Potentially suboptimal performance in r2r transforms?

[jishnub]

Currently, the Chebyshev plans internally use FFTW.r2r plans, eg.

FastTransforms.jl/src/chebyshevtransform.jl

Lines 37 to 43 in 59774ab

	function plan_chebyshevtransform(x::AbstractArray{T,N}, ::Val{1}, dims...; kws...) where {T<:fftwNumber,N}
	if isempty(x)
	ChebyshevTransformPlan{T,1,kindtuple(FIRSTKIND,N,dims...),false,N,isempty(dims) ? UnitRange{Int} : typeof(dims)}()
	else
	ChebyshevTransformPlan{T,1,kindtuple(FIRSTKIND,N,dims...)}(FFTW.plan_r2r(x, FIRSTKIND, dims...; kws...))
	end
	end

However, from what I understand after reading Steven G. Johnson's post on discourse, the r2r transforms are not as optimized as the r2c ones, and are not recommended over the alternatives. I wonder if a rewrite of the r2r transforms (here or elsewhere) might improve performance?

[MikaelSlevinsky]

I guess it could be worth benchmarking, though it feels more like an FFTW issue to me.