Closed MilesCranmer closed 6 months ago
The only tricky thing is that if the user modifies the underlying array (directly), these assumptions will get violated...
The original idea of ReadOnlyArrays.jl was to protect from unwanted mutation. As you comment we currently do not guarantee that the underlying array will not change.
Yeah good point. I guess this is better left for a core array operation (https://github.com/JuliaLang/julia/pull/44381 — though that has stalled).
I think some caching of metadata will help ReadOnlyArrays to start to get performance improvements over regular arrays.
Since a particular
ReadOnlyArray
cannot change shape, you can cache its shape information in the struct itself as an immutable value. Note that this is not stored as a type parameter – which is what StaticArrays.jl does – this change instead stores it as a value. But, since this value is immutable, LLVM should be able to do things like infer that two loops, one after the other, have the same number of iterations, and do some optimizations with that information.While this does not yet take advantage of the fact that the contents are unchanging (see https://github.com/JuliaLang/julia/pull/44381 for that), it will hopefully get some perf improvements from cached shape info.