Failure of traversing through nested structures

[zuhengxu]

As described in TuringLang/NormalizingFlows.jl#53 (comment)

[zuhengxu]

I believe the bugs are introduced in Mooncake v0.4.124 (see TuringLang/NormalizingFlows.jl#53 (comment))

[yebai]

It is likely caused by this change

https://github.com/chalk-lab/Mooncake.jl/pull/594/files#diff-0163b66510af82fcd75ca009591262b191accc7873909a2e7f3b12dfef7976cbR36-R55

@zuhengxu / @MilesCranmer anything obvious that could be incorrect?

[MilesCranmer]

"Unreachable reached" usually points to a Julia bug: https://discourse.julialang.org/t/unreachable-reached-at-0x13929e935/38528/6. It is possible it just appeared from that commit but might not be directly related.

In this code:

Mooncake.jl/src/rrules/new.jl

Lines 36 to 57 in a26b5c3

	@inline function build_fdata(::Type{P}, x::Tuple, fdata::Tuple) where {P}
	return _build_fdata_cartesian(P, x, fdata, Val(fieldcount(P)), Val(fieldnames(P)))
	end
	@generated function _build_fdata_cartesian(
	::Type{P}, x::Tuple, fdata::Tuple{Vararg{Any,N}}, ::Val{nfield}, ::Val{names}
	) where {P,N,nfield,names}
	quote
	processed_fdata = Base.Cartesian.@ntuple(
	$nfield, n -> let
	F = fdata_field_type(P, n)
	if n <= $N
	data = __get_data(P, x, fdata, n)
	F <: PossiblyUninitTangent ? F(data) : data
	else
	F()
	end
	end
	)
	F_out = fdata_type(tangent_type(P))
	return F_out(NamedTuple{$names}(processed_fdata))
	end
	end

nothing looks problematic in terms of generated function rules. The only thing being used in the generated function construction is Val(n) for n the number of fields, and Val(names) for names the names of those fields. So I don't think that specifically is an issue. Although "Unreachable reached" might be caused by something else and that particular change just caused a particular branch of code to get executed in Julia.

@zuhengxu might be useful to run this on previous versions of Julia and see what the bisect points to there?

[yebai]

I confirmed changes to build_fdata in #594 introduced this bug. Reverting related changes in #663 fixes example in TuringLang/NormalizingFlows.jl#53 (comment)

[MilesCranmer]

I don't think this is fixing the issue though, it's just temporarily concealing a symptom.

The function itself has no visible issues, and also this bug is related to undefined behavior, so it's either caused by something else in the code, or is a Julia bug. It just so happens that a certain combination of Julia code (for this particular Julia version, etc.) brings out the symptom. But the issue is still lurking somewhere

[yebai]

I agree; the changes to build_fdata in #594 are also necessary for DynamicExpressions to work correctly, so we cannot simply revert them.

The bug is, however, likely beyond my depth, so it requires someone with deeper Julia knowledge to figure out.

[MilesCranmer]

Right; also the changes to build_fdata were themselves a fix of world age issues seen elsewhere. It looks like those regression tests of world age issues are now correctly failing after the draft revert

[MilesCranmer]

I would recommend making an issue in the JuliaLang repo which is what Stefan recommends in https://discourse.julialang.org/t/unreachable-reached-at-0x13929e935/38528/6

[zuhengxu]

Thank you @yebai @MilesCranmer for taking into this.

might be useful to run this on previous versions of Julia and see what the bisect points to there?

I can reproduce the bug on julia v1.11.3 with Mooncake v0.4.124. I'm not sure if I can pinpoint the exact julia version/code that triggers the issue, which is beyond my knowledge. I'm going to open an issue in JuliaLang as suggested by @MilesCranmer.

[vchuravy]

I would recommend making an issue in the JuliaLang repo

It's less clear for packages like Enzyme and Mooncake that do shenanigans beyond the Julia compiler. A JuliaLang base issue should only be opened once their is a reproducer "without funny business"

[yebai]

This is fixed by #663.