Is it possible to handle headers with __float128?

[amontoison]

Some Fortran libraries, like CUTEst, can be easily compiled in quadruple precision with the gfortran compiler.
The associated type in the header file to call the routines from C is __float128 and is equivalent to Float128 (implemented in Quadmath.jl).

It will be a nice feature to add an option to generate Julia wrappers for this precision. We just need to add using Quadmath at the top of the generated file if Julia wrappers can be generated.

I have a package where I need to write these additional wrappers by hand:

• https://github.com/JuliaSmoothOptimizers/CUTEst.jl/tree/main/gen
• https://github.com/JuliaSmoothOptimizers/CUTEst.jl/blob/main/src/libcutest.jl

[Gnimuc]

have you tried @add_def __float128 before creating the context?

[Gnimuc]

looks like we need to uncomment this line

Clang.jl/src/generator/jltypes.jl

Line 145 in 9c77fd6

# tojulia(x::CLFloat128) = JuliaCfloat128() # see Quadmath.jl

[amontoison]

The Julia type for __float128 is Cfloat128:
https://github.com/JuliaMath/Quadmath.jl/blob/master/src/Quadmath.jl#L82

Float128 is working with Fortran routines because we pass everything by reference but if we pass by value, Cfloat128 is needed.

The macro @ccall is also modified if a Cfloat128 is returned:
https://github.com/JuliaMath/Quadmath.jl/blob/master/src/Quadmath.jl#L68

So it should be safe to handle all headers with __float128 in Clang.jl if the routines don't return a __float128.

[Gnimuc]

is there a general agreement of defining Cfloat128 as NTuple{2,VecElement{Float64}}, but not a primitive type or an alternative reprensentationNTuple{2,Float64}, etc?

if not, package authors should overload translate(jlty:: JuliaCfloat128, options=Dict()) = :TheirCfloat128Symbol to control what should be generated.

[amontoison]

@ViralBShah Should we add a definition of Float128 / Cfloat128 like Float16 in the Julia language directly so that we have a general agreement on how we should represent this type in Julia?

[ViralBShah]

It may be worth bringing up on discourse. For example with BFloat16s.jl, the preference has been to keep it outside to allow for flexibility.