Dispatching on T where Core.Compiler.iskindtype(T) is broken

[mauro3]

This testcase has a dispatch error in it:

# This just add debug info to Base.isbits(t::DataType) 
function Base.isbits(t::DataType)
    if !isa(t,DataType)
        @show typeof(t)
        error("Shouldn't dispatch to here!")
    end
    !t.mutable & t.pointerfree & isleaftype(t)
end
## # this fixes it:
## function Base.isbits(t::TypeConstructor)
##     false
## end

g() = Integer # needs to return a Type
gg(::Int) = nothing
tmm = first(methods(gg))
sig = Tuple{tmm.sig.parameters[2:end]...}
#    sig = (tmm.sig[2:end]...) # 0.3 version works
fret_typ = Base.return_types(g, sig)
println("$fret_typ")  # without this line no error is thrown later
@show methods(getindex) # error thrown here

This throws the error:

ERROR: LoadError: Shouldn't dispatch to here!

The indexing into the tuple is essential to trigger the bug.
The equivalent in 0.3 works.

(edit: slight update)

[simonster]

This may be the same bug that is breaking examples/typetree.jl:

➜  julia git:(master) ✗ julia examples/typetree.jl 
+- Any << abstract immutable size:0 >>
.  +- => = Pair{A,B} << concrete immutable size:16 >>
.  +- AbstractArray = AbstractArray{T,2} << abstract immutable size:0 >>
.  .  +- AbstractArray{T,1} 
ERROR: LoadError: type TypeConstructor has no field mutable
 in isbits at reflection.jl:75
 in show_delim_array at ./show.jl:216
 in show at ./show.jl:73
 in print_to_string at strings/io.jl:19
 in print_tree at /usr/local/julia/examples/typetree.jl:103
 in print_tree at /usr/local/julia/examples/typetree.jl:108 (repeats 3 times)
 in print_tree at /usr/local/julia/examples/typetree.jl:101
 in include at ./boot.jl:254
 in include_from_node1 at ./loading.jl:263
 in process_options at ./client.jl:308
 in _start at ./client.jl:411
while loading /usr/local/julia/examples/typetree.jl, in expression starting on line 127

[simonster]

Reduced test case:

julia> f(::Type) = "Type";

julia> f(::DataType) = "DataType";

julia> f(Type)
"DataType"

julia> f(AbstractVector)
"DataType"

If the order of the calls is reversed, then the result is correct:

julia> f(::Type) = "Type";

julia> f(::DataType) = "DataType";

julia> f(AbstractVector)
"Type"

julia> f(Type)
"DataType"

[joehuchette]

I'm seeing the type TypeConstructor has no field mutable error, followed by a nice segfault:

ERROR: LoadError: LoadError: MethodError: `copy` has no method matching copy(::Array{JuMP.Variable,1}, ::JuMP.Model)
 in copy at /Users/huchette/.julia/v0.4/JuMP/src/JuMP.jl:258
 in anonymous at /Users/huchette/.julia/v0.4/JuMP/test/model.jl:354
 in facts at /Users/huchette/.julia/v0.4/FactCheck/src/FactCheck.jl:391
 in include at ./boot.jl:259
 in include_from_node1 at ./loading.jl:266
 in include at ./boot.jl:259
 in include_from_node1 at ./loading.jl:266
 in process_options at ./client.jl:308
 in _start at ./client.jl:411fatal: error thrown and no exception handler available.
ErrorException("type TypeConstructor has no field mutable")
rec_backtrace at /Users/huchette/julia/src/task.c:633
jl_vexceptionf at /Users/huchette/julia/usr/lib/libjulia.dylib (unknown line)
jl_errorf at /Users/huchette/julia/usr/lib/libjulia.dylib (unknown line)
jl_field_index at /Users/huchette/julia/usr/lib/libjulia.dylib (unknown line)
jl_f_get_field at /Users/huchette/julia/src/builtins.c:703
isbits at reflection.jl:75
jlcall_isbits_21994 at  (unknown line)
show_delim_array at show.jl:177
show_delim_array at show.jl:164
jlcall_show_delim_array_21991 at  (unknown line)
show at show.jl:74
show_type_parameter at ./show.jl:83
show at ./show.jl:93
jlcall_show_16978 at /Users/huchette/julia/usr/lib/julia/sys.dylib (unknown line)
show_type_parameter at ./show.jl:83
show at ./show.jl:93
jlcall_show_16978 at /Users/huchette/julia/usr/lib/julia/sys.dylib (unknown line)
show at expr.jl:56
show_type_parameter at ./show.jl:83
show at ./show.jl:93
jlcall_show_16978 at /Users/huchette/julia/usr/lib/julia/sys.dylib (unknown line)
print at strings/io.jl:5
print_to_string at ./strings/io.jl:19
string at strings/io.jl:26
show_method_candidates at replutil.jl:247
showerror at replutil.jl:191
showerror at replutil.jl:83
jlcall___showerror#156___27344 at  (unknown line)
jl_apply at /Users/huchette/julia/src/gf.c:1661
julia_showerror_27343 at  (unknown line)
julia_showerror_27342 at  (unknown line)
jlcall_showerror_27342 at  (unknown line)
showerror at replutil.jl:91
julia_showerror_27340 at  (unknown line)
jlcall_showerror_27340 at  (unknown line)
jl_apply at /Users/huchette/julia/src/gf.c:1661
anonymous at client.jl:88
with_output_color at util.jl:313
display_error at client.jl:86
jl_apply at /Users/huchette/julia/src/gf.c:1661
_start at ./client.jl:458
jlcall__start_18534 at /Users/huchette/julia/usr/lib/julia/sys.dylib (unknown line)
true_main at /Users/huchette/julia/julia (unknown line)
main at /Users/huchette/julia/julia (unknown line)

This is from the JuMP tests running in FactCheck; I haven't found a way to isolate the problem.

[vtjnash]

minimalist testcase:

julia> f(::DataType) = 1; f(x::Type) = 2
f (generic function with 2 methods)

julia> f(Vector.body) # this "poisons" the cache for when the identical TypeConstructor shows up later
1

julia> f(Vector) # this should return 2 since Vector is a TypeConstructor not a DataType
1

note that ::Type was necessary on at least one method to prevent

julia/src/gf.c

Line 639 in 8ea2ad7

jl_svecset(newparams, i, decl_i);

from de-specializing the cache method.

plus, we already missed out on some of the TypeConstructor "kind" cache fixup logic just above it because very_general_type(decl_i) returns false for DataType.

[vtjnash]

The test for this is broken and thus disabled. I think I know a fix for it.

[JeffBezanson]

closed by #18457 (I think? Might have missed something here.)

[vtjnash]

no, this isn't a subtyping bug.

[KristofferC]

Code in OP works fine now.

[tkelman]

no, this isn't a subtyping bug.

What is it then?