Indices, not Pointers

But in C, there's not really any difference between pointers and indices. You can access array elements using either method, and they're both equally efficient.

Depending on how many elements you have, you can save some space using 32-bit or even 16-bit indices in place of 64-bit pointers. (Just make sure there isn't any route to overflow the index type.)

There are some differences.

- You can check indices that are out of bounds without running into formal undefined behavior. ISO C does not require pointers to distinct objects to be comparable via inequality, only exact equality. (In practice it works fine in any flat-address-space implementation and may be regarded as a common extension.)

- Indices are implicitlyl scaled. If you have done a range check that index is valid, then it refers to an entry in your array. At worst it is some unoccupied/free entry that the caller shouldn't be using. If you have checked that a pointer points into the array, you don't know that it's valid; you also have to check that its displacement from the base of the array is a multiple of the element size; i.e. it is aligned.

The distinction in the article really is between calling malloc for every added node in your data structure (“pointer”) or using the pre-allocated memory in the next element of an array (“index”).

You do have to take care of the ABA problem - if you access memory using an index that became invalid before and another object is using instead, you will have some weird hard-to-debug logic errors (worse than use-after-free, since even Valgrind can't save you). To prevent this you need another generational counter to store along with your id (which is either incremented for every usage or assigned a random hash)

After being translated literally in machine language, they are not equally efficient.

However, a C compiler may choose to use in machine language whichever of indices or pointers is more efficient on the target machine, regardless of whether the source program uses indices or pointers.

This matters only for shared data structures. It is irrelevant for thread-local data.

For shared data structures, you have more to worry about, so regardless if you use indices or pointers you must use either atomic operations or means to ensure exclusive access to the entire data structure or means to detect the need for retries when using optimistic accesses.

This is only one of the advantages discussed in TFA. The others are those due to using indices instead of pointers (like smaller size, cache locality, range checking, possibility of data exchange between systems with distinct address spaces).

Well, solving/mitigating the ABA ambiguity can debug use-after-free errors in single-threaded programs also. Because when a pointer A is freed to B, and then recycled again for a new object, we can make it into a different pointer A' (e.g. with a tagging scheme). So then when the old A pointer copies are lingering around, we can tell they are invalid due to having the wrong tag.

Solving ABA is probably a point in favor of indices (if we are working in a higher level language) because their type supports the bit operations for tagging. However, some hardware has support for hardware tagging for pointers. E.g. ARM; Android uses it.

With indices what you say can be implemented trivially, much simpler than with pointers, by always incrementing a reallocated index (i.e. an index extracted from the free list) with the array size and always addressing the array with the indices modulo the array size.

With the array size chosen to be a power of two, this adds negligible overhead in time and no overhead in space.

That's equivalent to packing two counters into one.

True.