invalid usage of @assume_effects in base/irrationals.jl

[nsajko]

There are four methods with @assume_effects :total applied to the entire method in base/irrationals.jl:

julia/base/irrationals.jl

Line 55 in 6e5e87b

@assume_effects :total function Rational{T}(x::AbstractIrrational) where T<:Integer

julia/base/irrationals.jl

Lines 70 to 71 in 6e5e87b

	@assume_effects :total function (t::Type{T})(x::AbstractIrrational, r::RoundingMode) where T<:Union{Float32,Float64}

julia/base/irrationals.jl

Lines 113 to 120 in 6e5e87b

	@assume_effects :total function rationalize(::Type{T}, x::AbstractIrrational; tol::Real=0) where T
	return rationalize(T, big(x), tol=tol)
	end
	@assume_effects :total function lessrational(rx::Rational{<:Integer}, x::AbstractIrrational)
	# an @assume_effects :total version of `<` for determining if the rationalization of
	# an irrational number required rounding up or down
	return rx < big(x)
	end

Issues:

• The :total effect is too powerful and not future-proof. I guess what's actually desired is just :foldable.
• It's not valid to apply @assume_effects to the ::AbstractIrrational methods, because users may define custom subtypes of AbstractIrrational, and thus calls like BigFloat(::AbstractIrrational) or big(::AbstractIrrational) may execute arbitrary code. That said ...
• It's not valid to apply @assume_effects even just for ::Irrational, because users may define custom subtypes of Irrational! Example:

# in a user package
struct IrrationalSymbol end
function big(::Irrational{IrrationalSymbol})
    execute_arbitrary_code()
end

This seems perfectly legal, or at least it's not type piracy.

The solution seems simple: define a Union of all known irrational constants, something like the following, then dispatch on it for methods that have @assume_effects, instead of on AbstractIrrational or on Irrational:

const _KnownIrrational = (ℯ, π, MathConstants.γ, MathConstants.catalan, MathConstants.φ)

This will ensure calls involving the known constants are foldable, and allow the unsafe @assume_effects annotations to be deleted. Will make a PR later.

[aviatesk]

After looking at the blame, it seems that those @assume_effects were added by me in #44776 as replacements for the previous @pure, but I agree that @assume_effects :total was too strong. It should have been replaced with :foldable, and we also need to restrict the annotation to known concrete types as suggested.

[vtjnash]

It isn't technically valid to be foldable either (there isn't technically any valid annotations to assume here), as it references global state (the precision) and returns a new allocation (a BigNum), but at least any reduction in that set of invalid annotations seems beneficial

[nsajko]

references global state (the precision)

I think this could at least partly be addressed by using setprecision(::Function, ::Type{BigFloat}, ::Int). One such PR is #55853.