[<Fact>]
    let ``nested push32`` () =
        let var = ref None
        let var2 = ref None
        let var3 = ref None
        let arg1 = ref <| Some "arg1"
        let arg2 = ref <| Some "arg2"

        let type_here = Push32 <| I 32
        // now we intro the push32 type who's variable-sized!
        let type_here = Push32 <| type_here

        // let's design a vivid case:
        // head: i32 = 2
        //   - subhead1: i32 = 3
        //       1 : i32
        //       2 : i32
        //       3 : i32
        //   - subhead2: i32 = 2
        //       1 : i32
        //       2 : i32
        // the size of them (in bytes) is
        // 3 * 4(3 i32 heads) + 5 * 4(3 + 2 elems) = 32
        // so we should allocate 32/4 = 8
        let num = 8

        let body = Suite [

            // allocate 2 to gain a 12-bytes memory piece(4 + 8).
            Tmp(var, Alloca(type_here, Some num))

             //prepare data
            Let(var2,
                Bitcast(Get var, Ptr <| I 32),
                let prepare_data a b =
                    Store(GEP(Get var2, integer 32 a, []),
                          integer 32 b)
                Suite [
                    prepare_data 0 2 // head = 2
                    prepare_data 1 3 // subhead1 = 3
                    prepare_data 2 10 // subhead[1] = 10
                    prepare_data 3 20 // subhead[2] = 20
                    prepare_data 4 30 // subhead[3] = 30
                    prepare_data 5 2 // subhead2 = 2
                    prepare_data 6 40 // subhead2[1] = 40
                    prepare_data 7 50 // subhead2[2] = 50
                ])

            (let index_fst_level =
                GEPEx(Get var, integer 32 0, Some <| Get arg1)
             let index_snd_level =
                GEPEx(index_fst_level, integer 32 0, Some <| Get arg2)
             Load index_snd_level)
            ]
        let defun = Defun("my_func2", ["arg1", I 32; "arg2", I 32], I 32, body)
        let defun = Suite [
               Decl("printf", [Ptr <| I 8; I 32], I 32)
               defun

            ]
        let my_module =
            try
            h.compile defun
            |> h.Parse
            with LLException e ->
                failwithf "%A" e
        //h.Optimize my_module
        //printf1 "%A" my_module
        h.Load my_module
        let fp = h.FindSymbol "my_func2"
        let f = Marshal.GetDelegateForFunctionPointer(fp.Address, typeof<intint2int>) :?> intint2int

        for i = 0 to 1 do
            for j = 0 to 2 do
                printf1 "%d\n" <| f.Invoke(i, j)

        let res = f.Invoke(0, 0) //subhead1[1]
        Assert.Equal(res, 10)

        let res = f.Invoke(0, 1) //subhead1[2]
        Assert.Equal(res, 20)

        let res = f.Invoke(0, 2) //subhead1[3]
        Assert.Equal(res, 30)

        let res = f.Invoke(1, 0) //subhead2[1]
        Assert.Equal(res, 40)

        let res = f.Invoke(1, 1) //subhead2[2]
        Assert.Equal(res, 50)

        h.Unload my_module