docs: improve runtime activity FAQ and fix MWE

[gdalle]

The function in the MWE was missing an end, so I made it executable to prevent this kind of bug (I didn't go for a full doctest because of the exception stack trace)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 74.92%. Comparing base (3a6776f) to head (ee2fd98).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2559   +/-   ##
=======================================
  Coverage   74.92%   74.92%           
=======================================
  Files          56       56           
  Lines       17437    17437           
=======================================
  Hits        13065    13065           
  Misses       4372     4372           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[gdalle]

Running into JuliaDocs/Documenter.jl#1420, will fix

[wsmoses]

let's still call it condition to not confuse the name of the function parameter from the input to the call

[wsmoses]

add the comment "and pointer-like"

[gdalle]

I wanted to ask: what is the precise definition of "pointer-like" in Julia terms, and what does it mean for two objects to be aliased in that way? I thought that it meant either a === b or pointer(a) == pointer(b), as your docs suggest, but I found out that these conditions are not equivalent, and that the latter can only be checked for Arrays and a few other concrete types. See JuliaDiff/DifferentiationInterface.jl#666 (comment) for details and JuliaDiff/DifferentiationInterface.jl#850 for my current attempt in DI.

[wsmoses]

whether the layout of the julia object in memory is a pointer

[gdalle]

Okay, then what is the proper way to test aliasing for Enzyme's runtime activity feature in general? Because === seems wrong based on my tests and pointer works for a very limited set of types

[wsmoses]

not sure offhand julia-side, I'm guessing there's some feature for determining type layout

[gdalle]

For instance, in this case, the check a === b doesn't suffice to detect aliasing:

julia> x = [1.0];

julia> mutable struct S
           x::Vector{Float64}
       end

julia> a = S(x);

julia> b = S(x);

julia> a === b
false

[wsmoses]

convert(LLVMType, julia_type; allow_boxed=true) isa LLVM.PointerType ought work always but not an answer you want @gdalle

[wsmoses]

Base.DataTypeFieldDesc may also be helpful here

[gdalle]

convert(LLVMType, julia_type; allow_boxed=true) isa LLVM.PointerType ought work always but not an answer you want @gdalle

How am I supposed to use it?

julia> using LLVM

julia> convert(LLVMType, Vector{Float64}; allow_boxed=true)
ERROR: No LLVM context is active
Stacktrace:
 [1] error(s::String)
   @ Base ./error.jl:35
 [2] context(; throw_error::Bool)
   @ LLVM ~/.julia/packages/LLVM/UFrs4/src/state.jl:22
 [3] context
   @ ~/.julia/packages/LLVM/UFrs4/src/state.jl:20 [inlined]
 [4] convert(::Type{LLVMType}, typ::Type; allow_boxed::Bool)
   @ LLVM.Interop ~/.julia/packages/LLVM/UFrs4/src/interop/base.jl:73
 [5] top-level scope
   @ REPL[25]:1

Base.DataTypeFieldDesc may also be helpful here

I don't know what this does and it seems private?

help?> Base.DataTypeFieldDesc
  │ Warning
  │
  │  The following bindings may be internal; they may change or be removed in future versions:
  │
  │    •  Base.DataTypeFieldDesc

  No documentation found for private symbol.

Besides, users would also need a way to compare the locations to which these "pointer-like" objects point, in order to detect aliasing

[gdalle]

I'd be happy if Enzyme provided Enzyme.is_pointerlike(x::Any) and Enzyme.are_aliased(x::T, y::T) where {T} with the precise semantics that you want, to use in runtime activity correctness checking. Expecting the users to go through this level of sophistication seems doomed to fail

Have opened #2587 to keep track so that we may merge this docs fix