Cthulu.jl + JET.jl workshop

This repository contains some examples and code snippets. Tested and developed with Julia version 1.12.0-rc3 and:

Cthulhu@v2.17.8
JET@v0.10.7
JuliaSyntax@v1.0.2

Common issues related to --trim

Specific code issues:

1. (vec...,)
   • this operation requires an awkward for-loop code to --trim
   • often causes downstream inference issues due to Tuple{Vararg{...}}
2. [vec_a..., vec_b...]
   • this is "Vector{Int}" all the way down, but poor apply_iterate support breaks this (takeaway: only ever splat Tuples, not Vectors)
   • JuliaLang/julia#57830
3. return (; attributes...)
   • very easy for this to become type-unstable when considering multiple return values
   • best to explicitly write the @NamedTuple{...} type you want to return and use that (similar to an anonymous struct)
4. eval(...) (CPUSummary.jl / HostCPUFeatures.jl are the notable examples in the community)
   • CPUSummary.jl has been fixed already, HostCPUFeatures.jl has not

Common patterns that cause issues:

1. Returning different types based on a runtime value
   • c.f. SciML/OrdinaryDiffEq.jl#2855
2. De-serializing data
   • if it can be done at top-level, this is a "compile-time" operation and (mostly) trimmable for free!
   • be careful about using heavy de-serialization libraries, since the binaries will ship with your --trim code
3. Manual specialization required
   • foo(f::F) where {F <: Function} (versus foo(f::Function))
     • JuliaLang/julia#57660
   • foo(::Type{T}) where T (versus foo(T))
     • JuliaLang/LinearAlgebra.jl#1419
4. (Advanced) Manual de-specialization required
   • a function that only constructs an object (w/ no type-parameters)
     • https://github.com/JuliaLang/julia/blob/c4683c41a8f36123de18a0d68092f23e2427f55c/contrib/juliac/juliac-trim-base.jl#L6
   • a function that constructs a type and then compares it using other polymorphic operations (e.g. === / !== / isa)
     • JuliaLang/julia#56120

Good patterns:

1. Returning a Const(...) value (Tuple / Integer / Symbol / Type) based on a value's type
   • this is the "Tim Holy Trait trick"
2. Make all struct fields "concretely-typed" (i.e. parameterized or a specific concrete type)

Good "natural-feeling" examples:

• TOML stdlib, before --trim fixes (JuliaLang/julia#55016)