`view` emits a boundscheck that can't be turned off.

[vchuravy]

using StaticArrays
@inbounds function f(data, idx)
         arr = MArray{Tuple{3,3}}(data...)
         va = view(arr, :, idx)
         va[2] = 2
         return SArray(arr)
end

emits a boundscheck despite the @inbounds.

julia/base/subarray.jl

Line 165 in 642a88c

@_inline_meta

should probably use @propagate_inbounds.

;  @ REPL[23]:3 within `f'
; ┌ @ subarray.jl:167 within `view'
   %24 = getelementptr inbounds { i64 }, { i64 }* %4, i64 0, i32 0
   store i64 %2, i64* %24, align 8
; │┌ @ abstractarray.jl:610 within `checkbounds' @ abstractarray.jl:589
; ││┌ @ abstractarray.jl:636 within `checkbounds_indices' @ abstractarray.jl:636
; │││┌ @ abstractarray.jl:667 within `checkindex'
; ││││┌ @ bool.jl:36 within `&'
       %.off = add i64 %2, -1
       %25 = icmp ugt i64 %.off, 2
; ││└└└
; ││ @ abstractarray.jl:610 within `checkbounds'
    br i1 %25, label %L21, label %L24

L21:                                              ; preds = %top
    %26 = getelementptr inbounds [3 x {}*], [3 x {}*]* %gcframe61, i64 0, i64 2
    store {}* %18, {}** %26, align 16
    %27 = call nonnull {}* @j_throw_boundserror_711({}* nonnull %18, { i64 }* nocapture readonly %4)
    call void @llvm.trap()
    unreachable

The bounds-check itself wouldn't be that bad if it in turn wouldn't cause us to capture the MArray and thus preventing allocating it on the stack, so an alternative fix would be for Base.throw_boundserror to not capture the array it's based off.

[KristofferC]

so an alternative fix would be for Base.throw_boundserror to not capture the array it's based off.

Heh, #29867 -> #35748.

[KristofferC]

Tangentially, please provide input arguments for issues like this.

[vchuravy]

f((rand(9)...,), 2)

should be an appropriate input.

[mbauman]

This is a StaticArrays issue:

https://github.com/JuliaArrays/StaticArrays.jl/blob/59f92e0ca7ac391a850a6e7a2ce1eb53aa237fc4/src/MArray.jl#L273-L280

This should do the @inline/@boundscheck/@inbounds dance (or perhaps @propagate_inbounds if you wish).

[mbauman]

But also, does this really work?

@inbounds function f(data, idx)
         #...
end

My understanding was that @inbounds was a runtime-thing, not a parse-thing. That is, this'll turn off bounds checking for the definition of the function, not the running of it.

[KristofferC]

But also, does this really work?

Nope.