hcat of many SparseVectors freezes / crashes in julia 1.9

[jtackm]

Running this toy code that concatenates a large number of sparse vectors

nsamp, notus = (8000, 2500)
vecs = [sparse(zeros(nsamp)) for i in 1:notus]
@time hcat(vecs...)

works fine in julia 1.8.2 (~5.5s runtime), but hangs for a long time in julia 1.9. If I ommit sparse, everything works correctly. In a real-world example julia usually hangs, but sometimes finishes quickly, crashes, and/or produces stack overflows like this one:

Internal error: stack overflow in type inference of _is_fixed(SparseArrays.SparseVector{Int64, Int64}, SparseArrays.SparseVector{Int64, Int64}, SparseArrays.SparseVector{Int64, Int64}, SparseArrays.SparseVector{Int64, Int64}, 
[ many repetitions ]
SparseArrays.SparseVector{Int64, Int64}, SparseArrays.SparseVector{Int64, Int64}, SparseArrays.SparseVector{Int64, Int64}, SparseArrays.SparseVector{Int64, Int64}).
This might be caused by recursion over very long tuples or argument lists.

My version:

julia> versioninfo()
Julia Version 1.9.0
Commit 8e630552924 (2023-05-07 11:25 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 256 × AMD EPYC 7742 64-Core Processor
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, znver2)
  Threads: 96 on 256 virtual cores
Environment:
  JULIA_WORKER_TIMEOUT = 3600
  JULIA_LOAD_PATH = [ removed ]
  JULIA_DEPOT_PATH = [ removed ]
  LD_GOLD = [ removed ]
  JULIA_NUM_THREADS = 96

[vtjnash]

Looks like it is caused by JuliaSparse/SparseArrays.jl@43b4d01a. We should probably consider tring to use afoldl in more cases, such as this one, instead of pure recursion.

[VPetukhov]

Same problem here. It was fine in julia 1.8.5, and it's broken now. Replacing hcat(vecs...) with reduce(hcat, vecs) fixes a part of it, but sometimes a new error is thrown (I don't have access to the server atm, so can't post the exact error).

[KristofferC]

Generally, splatting an unknown number of items is a bad idea. Would be good to know what the error is when using reduce though.

[jtackm]

Thanks! Okay, I thought my example is the canonic way of concatenating vectors. So is reduce the recommended approach then or am I missing something else? (I can ask on discourse if this is too tangential)

[stevengj]

The canonical way, as of Julia 1.9, is stack.

[Minhvt34]

The canonical way, as of Julia 1.9, is stack.

I am using Julia 1.11. I replaced A = hcat(A_list...) with stack(A_list) and it worked. Thank you!