DiffResults objects are not re-aliased properly

[gdalle]

The DiffResults documentation clearly states that whenever a DiffResult object is used, it must be realiased at the end of the function call:

result = do_stuff!(result, args...)

See for instance the docstrings in https://juliadiff.org/DiffResults.jl/stable/#Mutating-a-DiffResult, or the issue DiffResults#17

This is not done by ForwardDiff, for instance here:

ForwardDiff.jl/src/gradient.jl

Lines 35 to 44 in 42b1554

	function gradient!(result::Union{AbstractArray,DiffResult}, f::F, x::AbstractArray, cfg::GradientConfig{T} = GradientConfig(f, x), ::Val{CHK}=Val{true}()) where {T, CHK, F}
	result isa DiffResult ? require_one_based_indexing(x) : require_one_based_indexing(result, x)
	CHK && checktag(T, f, x)
	if chunksize(cfg) == length(x)
	vector_mode_gradient!(result, f, x, cfg)
	else
	chunk_mode_gradient!(result, f, x, cfg)
	end
	return result
	end

Am I right in deducing that it can lead to incorrectness?

[gdalle]

See also JuliaDiff/ReverseDiff.jl#251

[gdalle]

I can't manage to get an actual bug out of this, presumably because the StaticArrays extension of ForwardDiff overloads the vulnerable parts I've tried. But for ReverseDiff this does lead to wrong behavior

[devmotion]

I noticed that it's actually quite easy to construct an actual bug, one just has to call e.g. ForwardDiff.gradient! with an immutable DiffResult and a non-static input array (since otherwise the code path in the StaticArrays extension is taken):

julia> using ForwardDiff, DiffResults, StaticArrays

julia> result = DiffResults.GradientResult(@SVector zeros(4));

# Uses the buggy implementation pointed out above
julia> ForwardDiff.gradient!(result, sum, ones(4))
ImmutableDiffResult(0.0, ([0.0, 0.0, 0.0, 0.0],))

# The following works since then the implementation in the StaticArrays extension is used
julia> ForwardDiff.gradient!(result, sum, @SVector(ones(4)))
ImmutableDiffResult(4.0, ([1.0, 1.0, 1.0, 1.0],))