A lot of functions in linalg/factorization.jl (and perhaps elsewhere in LinAlg) have methods like:

function eigfact!{T<:BlasReal}(A::StridedMatrix{T}) ... end
function eigfact!{T<:BlasComplex}(A::StridedMatrix{T}) ... end
eigfact!(A::StridedMatrix) = eigfact!(float(A))

This gives a stack overflow if float(A) does not return a BlasReal or BlasComplex, e.g. if eltype(A) is Float16 or BigFloat.