Performance regression in ccall on master and 1.12-beta

[thofma]

We noticed a large increase in runtime and allocations in https://github.com/Nemocas/Nemo.jl, due to the following regression in a simple ccall:

On 1.11.5:

julia> using BenchmarkTools, Nemo # 0.50.0

julia> x = ZZ(1);

julia> f(z) = ccall((:fmpz_mul, Nemo.libflint), Nothing, (Ref{ZZRingElem}, Ref{ZZRingElem}, Ref{ZZRingElem}), z, z, z);

julia> @btime f($x)
  5.997 ns (0 allocations: 0 bytes)

On master and 1.12-beta:

julia> @btime f($x)
  412.824 ns (3 allocations: 48 bytes)

Bisect points to 8bf2802 (@vtjnash)

[vtjnash]

Looks like we didn't keep around the code for Base.cconvert(::Type{Ref{ZZRingElem}}, ::ZZRingElem), leading us to conclude that must be a dynamic dispatch to cconvert(::Type, ::ZZRingElem) instead, which isn't really the intent there.