Arpack.svds is not type stable

[mileslucas]

Inferred types from Arpack.svds on dense matrix inputs is failing

MWE:

julia> using Arpack
julia> X = randn(10, 1000);
julia> @code_warntype Arpack.svds(X; nsv=3)
Variables
  #unused#::Core.Const(Arpack.var"#svds##kw"())
  @_2::NamedTuple{(:nsv,), Tuple{Int64}}
  @_3::Core.Const(Arpack.svds)
  A::Matrix{Float64}
  kwargs...::Base.Iterators.Pairs{Symbol, Int64, Tuple{Symbol}, NamedTuple{(:nsv,), Tuple{Int64}}}

Body::Tuple{LinearAlgebra.SVD, Any, Any, Any, Any}
1 ─      (kwargs... = Base.pairs(@_2))
│   %2 = Arpack.:(var"#svds#40")(kwargs...::Core.PartialStruct(Base.Iterators.Pairs{Symbol, Int64, Tuple{Symbol}, NamedTuple{(:nsv,), Tuple{Int64}}}, Any[NamedTuple{(:nsv,), Tuple{Int64}}, Core.Const((:nsv,))]), @_3, A)::Tuple{LinearAlgebra.SVD, Any, Any, Any, Any}
└──      return %2

(in the REPL the Body return type is red, and Test.@inferred calls fail)

system info:

Julia Version 1.6.0
Commit f9720dc2eb* (2021-03-24 12:55 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin20.3.0)
  CPU: Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.1 (ORCJIT, skylake)

[mileslucas]

Here's some traces from Cthulhu.@descend_code_warntype

│ ─ %-1  = invoke #_svds#47(::Int64,::Bool,::Float64,::Int64,::Int64,::Vector{Float64},::Int64,::typeof(Arpack._svds),::Matrix{Float64})::Tuple{LinearAlgebra.SVD, Any, Any, Any, Any}
Variables
  #_svds#47::Core.Const(Arpack.var"#_svds#47")
  nsv@_2::Int64
  ritzvec::Bool
  tol::Float64
  maxiter::Int64
  ncv::Int64
  v0::Vector{Float64}
  check::Int64
  @_9::Core.Const(Arpack._svds)
  X::Matrix{Float64}
  #48::Arpack.var"#48#49"{_A} where _A
  @_12::Int64
  U::Any
  V::Any
  svals::Any
  r::Any
  firstzero::Any
  threshold::Any
  realex1::Any
  ex::Tuple
  min_mn::Any
  otype::Type{Float64}
  n::Int64
  m::Int64
  nsv@_25::Any
  @_26::Any

it looks like

Arpack.jl/src/Arpack.jl

Line 333 in 6baea77

function _svds(X; nsv::Int = 6, ritzvec::Bool = true, tol::Float64 = 0.0, maxiter::Int = 1000, ncv::Int = 2*nsv, v0::Vector=zeros(eltype(X),(0,)), check::Integer=0)

is the culprit method.

[ViralBShah]

Issue is #136

[ViralBShah]

Closing as dup of #136.