jl_value_ptr function and '&' operator errror on Windows

[joa-quim]

This error came out on the tread

https://groups.google.com/forum/?fromgroups=#!topic/julia-dev/-w5SRAMNSmU

and can be resumed in

julia> type Foo; bar::Int; end
julia> foo = Foo(1)
Foo(1)

julia> ccall(:jl_value_ptr, Ptr{Void}, (Ptr{Foo},), &(foo::Foo))
ERROR: error compiling anonymous: unsupported or misplaced expression & in function anonymous

Apparently this occurs only on Windows. It did with
Version 0.2.0
Commit e912b26

[pao: formatting]

[stevengj]

I can confirm that this code works on MacOS 10.7.5 and on Debian GNU/Linux (x86_64) with commit b66fcaa, so this seems to be a Windows-specific problem.

[pao]

cc @loladiro @vtjnash

[vtjnash]

ah, jl_value_ptr is a special intrinsic function and is not supposed to be used like that. It's only supposed to be called through the pointer_from_objref function. the error detected on windows is correct.

[stevengj]

This has nothing to do with jl_value_ptr (which is just an identity function, not an intrinsic last I checked, and gets no special treatment in ccall as far as I can tell). The error occurs in the parser before any function is called.

[vtjnash]

ccall tries to inline that function (jl_value_ptr) -- but it is currently broken/disabled

[stevengj]

I see, ccall.cpp has special-case code:

    if (fptr == &jl_value_ptr) {
        assert(lrt->isPointerTy());
        Value *ary = emit_expr(args[4], ctx);
        JL_GC_POP();
        return mark_julia_type(builder.CreateBitCast(ary,lrt),rt);
    }

My inclination would be to remove this code, and just add a pointerfromref intrinsic (like pointertoref), since what we really want here is an intrinsic function and not a "blessed" C API function.

The underlying issue is that ccall can do magical things that ordinary Julia code cannot. It can convert jl_value_t* pointers to/from references, and it can use & to get pointers to struct-like member data. An identity function in C (like jl_value_ptr) provides a workaround to access this functionality from Julia.

Right now, PyCall is using jl_value_ptr to get access to the &foo pointer, in order to insert a pointer to a struct (Julia composite type) in another struct. A much cleaner solution will be provided by #2818 once it lands, but until then the above special-case code apparently breaks this on Windows.

[vtjnash]

I see it is also causing this test to fail: @test unsafe_pointer_to_objref(pointer_from_objref(e+pi)) == e+pi, presumably because the special-case code assumes that the object is actually an object (but is missing a call to boxed)