alt Hacker NewsNull conditional assignment is bunk.
When you have an expression P which names a mutable place, and you execute P := X, the contract says that P now exhibits the value X, until it is assigned another value.
Conditional assignment fucks this up. When P doesn't exist, X is not stored. (Worse, it may even be that the expression X is not evaluated, depending on how deep the fuckery goes.)
Then when you access the same expression P, the conditional assignment becomes conditional access and you get back some default value like a nil.
Store X, get back nil.
That's like a hardware register, not a durable memory model.
It's okay for a config.connection?.retryPolicy to come up nil when there is no config.connection. It can be that the design makes nil a valid retry policy, representing some default. Or it could be that it is not the case, but the code which uses connection? handles the nil soon afterward.
But a store to config.connection?.retryPolicy not having an effect; that is dodgy.
What you need for config.connection? to do when the expression is being used to calculate a mutable place to assign to is to check that config.connection is null, and in that case, instantiate a representative instance of something which is then assigned to config.connnection, such that the config.connection.retryPolicy place then exists and the assignment can proceed.
This recognizable as a variation on COW (copy-on-write); having some default instance for reading, but allocating something on writing.
In a virtual memory system, freshly allocated memory can appear to contain zero bytes on access due to all of its pages being mapped to a single all-zero frame that exists in the entire system. Conceptually, the hardware could do away with even that all-zero frame and just have a page table entry which says "this is a zero-filled page", so the processor then fakes out the zero values without accessing anything. When the nonexistent page is written, then it gets the backing storage.
In order to instantiate settings.connection? we need to know what that has to be. If we have a static type system, it can come from that: the connection member is of some declared type of which a representative instance can be produced with all constructor parameters defaulted. Under a dynamic paradigm, the settings object can have a handler for this: a request to materialize a field of the object that is required for an assignment.
If you don't want a representative config.connection to be created when config.connection?.retryPolicy is assigned, preferring instead that config.connection stays null, and the assignment is sent to the bit buckets, you have incredibly bad taste and a poor understanding of software engineering and programming language design --- and the design of your program is scatter-brained accordingly.
Replies
This is syntactic sugar only. No change in semantics. So I don't buy into any argument about what this change 'means'.
And we all get to choose what we find ridiculous:
i = i + 1 ? No it does not. Never has, never will.
Connection is null? It's insane to type it as Connection then. null has type Null.
A representative config.connection being made out of nothing sounds pretty bad to me. If you want to make sure the value doesn't disappear, you shouldn't be using conditional assignment in the first place.
The config example isn't the best, but instead imagine if it was just connection.?retryPolicy. After you set connection?.retryPolicy it would be weird for reading it back to be null. But it would be just as weird for connection?.retryPolicy to not be null when we never established a connection in the first place.
The copy on write analogy is tempting but what you're describing only works when the default value is entirely made of nulls. If you need anything that isn't null, you need to actually make an object (either upfront or on first access). And if you do that, you don't need ?. anymore.
Apparently you do t use if in your code?
> When you have an expression P which names a mutable place, and you execute P := X
This isn't the case, though, is it? A normal member access (or indexer) expression may point to a mutable location (field, property). However, with conditional access expressions you get either a member access _or nothing_. And that nothing is not a mutable place.
When you use any of the conditional operators, you split the following code into two paths, and dropping the assignment (since there's nothing to assign to) seems pretty consistent to me, since you'd also drop an invocation, or a property evaluation in similar expressions.
If you require P to point to something that actually exists because you want the assignment to succeed, then write code to ensure that P exists because the assignment has no way of knowing what the intention was on the left side.