Concrete return type of BMI.get_value_ptr
#2018
Conversation
This aligns better with the updated expectations in https://github.com/Deltares/BasicModelInterface.jl/releases/tag/v0.1.1. This means we don't return `SubArrays` or similar anymore, but always a simple `Vector{Float64}`.
core/src/util.jl
Outdated
| function unsafe_array( | ||
| A::SubArray{Float64, 1, Vector{Float64}, Tuple{UnitRange{Int64}}, true}, | ||
| )::Vector{Float64} | ||
| unsafe_wrap(Array, pointer(A), length(A)) |
There was a problem hiding this comment.
This probably requires a GC.@preserve A?
There was a problem hiding this comment.
That would only be needed if A is at risk of GC right? In which case this shouldn't be called anyway.
There was a problem hiding this comment.
Well I rather program defensively than making those assumptions. Worst case you don't get a crash, but silent corruption. I don't think the macro has downsides for us here.
There was a problem hiding this comment.
It doesn't have downsides, but I don't think it helps us here. The docs mention:
Mark the objects x1, x2, ... as being in use during the evaluation of the expression expr.
So if we return from this function this won't apply anymore I think.
I confirmed this with a little experiment:
f(a) = unsafe_wrap(Array, pointer(a), length(a))
g(a) = GC.@preserve a unsafe_wrap(Array, pointer(a), length(a))
a = collect(1:9)
b = f(a) # f or g give the same results
a = 2 # deref the original a
GC.gc() # a bunch of times
b # the data changes after a while
The type stability was indeed not directly giving us problems that I know of. But I want to keep those functions as simple as possible, and when trying out |
This aligns better with the updated expectations in https://github.com/Deltares/BasicModelInterface.jl/releases/tag/v0.1.1.
This means we don't return
SubArraysor similar anymore, but always a simpleVector{Float64}.