subtyping bug for constrained varargs

[vtjnash]

This is bad:

julia> T = SubArray{Int64,0,Array{Int64,3},Tuple{Int64,Int64,Int64},false};
julia> methods(Base._getindex, (IndexCartesian, T, Int, Vararg{Int}), )
# 1 method for generic function "_getindex":
[1] _getindex(::IndexCartesian, A::AbstractArray{T,N}, I::Vararg{Int64,N}) where {T, N} in Base at abstractarray.jl:943

This method can never be selected for this signature, and methods is claiming it will always be selected.

This is correct:

julia> methods(Base._getindex, (IndexCartesian, T, Int,), )
# 1 method for generic function "_getindex":
[1] _getindex(::IndexCartesian, A::AbstractArray, I::Vararg{Int64,M}) where M in Base at abstractarray.jl:937

We expect to need to reduce I at runtime until it matches the length of A dims (0).

This is what happens when get this wrong (reduced from a subarray test optimization):

julia> B = ones(Int64, (1, 1, 1));

julia> A = view(B, 1, 1, 1);

julia> function _test_mixed(A)
                  for J in CartesianIndices(size(A)[2:end])
                  A[1, J]
                  end
                  nothing
              end
_test_mixed (generic function with 1 method)

julia> _test_mixed(A)
Unreachable reached at 0x113b799a3

signal (4): Illegal instruction: 4
in expression starting at no file:0
getindex at ./abstractarray.jl:904 [inlined]
_test_mixed at ./REPL[4]:3
jl_fptr_args at /Users/jameson/julia-master/src/gf.c:1823
jl_fptr_trampoline at /Users/jameson/julia-master/src/gf.c:1813
jl_apply_generic at /Users/jameson/julia-master/src/gf.c:2142
do_call at /Users/jameson/julia-master/src/interpreter.c:324
eval_value at /Users/jameson/julia-master/src/interpreter.c:409
eval_stmt_value at /Users/jameson/julia-master/src/interpreter.c:363
eval_body at /Users/jameson/julia-master/src/interpreter.c:671
jl_interpret_toplevel_thunk_callback at /Users/jameson/julia-master/src/interpreter.c:788
unknown function (ip: 0xfffffffffffffffe)
unknown function (ip: 0x11fe3712f)
unknown function (ip: 0xffffffffffffffff)
jl_interpret_toplevel_thunk at /Users/jameson/julia-master/src/interpreter.c:797
jl_toplevel_eval_flex at /Users/jameson/julia-master/src/toplevel.c:814
jl_toplevel_eval at /Users/jameson/julia-master/src/toplevel.c:823
jl_toplevel_eval_in at /Users/jameson/julia-master/src/builtins.c:631
eval at ./boot.jl:319
jl_fptr_args at /Users/jameson/julia-master/src/gf.c:1823
jl_apply_generic at /Users/jameson/julia-master/src/gf.c:2142
eval_user_input at /Users/jameson/julia-master/usr/share/julia/stdlib/v0.7/REPL/src/REPL.jl:85
macro expansion at /Users/jameson/julia-master/usr/share/julia/stdlib/v0.7/REPL/src/REPL.jl:116 [inlined]
#28 at ./task.jl:257
jl_fptr_args at /Users/jameson/julia-master/src/gf.c:1823
jl_fptr_trampoline at /Users/jameson/julia-master/src/gf.c:1813
jl_apply_generic at /Users/jameson/julia-master/src/gf.c:2142
jl_apply at /Users/jameson/julia-master/src/./julia.h:1537
start_task at /Users/jameson/julia-master/src/task.c:268
Allocations: 9571377 (Pool: 9569985; Big: 1392); GC: 20
Illegal instruction: 4

[vtjnash]

More precise repro of the subtyping bug that leads to this:

julia> Tuple{Array{Int,0}, Int, Vararg{Int}} <: Tuple{AbstractArray{T,N}, Vararg{Int,N}} where {T, N}
true

However, this must return false, since the LHS is populated but can never be a subtype of the RHS.