isassigned is slow for out-of-range indices

[gaurav-arya]

MWE:

using StaticArrays, BenchmarkTools

arr = SA[1, 2, 3]

@btime isassigned(arr, 4) # 17 μs

This is because Base's fallback implementation uses a try/catch, which is slow in the catch case:
https://github.com/JuliaLang/julia/blob/def2ddacc9a9b064d06c24d12885427fb0502465/base/abstractarray.jl#L605-L616

This really hurts performance when isassigned is being used in a case where the index is really expected to be out-of-range sometimes. I even wonder if Base's fallback implementation is just a bad idea, because it will always be slow in this case: something like i in eachindex(arr) would at least be faster for static arrays.

Using Julia 1.9 and StaticArrays v.1.5.12

[thchr]

There used to be a (faulty) isassigned specialization for StaticArrays but it was removed recently (#1096), partly due to incorrectness, invalidations, and because it was not clear why performance would matter here.

What do you need a performant isassigned for?

[gaurav-arya]

The logic was something like "given an array and an index, see if index + 1 is a valid index and access array there if so". I was using it in my package here:

https://github.com/gaurav-arya/StochasticAD.jl/blob/892486bec7bb0757a2fef6e6e8afd12374d4ebe4/src/general_rules.jl#L259

Although I've since replaced it with i in eachindex(arr) because of this issue

[thchr]

Yeah, that seems better: isassigned is intended for more than just bounds checking (i.e., it also returns false if the index is #undef).

See also checkbounds and the Base docs.

Closing cf. earlier discussion in #1096.

[gaurav-arya]

Thanks, that makes sense, checkbounds seems like the right thing to use indeed.

(Although note that Base's Array does have its own specialization: https://github.com/JuliaLang/julia/blob/1aa65c33f24e7a6c7a341ce4a13202bfb6ac4811/base/array.jl#L258)