[kazinator]

Any situation where the program depends on the expression producing a new string each time it is evaluated, rather than returning the same string. The program may be modifying the string, on the assumption that nothing else has access to it, since it is brand new. The program could also be relying on the string having a unique identity, not comparing equal to any previously seen object. (E.g. assuming that each time the expression is evaluated, it produces an object that can serve as a unique key in an EQ hash table).

Any situation in which these behaviors cannot be ruled out (because the object escapes beyond the scope of analysis), the optimization cannot be applied.

[basil-rash]

Ah, well all JS strings are always immutable and only value-referable (you have no access to the underlying memory location), so that’s not a concern here.

[kazinator]

What about the identity side of it? Does the JS specification say that an operation like "a" + "b" is not required to create a new object? Regardless of whether there is such a spec, you can write code that is sensitive to the difference.

[basil-rash]

In fact, you cannot. But I encourage you to try!

[kazinator]

It looks like JS doesn't expose equality operator which can distinguish different strings. Thus "abc" and "abc" are the same object, no matter how they are produced, even if under the hood they are separate instances.

[basil-rash]

Pretty much, though some would contest calling strings “objects”.

The spec indeed goes through some trouble to ensure they are pure value-types and do not exhibit any reference-like semantics, for instance by prohibiting their use as keys of WeakMaps and WeakSets - along with numbers, booleans, nullish values, and bignums.