mixed activity for jl_new_struct

[Crown421]

Hey, I noticed that 0.11.6 seems to have broken the second derivative of metrics from Distances.jl.

The following works with 0.11.5, but not with 0.11.6, 0.11.7 and main

using Distances
using Enzyme
# metric = Euclidean()
metric = SqEuclidean()
metric(1.0, 1.0 + 0.1)

d2metric(x, y) = only(autodiff(
    Forward,
    yt -> only(autodiff_deferred(Forward, metric, DuplicatedNoNeed, Duplicated(x, 1.0), yt)),
    DuplicatedNoNeed,
    Duplicated(y, 1.0)))
d2metric(1.0, 1.0 + 0.1)

The error is

ERROR: Enzyme execution failed.
Enzyme: Not yet implemented, mixed activity for jl_new_struct constants=Bool[1, 0, 0]   %9 = call noalias nonnull {} addrspace(10)* ({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)*, {} addrspace(10)*, ...) @julia.call({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32)* noundef nonnull @jl_f_tuple, {} addrspace(10)* noundef null, {} addrspace(10)* nonnull %1, {} addrspace(10)* nonnull %5, {} addrspace(10)* nonnull %7) #11

Stacktrace:
  [1] throwerr(cstr::Cstring)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:3014
  [2] macro expansion
    @ ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:9661 [inlined]
  [3] enzyme_call
    @ ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:9339 [inlined]
  [4] ForwardModeThunk
    @ ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:9305 [inlined]
  [5] runtime_generic_fwd(activity::Type{Val{(false, false, false, false, true, true)}}, width::Val{1}, RT::Val{NamedTuple{(Symbol("1"), Symbol("2")), Tuple{Any, Any}}}, f::typeof(autodiff_deferred), df::Nothing, primal_1::ForwardMode{FFIABI}, shadow_1_1::Nothing, primal_2::SqEuclidean, shadow_2_1::Nothing, primal_3::Type{DuplicatedNoNeed}, shadow_3_1::Nothing, primal_4::Duplicated{Float64}, shadow_4_1::Duplicated{Float64}, primal_5::Float64, shadow_5_1::Float64)
    @ Enzyme.Compiler ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:1310
  [6] macro expansion
    @ ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:9661 [inlined]
  [7] enzyme_call
    @ ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:9339 [inlined]
  [8] ForwardModeThunk
    @ ~/.julia/packages/Enzyme/dByM2/src/compiler.jl:9305 [inlined]
  [9] autodiff
    @ ~/.julia/packages/Enzyme/dByM2/src/Enzyme.jl:330 [inlined]
 [10] autodiff
    @ ~/.julia/packages/Enzyme/dByM2/src/Enzyme.jl:222 [inlined]

The same happens with forward over reverse

d2metric(x, y) = autodiff(
    Forward,
    (x, y) ->
        first(only(autodiff_deferred(Reverse, metric, Active, Active(x), Const(y)))),
    DuplicatedNoNeed,
    Const(x),
    DuplicatedNoNeed(y, 1.0),
)

Enzyme.API.runtimeActivity!(true) does not help (not sure if it should, but I tried it).

Not sure if this issues should rather be in Distances.jl?

[wsmoses]

Yeah what happened here isn't that v11.6+ broke it, but rather 11.6+ will error that we could get that answer wrong (which 11.5 and before could as well, just we didn't throw the error here like we should have).

In essence this is a type unstable, multiple activity input object construction, which we don't presently handle (even with the runtimeActivity flag) yet.

In the interim, if you make that part of the code type stable, this error should disappear.

[Crown421]

I see, thank you. In this case I will see what can be done on the side of Distances.jl, I somewhat surprised that it isn't already type stable.

[Crown421]

After a little experimentation, I noticed that making it const metric fixes the issue. The type instability is introduced by the metric, which is called via metric(x, y). Only an issue when working in the REPL.

Closing this as there is no more need.