensureCapacity() methods in the
> collection implementations.  Given that a part of the motivation for this
> change was to leave room in the namespace for both the precondition-enforcing
> variety (barf on null) and the postcondition-ensuring variety, aka the
> carpet-sweeping variety ("if it is null, make it non-null") ensureNonNull
> sounds a lot more like the the carpet-sweeping version than the version being
> discussed (barf on null).

I was actually arguing that the barfing version should be named
ensureNonNull.

(Wow, am I actually writing about barfing methods?)

> The r/e/m framework seems to support require for the throwing version (as
> implemented by this patch): for the throwing version, non-nullity is a
> precondition of the check method (if the condition is not met, error), whereas
> for the carpet-sweeping version, it is a postcondition of the check method (if
> the check method can come up with a reasonable default value).  (It happens to
> be a postcondition of both, but the significant behavior and use of the
> throwing version currently in Objects is to enforce an error when the
> precondition is not met.)  Therefore:
> 
>   requireNonNull(x) -> throw if x == null
>   ensureNonNull(x)  -> convert x to a non-null value if null
> 
> seems like the right taxonomy.

It depends on your perspective.

In a nested method invocation, the postconditions of the inner method
must relate to the preconditions of the outer method.  That is, for the
expression

    foo(bar(x))

to be correct then the postconditions of the bar method must imply the
preconditions of foo method.

Now suppose that bar is intended to be the barf-on-null method.

If we consider how to name bar on its own, outside of its intended usage
as in this example, then it's natural (for me, anyway) to conclude that
"ensuresNonNull" is the right name.  What's important about bar is its
postcondition, namely that its return value (x) is not null.  Naming it
"requireNonNull" seems backwards on this view, because "requires" is
always about preconditions, and it's definitely not a precondition of
this method that its argument not be null.

If, by contrast, we consider how to name bar in the context of the above
example then it's easy to see that "requireNonNull" is the better name,
because what's important in the overall expression is satisfying foo's
precondition that its argument not be null, and so

    foo(requireNonNull(x))

reads very fluently.

A little utility method like this will never (well, hardly ever) be used
in isolation, so I agree now that it makes sense to prefer the second
form, i.e., requireNonNull.

I'm not sure I agree that the corresponding carpet-sweeping method is
best named ensureNonNull; treating that as a conversion, i.e., asNonNull,
may be a better choice.  That, however, is a discussion for another
release.

- Mark