Julia: bounds error on empty splat

[olof3]

As stated in the documentation, one should use a semicolon before keyword arguments. However some packages, perhaps for historical reasons (?), use a comma instead (including a few spurious usages in stdlib).

In nightly, the following mwe throws a cryptic error message, while it worked in v1.5.

 function eigvals_mod(A::AbstractMatrix; kwargs...)
     @assert size(A)[1] == size(A)[2] "Matrix A needs to be square"
     eigvals!(copy(A), kwargs...)
 end

 > eigvals_mod(randn(2,2))

ERROR: BoundsError: attempt to access () at index [1]
Stacktrace:
 [1] getindex(t::Tuple, i::Int64)
   @ Base ./tuple.jl:29
 [2] iterate
   @ ./tuple.jl:66 [inlined]
 [3] iterate
   @ ./iterators.jl:254 [inlined]
 [4] #eigvals_mod#13
   @ ./REPL[15]:3 [inlined]
 [5] eigvals_mod(A::Matrix{Float64})
   @ Main ./REPL[15]:2
 [6] top-level scope
   @ REPL[16]:1

If the assertion is removed or the comma is replaced by a semicolon it works in nightly as well.

Perhaps it is related to #36774 ?

i guess it's not too much of an issue that incorrect usage gives rise to errors, but I guess it would be good to fix the handful of incorrect usages in stdlib, and possibly provide a better error message?

[c42f]

This is not a problem with syntax. It's ok to list keyword arguments without a ; when calling a function with keywords. (See https://docs.julialang.org/en/v1/manual/functions/#Keyword-Arguments)

Oddly enough, if I forget to do using LinearAlgebra before running your example and then correct the mistake interactively, this code works fine as written. Example session:

julia> function eigvals_mod(A::AbstractMatrix; kwargs...)
           @assert size(A)[1] == size(A)[2] "Matrix A needs to be square"
           eigvals!(copy(A), kwargs...)
       end
eigvals_mod (generic function with 1 method)

julia> eigvals_mod(randn(2,2))
ERROR: UndefVarError: eigvals! not defined
Stacktrace:
 [1] #eigvals_mod#7
   @ ./REPL[1]:3 [inlined]
 [2] eigvals_mod(A::Matrix{Float64})
   @ Main ./REPL[1]:2
 [3] top-level scope
   @ REPL[2]:1

julia> using LinearAlgebra

julia> eigvals_mod(randn(2,2))
2-element Vector{Float64}:
 -0.1764691396097775
  2.5547516089663445

So I think this is a compiler bug involving inference and iteration. @Keno could this be due to the changes in #36684?

[JeffBezanson]

Yes it looks like a compiler bug. Works fine with --compile=min.