ComputedSet (values-only ComputedPairs) about fusion

[boatbomber]

Potentially pretty valuable (especially if we can get ipairs internally for some extra speed), but the workaround of [val] = true seems good enough for now. This definitely isn't a blocking issue, but could be implemented in the second release!

[dphfox]

Ideal for transforming unordered arrays, where order does not matter but duplicates are allowed.

In ForValues, you say this. But doesn't this introduce the initial issue of subtle recalculations?

Nope! The initial issue of subtle recalculations is down to keys being assumed to carry meaning, which they don't in ForValues.

To illustrate:

local array = Value({"Red", "Green", "Blue"})

ComputedPairs(array, function(key, value)
    -- notice ComputedPairs gives you the key
    -- ComputedPairs has to assume the key could be used to calculate the returned value here
    -- therefore, we can't reuse a value if it appears at a different key
    return {
        colour = value
    }
end))

ForValues(array, function(value)
    -- here, ForValues doesn't give us the key
    -- the key could never be used to calculate the returned value here
    -- therefore, we *can* reuse a value if it appears at a different key
    return {
        colour = value
    }
end))

Specifically with respect to duplicate values, the only condition is that two values in the output array never point to the same cached object. ForValues doesn't have to make any further guarantees.

[dphfox]

I've decided on an API design for this after much consideration with users.

I plan to replace ComputedPairs with 3 new mapping objects (names being bikeshedded).

---

ForKeys

Iterates over every pair of the input table. Transforms the key using the processor function (KI) -> KO. Puts transformed key and original value in output table.

Processing only occurs when a key is added. Transformed keys are cached. Values do not trigger processing but are forwarded to the output table directly.

Ideal for transforming sets, where elements are stored as keys.

ForValues

Iterates over every pair of the input table. Transforms the value using the processor function (VI) -> VO. Puts original key and transformed value in output table.

Processing only occurs when a key is added or when a key's value changes to a non-nil value. Transformed values are cached, though care needs to be taken with handling duplicate input values to ensure every value in the output table is unique.

Ideal for transforming unordered arrays, where order does not matter but duplicates are allowed.

ForPairs

Iterates over every pair of the input table. Transforms the pair using the processor function (KI, VI) -> (KO, VO). Puts transformed pair in output table.

Processing only occurs when a key is added or when a key's value changes to a non-nil value. Transformed pairs are cached by key.

Ideal for transforming ordered arrays and dictionaries, where both keys and values carry meaning.

---

Under this new system there is no direct replacement for ComputedPairs. While this may sound bad, it's actually great - it forces users to consider which behaviour they need, which means Fusion can better optimise for each case without user intervention or cognitive load.

[nottirb]

Do we want the optional destructor to be called with the output of For*? Since that's how ComputedPairs currently works:

processor: (K, VI) -> VO,
destructor: (VO) -> ()?

So the idea would be as follows,

For ForKeys:

processor: (KI) -> KO,
destructor: (KO) -> ()?

For ForValues:

processor: (VI) -> VO,
destructor: (VO) -> ()?

For ForPairs:

processor: (KI, VI) -> (KO, VO)
destructor: (KO, VO) -> ()?

[Dionysusnu]

If we're on it, this is a good time to implement #97 along with it. But since that's not quite approved yet, maybe this issue should wait a bit.

[dphfox]

Yes - I think that the destructor should be forwarded all returned values from the processor function, including any extra metadata the user wishes to provide (see #97).

[Dionysusnu]

Ideal for transforming unordered arrays, where order does not matter but duplicates are allowed.

In ForValues, you say this. But doesn't this introduce the initial issue of subtle recalculations?