type inference problem (regression?) in nightly

[sbromberger]

For the last several days, our CI for nightly has been failing with an unexpected type inference issue. This is as minimal a WE as I can come up with right now (and it's not really minimal, so my apologies):

using LightGraphs, LightGraphs.Parallel, LightGraphs.VertexSubsets
g = Graph(0x10, 20)

julia> @code_warntype Parallel.vertex_cover(g, 4, RandomVertexCover(); parallel=:threads)
Variables
  #unused#::Core.Compiler.Const(LightGraphs.Parallel.var"#vertex_cover##kw"(), false)
  @_2::NamedTuple{(:parallel,),Tuple{Symbol}}
  @_3::Core.Compiler.Const(LightGraphs.Parallel.vertex_cover, false)
  g::LightGraphs.SimpleGraphsCore.SimpleGraph{UInt8}
  reps::Int64
  alg::Core.Compiler.Const(LightGraphs.VertexSubsets.RandomSubset{Random._GLOBAL_RNG}(Random._GLOBAL_RNG()), false)
  parallel::Symbol
  seed::Int64
  @_9::Symbol
  @_10::Int64

Body::Any
1 ─ %1  = Base.haskey(@_2, :parallel)::Core.Compiler.Const(true, false)
│         %1
│         (@_9 = Base.getindex(@_2, :parallel))
└──       goto #3
2 ─       Core.Compiler.Const(:(@_9 = :threads), false)
3 ┄       (parallel = @_9)
│   %7  = Base.haskey(@_2, :seed)::Core.Compiler.Const(false, false)
└──       goto #5 if not %7
4 ─       Core.Compiler.Const(:(@_10 = Base.getindex(@_2, :seed)), false)
└──       Core.Compiler.Const(:(goto %12), false)
5 ┄       (@_10 = -1)
│         (seed = @_10::Core.Compiler.Const(-1, false))
│   %13 = (:parallel, :seed)::Core.Compiler.Const((:parallel, :seed), false)
│   %14 = Core.apply_type(Core.NamedTuple, %13)::Core.Compiler.Const(NamedTuple{(:parallel, :seed),T} where T<:Tuple, false)
│   %15 = Base.structdiff(@_2, %14)::Core.Compiler.Const(NamedTuple(), false)
│   %16 = Base.pairs(%15)::Core.Compiler.Const(Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}(), false)
│   %17 = Base.isempty(%16)::Core.Compiler.Const(true, false)
│         %17
└──       goto #7
6 ─       Core.Compiler.Const(:(Base.kwerr(@_2, @_3, g, reps, alg)), false)
7 ┄ %21 = LightGraphs.Parallel.:(var"#vertex_cover#16")(parallel, seed::Core.Compiler.Const(-1, false), @_3, g, reps, alg)::Any
└──       return %21

This infers correctly on 1.4.1 but does not on 1.6.0-dev.

The call is actually just this:

alg = parallel == :threads ?  LightGraphs.VertexSubsets.ThreadedRandomSubset(reps, rng=rng) : LightGraphs.VertexSubsets.DistributedRandomSubset(reps, rng=rng)

LightGraphs.VertexSubsets.vertex_cover(g, alg)

and each possibility returns a Vector{eltype(g)}. (Edit: that is, if you run the function in each branch by itself, it infers correctly.)

Again, this is fine in 1.4.1 but breaks in 1.6.0.

[sbromberger]

Here's the code_warntype for 1.4.1, for completeness:

julia> @code_warntype Parallel.vertex_cover(g, 4, RandomVertexCover(); parallel=:threads)
Variables
  #unused#::Core.Compiler.Const(LightGraphs.Parallel.var"#vertex_cover##kw"(), false)
  @_2::NamedTuple{(:parallel,),Tuple{Symbol}}
  @_3::Core.Compiler.Const(LightGraphs.Parallel.vertex_cover, false)
  g::LightGraphs.SimpleGraphsCore.SimpleGraph{UInt8}
  reps::Int64
  alg::Core.Compiler.Const(LightGraphs.VertexSubsets.RandomSubset{Random._GLOBAL_RNG}(Random._GLOBAL_RNG()), false)
  parallel::Symbol
  seed::Int64
  @_9::Symbol
  @_10::Int64

Body::Array{UInt8,1}
1 ─ %1  = Base.haskey(@_2, :parallel)::Core.Compiler.Const(true, false)
│         %1
│         (@_9 = Base.getindex(@_2, :parallel))
└──       goto #3
2 ─       Core.Compiler.Const(:(@_9 = :threads), false)
3 ┄       (parallel = @_9)
│   %7  = Base.haskey(@_2, :seed)::Core.Compiler.Const(false, false)
└──       goto #5 if not %7
4 ─       Core.Compiler.Const(:(@_10 = Base.getindex(@_2, :seed)), false)
└──       Core.Compiler.Const(:(goto %12), false)
5 ┄       (@_10 = -1)
│         (seed = @_10::Core.Compiler.Const(-1, false))
│   %13 = (:parallel, :seed)::Core.Compiler.Const((:parallel, :seed), false)
│   %14 = Core.apply_type(Core.NamedTuple, %13)::Core.Compiler.Const(NamedTuple{(:parallel, :seed),T} where T<:Tuple, false)
│   %15 = Base.structdiff(@_2, %14)::Core.Compiler.Const(NamedTuple(), false)
│   %16 = Base.pairs(%15)::Core.Compiler.Const(Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}(), false)
│   %17 = Base.isempty(%16)::Core.Compiler.Const(true, false)
│         %17
└──       goto #7
6 ─       Core.Compiler.Const(:(Base.kwerr(@_2, @_3, g, reps, alg)), false)
7 ┄ %21 = LightGraphs.Parallel.:(var"#vertex_cover#16")(parallel, seed::Core.Compiler.Const(-1, false), @_3, g, reps, alg)::Array{UInt8,1}
└──       return %21

[JeffBezanson]

Looks similar in nature to #35800.

[sbromberger]

I don't think we saw these errors going back to May. It's just been in the last couple of weeks, but I'll need to check.

[sbromberger]

We were good as of June 4 at the earliest; perhaps even later than that: sbromberger/LightGraphs.jl#1440

[sbromberger]

@abhinavmehndiratta noticed that the inference issues only happen when the vector returned is empty. If the vector contains values, then no inference error is generated in our test code.