Splatting into push! for arrays of eltype Any is broken for length bigger than four after having loaded LoopVectorization

[mwlidar]

[Issue updated by @KristofferC to include info about LoopVectorization]

On v1.9.0-beta2 splatting into push! does not work for Arrays of eltype Any for more than four arguments

julia> using LoopVectorization

julia> aa = []
Any[]

julia> push!(aa, [1,2,3,4,5,6]...)
4-element Vector{Any}:
 1
 2
 3
 4

It works on 1.8.3 and for example for Arrays of type Int. Have not tried others so far...

This is on windows:

Julia Version 1.9.0-beta2
Commit 7daffeecb8 (2022-12-29 07:45 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: 8 × Intel(R) Core(TM) i7-8665U CPU @ 1.90GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, skylake)
  Threads: 4 on 8 virtual cores
Environment:
  JULIA_NUM_THREADS = 4
  JULIA_PKG_DEVDIR = H:\pc\julia
  JULIA_EDITOR = code

[giordano]

Works on master:

julia> aa = []
Any[]

julia> push!(aa, [1,2,3,4,5,6]...)
6-element Vector{Any}:
 1
 2
 3
 4
 5
 6

julia> aa
6-element Vector{Any}:
 1
 2
 3
 4
 5
 6

[KristofferC]

Works on 1.9.2 for me (Mac):

julia> aa = []
Any[]

julia> push!(aa, [1,2,3,4,5,6]...)
6-element Vector{Any}:
 1
 2
 3
 4
 5
 6

[mwlidar]

Ok, found the culprit...
On a pristine session the error does not occur. But I had loaded a few modules which I use as a standard and among them LoopVectorization. Without LoopVectorization everything is fine.

[KristofferC]

Before closing, we should probably figure out what is happening here.

Does not happen when running through debugger:

julia> using LoopVectorization

julia> aa=[]; push!(aa, 1,2,3,4,5,6)
4-element Vector{Any}:
 1
 2
 3
 4

julia> using Debugger

julia> aa=[]; @run push!(aa, 1,2,3,4,5,6)
6-element Vector{Any}:
 1
 2
 3
 4
 5
 6

Does not happen when running with --compiled-modules=no.

[KristofferC]

It seems to break when this method is defined https://github.com/JuliaSIMD/LoopVectorization.jl/blob/35f83103c12992ddd887cd709bf65e345db5ec9e/src/modeling/graphs.jl#L1653-L1667 (which is a Base.push! method)...

Regressed between 1.9.0-alpha1 and 1.9.0-beta2

[gbaraldi]

Bisected to 31df7c8
@vtjnash

[KristofferC]

MWE is:

A.jl:

module A
using B
function Base.push!(a,b,c,d) end
end

B.jl:

module B
f() = push!(a, nothing, nothing, nothing, nothing)
precompile(Tuple{typeof(f)})
end

REPL:

julia> push!(LOAD_PATH, ".");

julia> using A

julia> aa = []
Any[]

julia> push!(aa, [1,2,3,4,5,6]...)
4-element Vector{Any}:
 1
 2
 3
 4

[vtjnash]

Isolated to this

(rr) p jl_(jl_normalize_to_compilable_mi(mi))
push!(Array{Any, 1}, Any, Any, Any, Any) from push!(Array{Any, 1}, Any...)
$43 = void
(rr) p jl_(mi)
push!(Array{Any, 1}, Any, Any, Any, Any...) from push!(Array{Any, 1}, Any...)
$44 = void

This is not correct though, since jl_normalize_to_compilable_mi should be wider than the input type, but here it forgot to rewrap in the Vararg bound