Reverse mode incorrectly detecting duplicated returns for a type unstable function

[MasonProtter]

Whatever Enzyme is doing to detect if there are duplicated returns appears to be confused by type instabilities. Not sure if this is intentional or not, but here is a MWE:

using Enzyme

f_unstable = asin
f_stable(x) = asin(x)

function foo_unstable(x)
    y = sin(x)
    z = f_unstable(y) # untyped global variable
    cos(z)
end;

function foo_stable(x)
    y = sin(x)
    z = f_stable(y) # global const
    cos(z)
end;

julia> autodiff(Reverse, foo_stable, Active(1.0))
((-0.8414709848078965,),)

julia> autodiff(Reverse, foo_unstable, Active(1.0))
ERROR: Duplicated Returns not yet handled
Stacktrace:
 [1] autodiff
   @ ~/.julia/packages/Enzyme/haqjK/src/Enzyme.jl:409 [inlined]
 [2] autodiff
   @ ~/.julia/packages/Enzyme/haqjK/src/Enzyme.jl:544 [inlined]
 [3] autodiff(mode::ReverseMode{false, false, FFIABI, false, false}, f::typeof(foo_unstable), args::Active{Float64})
   @ Enzyme ~/.julia/packages/Enzyme/haqjK/src/Enzyme.jl:516
 [4] top-level scope
   @ REPL[29]:1

Note that Enzyme can differentiate this function even with the instability:

julia> let x = 1.0, ydy = Duplicated(Ref(0.0), Ref(0.0))
           autodiff(Reverse, ydy, Active(x)) do y, x
               y[] = foo_unstable(x[])
               y[] 
           end
       end
((nothing, -0.8414709848078965),)

system info:

(jl_GwHTSw) pkg> st
Status `/tmp/jl_GwHTSw/Project.toml`
  [7da242da] Enzyme v0.13.22

julia> versioninfo()
Julia Version 1.11.2
Commit 5e9a32e7af2 (2024-12-01 20:02 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 16 × AMD Ryzen 7 7840U w/ Radeon  780M Graphics
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, znver4)
Threads: 8 default, 0 interactive, 4 GC (on 16 virtual cores)
Environment:
  JULIA_NUM_THREADS = 8
  JULIA_EDITOR = emacsclient

[MasonProtter]

Okay, so update based on what I learned from @wsmoses's comment in #2195, this is only a problem if I don't assert the returned activity:

julia> autodiff(Reverse, foo_unstable, Active, Active(1.0))
((-0.8414709848078965,),)

Maybe we should at least give a hint in the error message as to what's going on here. E.g. something like

The function being differentiated has an inferred return type of Any, but autodiff(Reverse currently only handles functions which return isbits subtypes of Real.

[wsmoses]

Yeah for sure! The relevant part of the code is here:

Enzyme.jl/src/Enzyme.jl

Line 409 in 7c0823f

throw(ErrorException("Duplicated Returns not yet handled"))

(and we can even see in rt what the deduced return type was)