Unreachable reached

[englhardt]

I am running into a unreachable error when executing the following script bug.jl.
This issue seems similar to #31439 and #31406. However in my case, only Julia 1.0.3 works correctly, and the bug already appears with Julia 1.1.0 and also fails on the latest nightly 1.2.0-DEV.557 (Commit 8649af9abe). This is different from the bug in #31439 that works with 1.1.0 and only fails on 1.2.0-DEV

bug.jl

using GaussianMixtures

struct GMMWrapper
    gmm::GMM
end

function sample_gmm(g::GMMWrapper, n::Int)
    return rand(g.gmm, n)
end

g = GMMWrapper(rand(GMM, 1, 2))

# works
@show rand(g.gmm, 10)
# fails
@show sample_gmm(g, 10)

Output:

rand(g.gmm, 10) = [-2.1182340176250976 1.630642305577004; -1.5943359590743922 0.026937477698993795; -2.022431014003848 2.3052428001268304; -1.2051696749634346 0.7670988215906123; -2.2556905707970296 1.4146400755937694; 0.05866609286796365 -0.45946366339887723; -1.5514045527686888 1.4352168554403033; -2.1021557548997785 -0.2619369172736933; -0.58429883008154 -0.3474937399220318; -4.191643913606173 3.6064676997298957]
Unreachable reached at 0x7f5d4dec3eb4

signal (4): Illegal instruction
in expression starting at xxxxx/bug.jl:19
sample_gmm at xxxxx/bug.jl:11
unknown function (ip: 0x7f5d4dec3ed7)
jl_fptr_trampoline at /buildworker/worker/package_linux64/build/src/gf.c:1871
jl_apply_generic at /buildworker/worker/package_linux64/build/src/gf.c:2226
do_call at /buildworker/worker/package_linux64/build/src/interpreter.c:323
eval_value at /buildworker/worker/package_linux64/build/src/interpreter.c:411
eval_stmt_value at /buildworker/worker/package_linux64/build/src/interpreter.c:362 [inlined]
eval_body at /buildworker/worker/package_linux64/build/src/interpreter.c:772
jl_interpret_toplevel_thunk_callback at /buildworker/worker/package_linux64/build/src/interpreter.c:884
unknown function (ip: 0xfffffffffffffffe)
unknown function (ip: 0x7f5d48872b0f)
unknown function (ip: 0xffffffffffffffff)
jl_interpret_toplevel_thunk at /buildworker/worker/package_linux64/build/src/interpreter.c:893
jl_toplevel_eval_flex at /buildworker/worker/package_linux64/build/src/toplevel.c:797
jl_parse_eval_all at /buildworker/worker/package_linux64/build/src/ast.c:873
jl_load at /buildworker/worker/package_linux64/build/src/toplevel.c:859
include at ./boot.jl:325 [inlined]
include_relative at ./loading.jl:1042
include at ./Base.jl:31
jl_apply_generic at /buildworker/worker/package_linux64/build/src/gf.c:2226
exec_options at ./client.jl:295
_start at ./client.jl:464
jl_apply_generic at /buildworker/worker/package_linux64/build/src/gf.c:2226
jl_apply at /buildworker/worker/package_linux64/build/ui/../src/julia.h:1594 [inlined]
true_main at /buildworker/worker/package_linux64/build/ui/repl.c:96
main at /buildworker/worker/package_linux64/build/ui/repl.c:217
__libc_start_main at /usr/lib/libc.so.6 (unknown line)
_start at /home/adrian/Downloads/julia-8649af9abe/bin/julia (unknown line)
Allocations: 15082729 (Pool: 15079560; Big: 3169); GC: 33
[1]    19165 illegal hardware instruction (core dumped)  ~/Downloads/julia-8649af9abe/bin/julia bug.jl

Version info:

julia> versioninfo()
Julia Version 1.2.0-DEV.557
Commit 8649af9abe (2019-03-26 17:56 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i5-7300U CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, skylake)

[JeffBezanson]

Possible reduction:

DiagCov{T} = AbstractArray{T,2}
FullCov{T} = Vector{UpperTriangular{T,Matrix{T}}}
CovType{T} = Union{DiagCov{T}, FullCov{T}}

typeintersect(Pair{<:Any, <:AbstractMatrix},
              Pair{T,     <:CovType{T}} where T<:AbstractFloat)  # empty

[Jutho]

Not sure if this one is related, the error is caused by deepcopy itself!

However, I have not been able to reduce it to a simple test case using only Base code. It occurs in TensorKit.jl. Errors from julia 0.7 to latest freshly built master.

julia> using TensorKit

julia> W = TensorMap(rand, Float64, SU2Space(0=>3), SU2Space(0=>3))
julia> d1 = Base.deepcopy_internal(W.data, IdDict()) # works
julia> d2 = Base.deepcopy_internal(W.data, IdDict())::typeof(W.data) # works
julia> d3 = deepcopy(W.data)
Unreachable reached at 0x118bd9820

signal (4): Illegal instruction: 4
in expression starting at REPL[5]:1
deepcopy at ./deepcopy.jl:28
do_call at /usr/local/julia/src/interpreter.c:323
eval_stmt_value at /usr/local/julia/src/interpreter.c:362 [inlined]
eval_body at /usr/local/julia/src/interpreter.c:758
jl_interpret_toplevel_thunk_callback at /usr/local/julia/src/interpreter.c:884
unknown function (ip: 0xfffffffffffffffe)
unknown function (ip: 0x11a770c8f)
unknown function (ip: 0x0)
jl_interpret_toplevel_thunk at /usr/local/julia/src/interpreter.c:893
jl_toplevel_eval_flex at /usr/local/julia/src/toplevel.c:797
jl_toplevel_eval_flex at /usr/local/julia/src/toplevel.c:746
jl_toplevel_eval at /usr/local/julia/src/toplevel.c:806 [inlined]
jl_toplevel_eval_in at /usr/local/julia/src/toplevel.c:826
eval at ./boot.jl:330
eval_user_input at /usr/local/julia/usr/share/julia/stdlib/v1.3/REPL/src/REPL.jl:86
macro expansion at /usr/local/julia/usr/share/julia/stdlib/v1.3/REPL/src/REPL.jl:118 [inlined]
#26 at ./task.jl:268
jl_apply at /usr/local/julia/src/./julia.h:1604 [inlined]
start_task at /usr/local/julia/src/task.c:583
Allocations: 10811848 (Pool: 10809461; Big: 2387); GC: 24
Illegal instruction: 4

[Jutho]

Also

julia> T = typeof(W.data)
TensorKit.SortedVectorDict{SU₂,Array{Float64,2}}

julia> Base.return_types(Base.deepcopy_internal, Tuple{T, IdDict}) # ok
1-element Array{Any,1}:
 TensorKit.SortedVectorDict{SU₂,Array{Float64,2}}

julia> Base.return_types(deepcopy, Tuple{T})
1-element Array{Any,1}:
 Union{}

[Jutho]

But

julia> Base.return_types(Base.deepcopy_internal, Tuple{T, IdDict{Any,Any}})
1-element Array{Any,1}:
 Union{}

[Jutho]

Ok, I debugged this down to a method ambiguity in my own code with one in base:

julia> get(IdDict(), d1, :x)
ERROR: MethodError: get(::IdDict{Any,Any}, ::TensorKit.SortedVectorDict{SU₂,Array{Float64,2}}, ::Symbol) is ambiguous. Candidates:
  get(f, d::TensorKit.SortedVectorDict{K,V} where V, k) where K in TensorKit at /Users/jhaegemn/.julia/dev/TensorKit/src/auxiliary/dicts.jl:187
  get(d::IdDict{K,V}, key, default) where {K, V} in Base at abstractdict.jl:594
Possible fix, define
  get(::IdDict{K,V}, ::TensorKit.SortedVectorDict{K,V} where V, ::Any)

my apologies for the noise. Maybe something to be weary about if one defines new AbstractDict subtypes.

[JeffBezanson]

Hmm, it should not cause a crash though.

[Jutho]

Yes, an error message would have been helpful :-)
Compare

julia> get(IdDict(), d1, :x)
ERROR: MethodError: get(::IdDict{Any,Any}, ::TensorKit.SortedVectorDict{SU₂,Array{Float64,2}}, ::Symbol) is ambiguous. Candidates:
  get(f, d::TensorKit.SortedVectorDict{K,V} where V, k) where K in TensorKit at /Users/jhaegemn/.julia/dev/TensorKit/src/auxiliary/dicts.jl:187
  get(d::IdDict{K,V}, key, default) where {K, V} in Base at abstractdict.jl:594
Possible fix, define
  get(::IdDict{K,V}, ::TensorKit.SortedVectorDict{K,V} where V, ::Any)

versus

julia> methods(Base.get,Tuple{IdDict{Any,Any}, T, Symbol})
# 0 methods for generic function "get":

and from this

julia> Base.return_types(Base.get, Tuple{IdDict{Any,Any}, T, Symbol})
0-element Array{Any,1}

and thus

julia> code_typed(in, Tuple{T,Base.KeySet{Any,IdDict{Any,Any}}})
1-element Array{Any,1}:
 CodeInfo(
1 ─ %1 = Base.getfield(v, :dict)::IdDict{Any,Any}
│        Base.get(%1, k, Base.secret_table_token)::Union{}
└──      $(Expr(:unreachable))::Union{}
) => Union{}

However, this still works:

julia> in(d1, keys(IdDict()))
false

but this does not

julia> haskey(IdDict(), d1)
Segmentation fault: 11