Indexing out-of-bounds not throwing

[korsbo]

nn = SimpleChain(
  static(10),
  TurboDense(identity, 1),
)

nn(randn(10), SimpleChains.init_params(nn))[2]

This seems like a bit of a footgun. Indeed, my foot is not unscathed.

[chriselrod]

julia> StrideArraysCore.boundscheck() = true

julia> nn(randn(10), SimpleChains.init_params(nn))[2]
ERROR: BoundsError: attempt to access 1-element PtrArray{Tuple{StaticInt{1}}, (true,), Float64, 1, 1, 0, (1,), Tuple{StaticInt{8}}, Tuple{StaticInt{1}}} with indices 1:1:1 at index [2]
Stacktrace:
 [1] throw_boundserror(A::PtrArray{Tuple{StaticInt{1}}, (true,), Float64, 1, 1, 0, (1,), Tuple{StaticInt{8}}, Tuple{StaticInt{1}}}, I::Tuple{Int64})
   @ Base ./abstractarray.jl:703
 [2] checkbounds
   @ ./abstractarray.jl:668 [inlined]
 [3] getindex(A::PtrArray{Tuple{StaticInt{1}}, (true,), Float64, 1, 1, 0, (1,), Tuple{StaticInt{8}}, Tuple{StaticInt{1}}}, i::Int64)
   @ StrideArraysCore ~/.julia/packages/StrideArraysCore/pjIfM/src/ptr_array.jl:569
 [4] top-level scope
   @ REPL[8]:1

There are a ton of functions in the Julia ecosystem that don't have @inbounds because, even though they iterate based on the size of the arrays, "there's no guarantee that the array types actually implemented those functions correctly".
Forcing me to either reimplement the world, or be much more dangerous.

[chriselrod]

Note that if you start Julia with --check-bounds=yes, you don't need to manually set boundscheck() = true...

julia> using SimpleChains

julia> nn = SimpleChain(
         static(10),
         TurboDense(identity, 1),
       )
SimpleChain with the following layers:
TurboDense static(1) with bias.

julia> nn(randn(10), SimpleChains.init_params(nn))[2]
ERROR: BoundsError: attempt to access 1-element StrideArraysCore.PtrArray{Tuple{Static.StaticInt{1}}, (true,), Float64, 1, 1, 0, (1,), Tuple{Static.StaticInt{8}}, Tuple{Static.StaticInt{1}}} with indices 1:1:1 at index [2]
Stacktrace:
 [1] throw_boundserror(A::StrideArraysCore.PtrArray{Tuple{Static.StaticInt{1}}, (true,), Float64, 1, 1, 0, (1,), Tuple{Static.StaticInt{8}}, Tuple{Static.StaticInt{1}}}, I::Tuple{Int64})
   @ Base ./abstractarray.jl:703
 [2] checkbounds
   @ ./abstractarray.jl:668 [inlined]
 [3] getindex(A::StrideArraysCore.PtrArray{Tuple{Static.StaticInt{1}}, (true,), Float64, 1, 1, 0, (1,), Tuple{Static.StaticInt{8}}, Tuple{Static.StaticInt{1}}}, i::Int64)
   @ StrideArraysCore ~/.julia/packages/StrideArraysCore/pjIfM/src/ptr_array.jl:569
 [4] top-level scope
   @ REPL[3]:1

[korsbo]

Ok, so a general fix for this would impact performance too much - leaving it up to the documentation to be explicit about the danger?

[chriselrod]

Ok, so a general fix for this would impact performance too much - leaving it up to the documentation to be explicit about the danger?

I guess it isn't obvious to anyone where to look/maybe it should be plastered everywhere? https://github.com/JuliaSIMD/StrideArraysCore.jl#caution