> for arr[i] = 2, you'll still need to subtract 1

You can, but don't need to. Just have compiler store array pointer constant as (arr-1) instead of arr, et voila, zero runtime overhead.

Or, for modern(-ish) ISAs, often you can add/substract a small constant at runtime, with no extra cycles taken. For example, for x86_64:

  # rbp contains "true" pointer to arr
  # rax contains 1-based array index
  mov rax, QWORD PTR [rbp-8+rax*8]

Another option for users is to "sacrifice" the 0th entry of an array. Depending on the size(as in sizeof) of the entry, it can be worth it.

A benefit is that you can then use 0 as a sentinel value; for instance if you have a find() routine that surely can fail, it can just return 0 instead of having e.g. -1 (which can introduce minor issues).

In my experience, though, I am so used to 0-based index that switching schemes can cause stupid off-by-one bugs. I guess that's the main reason behind complains about Lua. It's not that "natural" arrays are thought of as bad, but mixing both schemes (often C an Lua) is error-prone.

It's basically only x86 among modern ISAs that lets you do base + literal + register * literal, aarch64 only gives you base + register shifted by a literal, and I believe RISC-V is similar: https://gcc.godbolt.org/z/dz768768z

>and I believe RISC-V is similar

Yes it is. It was evaluated, carefully weighted and discarded, as it was not worth it.

But it will still execute with likely no extra time at all due to OOE and how fast arithmetics are.

OOE doesn't necessarily save you if you end up with a hard dependency on the value of the read (and even if it did, the little cores on ARM SoCs are in-order). This is a pretty obvious candidate for macro-op fusion, but I'm not sure whether this actually happens (and if it happens on ARM little cores, etc.)

If it is not on the hot path, it is likely free, but not guaranteed. If it is on the hot path then it is wasting a whole cycle. And of course in highly ALU-dependent code it is another instruction, so a fraction of a clock.

What do you mean it wastes a whole cycle? It may indeed have worse performance due to blowing the instruction cache, but I don’t see why would out-of-order execution be slower on the hot path - I doubt there would be too many hot paths without any dependence on memory fetches outside specific benchmarks - the memory loads will take significantly more time even if they hit cache.

What if the variable is used for array indexing and something else, like displayed to the user?

Can't say I've really thought this through, but couldn't you just subtract 1 (*sizeof(X)) from the arr address?

What would memset(arr, 0, sizeof(arr)) do in that case? I can see myriad problems with having arr itself point outside the memory range allocated to arr[min..max].

Hmm, maybe?

You'd still have the odd subtract here or there, but for many use patterns you could probably ignore the overhead.

For most cases you could get away with some index transformation logic in the compiler (but probably your language would need special types for indices, so I do agree with you that it is not worth the trouble).