DiffResults objects are not re-aliased properly #251

Activity

1. gdalle commented on Mar 15, 2024

   

   gdalle

   on Mar 15, 2024

   MemberAuthor

   More actions

   See also JuliaDiff/ForwardDiff.jl#696
2. gdalle

   added

   bug

   on Mar 15, 2024
3. gdalle commented on Mar 15, 2024

   

   gdalle

   on Mar 15, 2024

   MemberAuthor

   More actions

   Here's an MWE demonstrating the bug:

   julia> using StaticArrays, ReverseDiff, DiffResults
   
   julia> x = MVector{2}(3.0, 4.0);
   
   julia> result_before = DiffResults.GradientResult(x)
   ImmutableDiffResult(3.0, ([3.0, 4.0],))
   
   julia> result_after = ReverseDiff.gradient!(result_before, sum, x)
   ImmutableDiffResult(3.0, ([1.0, 1.0],))  # the first value should be 7 = sum(x)
   
   julia> x = [3.0, 4.0];
   
   julia> result_before = DiffResults.GradientResult(x)
   MutableDiffResult(3.0, ([6.365987374e-314, 1.0e-323],))
   
   julia> result_after = ReverseDiff.gradient!(result_before, sum, x)
   MutableDiffResult(7.0, ([1.0, 1.0],))  # the first value is 7 = sum(x)
4. devmotion

   added a commit that references this issue on Sep 16, 2026

   Realias `result` on the Jacobian, Hessian and tuple paths (fixes #251)

   ...

   This issue will close once commit ac2f9c6 is merged into the 'master' branch.

   ac2f9c6
5. devmotion commented on Sep 16, 2026

   

   devmotion

   on Sep 16, 2026

   Member

   More actions

   Fixed by #295, which realiases the whole chain — gradients.jl, jacobians.jl, hessians.jl, tape.jl and utils.jl.

   The rebind has to be unbroken. The leaf extract_result!/extract_result_value! methods in utils.jl already realiased correctly, but every caller discarded the result, so patching any single layer leaves the bug in place.

   Tuple results needed a slightly different fix: those methods looped over result[i], which is not assignable, so the returned struct had nowhere to go. They now map over the tuple instead, which also removes the pattern that allowed the bug — map has no place to drop a return value. The result and input tuples are constrained to a common NTuple{N,Any} so a mismatched arity is a dispatch error rather than a silent truncation. Allocations and inference are unchanged.

   f(x, y) = sum(abs2, x) + sum(abs2, y)
   x, y = MVector{2}(3.0, 5.0), MVector{2}(2.0, 4.0)
   res = (DiffResults.GradientResult(MVector(x)), DiffResults.GradientResult(MVector(y)))
   res = ReverseDiff.gradient!(res, f, (x, y))
   
   map(DiffResults.value, res)     # was (3.0, 2.0), now (54.0, 54.0)
   map(DiffResults.gradient, res)  # ([6.0, 10.0], [4.0, 8.0])
6. devmotion

   added 2 commits that reference this issue on Sep 17, 2026

   Realias `result` on the gradient path (fixes #269)

   ...

   f79e366

   Realias `result` on the Jacobian, Hessian and tuple paths (fixes #251)

   ...

   0da149f
7. devmotion

   closed this as completedin 588d0efon Sep 17, 2026