Error from `has_chain_rrule` on an array of `Bool`

[mcabbott]

The bug reported here https://discourse.julialang.org/t/help-me-pin-this-bug-in-flux/81901 can be simplified a little further. Note that the error occurs on the forward pass, and requires an array of booleans:

julia> z, back = pullback([true, false]) do x
           y2 = reshape(x, 1, 2, 1)
           y3 = repeat(y2, 3)
           y3[1]
       end
ERROR: type Nothing has no field method
Stacktrace:
  [1] getproperty(x::Nothing, f::Symbol)
    @ Base ./Base.jl:37
  [2] matching_cr_sig(t::IRTools.Inner.Meta, s::Nothing)
    @ Zygote ~/.julia/packages/Zygote/DkIUK/src/compiler/chainrules.jl:72
  [3] has_chain_rrule(T::Type)
    @ Zygote ~/.julia/packages/Zygote/DkIUK/src/compiler/chainrules.jl:54
  [4] #s53#1205
    @ ~/.julia/packages/Zygote/DkIUK/src/compiler/interface2.jl:20 [inlined]
  [5] var"#s53#1205"(::Any, ctx::Any, f::Any, args::Any)
    @ Zygote ./none:0
  [6] (::Core.GeneratedFunctionStub)(::Any, ::Vararg{Any})
...

julia> gradient([true, false]) do x
           y2 = reshape(x, 1, 2, 1) .+ 0.0
           y3 = repeat(y2, 3)
           y3[1]
       end
(nothing,)

julia> gradient([1, 0]) do x
           y2 = reshape(x, 1, 2, 1)
           y3 = repeat(y2, 3)
           y3[1]
       end
([1.0, 0.0],)

In testing variants, I also got this error, which is unrelated I think. But similar to JuliaDiff/ChainRules.jl#627

julia> gradient([true, false]) do x
           y2 = reshape(x, 1, 2, 1)
           @show summary(y2)
           y3 = repeat(y2, outer=3)
           @show summary(y3)
           y3[1]
       end
summary(y2) = "1×2×1 Array{Bool, 3}"
ERROR: ArgumentError: number of outer repetitions (1) cannot be less than number of dimensions of input array (3)
Stacktrace:
  [1] check(arr::Array{Bool, 3}, inner::Tuple{Int64, Int64, Int64}, outer::Int64)
    @ Base._RepeatInnerOuter ./abstractarraymath.jl:453
  [2] repeat(arr::Array{Bool, 3}; inner::Tuple{Int64, Int64, Int64}, outer::Int64)
    @ Base._RepeatInnerOuter ./abstractarraymath.jl:399
  [3] repeat(A::Array{Bool, 3}; inner::Tuple{Int64, Int64, Int64}, outer::Int64)
    @ Base ./abstractarraymath.jl:393
  [4] #adjoint#511
    @ ~/.julia/packages/Zygote/DkIUK/src/lib/array.jl:120 [inlined]
  [5] _pullback
    @ ~/.julia/packages/ZygoteRules/AIbCs/src/adjoint.jl:71 [inlined]
...

julia> Zygote.gradient(rand(1,2,1)) do x
         repeat(x, outer=3)[1]
       end
ERROR: ArgumentError: number of outer repetitions (1) cannot be less than number of dimensions of input array (3)

[ToucheSir]

Somehow, https://github.com/FluxML/IRTools.jl/blob/v0.4.6/src/reflection/reflection.jl#L48-L50 is being triggered by https://github.com/FluxML/Zygote.jl/blob/v0.6.40/src/compiler/chainrules.jl#L24 or https://github.com/FluxML/Zygote.jl/blob/v0.6.40/src/compiler/chainrules.jl#L20. I don't understand this mechanism well enough to know whether we can simply say nothing implies a non-match, or if it truly is an exceptional circumstance higher up the stack.

[ToucheSir]

@oxinabox can we call on your expertise to make sense of this?

[oxinabox]

I don't really understand what is going on, but it seems like it is saying according to IRTools there are no methods.
As such it can't possibly match a relevent chainrule.
So we can introduce

matching_cr_sig(::Any, ::Nothing) = false  # https://github.com/FluxML/Zygote.jl/issues/1234

[ToucheSir]

Running with that method helps highlight what might be the underlying cause of the IRTools lookup failure:

julia> z, back = pullback([true, false]) do x
                  y2 = reshape(x, 1, 2, 1)
                  y3 = repeat(y2, 3)
                  y3[1]
              end
ERROR: MethodError: rrule(::typeof(repeat), ::Array{Bool, 3}, ::Int64) is ambiguous. Candidates:
  rrule(::typeof(repeat), xs::AbstractArray, counts::Integer...) in ChainRules at /home/brianc/.julia/packages/ChainRules/o1vND/src/rulesets/Base/array.jl:191
  rrule(::typeof(repeat), var"543"::AbstractArray{Bool}, var"544"...; repeat_pullback) in ChainRules at /home/brianc/.julia/packages/ChainRules/o1vND/src/rulesets/Base/nondiff.jl:65
Possible fix, define
  rrule(::typeof(repeat), ::AbstractArray{Bool}, ::Vararg{Integer})
Stacktrace:
 [1] rrule(::Zygote.ZygoteRuleConfig{Zygote.Context{false}}, ::Function, ::Array{Bool, 3}, ::Int64)
   @ ChainRulesCore ~/.julia/packages/ChainRulesCore/ctmSK/src/rules.jl:134
 [2] chain_rrule
   @ ~/projects/julia-dev/Zygote/src/compiler/chainrules.jl:218 [inlined]
 [3] macro expansion
   @ ~/projects/julia-dev/Zygote/src/compiler/interface2.jl:0 [inlined]
 [4] _pullback(::Zygote.Context{false}, ::typeof(repeat), ::Array{Bool, 3}, ::Int64)
   @ Zygote ~/projects/julia-dev/Zygote/src/compiler/interface2.jl:9
 [5] _pullback
   @ ./REPL[2]:3 [inlined]
 [6] _pullback(ctx::Zygote.Context{false}, f::var"#1#2", args::Vector{Bool})
   @ Zygote ~/projects/julia-dev/Zygote/src/compiler/interface2.jl:0
 [7] pullback(f::Function, cx::Zygote.Context{false}, args::Vector{Bool})
   @ Zygote ~/projects/julia-dev/Zygote/src/compiler/interface.jl:44
 [8] pullback(f::Function, args::Vector{Bool})
   @ Zygote ~/projects/julia-dev/Zygote/src/compiler/interface.jl:42
 [9] top-level scope
   @ REPL[2]:1

This is quite a bit more informative than an opaque error on getproperty! Looks like CR needs an update as well.

[mcabbott]

Ah OK. Then this is similar to #1263 (comment)

[oxinabox]

ah nice, this is because we basically surpress the error here saying nothing is wrong, then generate code that calls the ambigious method and errors.
We could alternatively error here ourself, but we wouldn't get the nice method error