Re: is it possible to make longjmp-free Lua?

["Juris Kalnins" <juris@...>]

On Mon, 27 Jul 2009 12:35:02 +0300, Jerome Vuarand  
<jerome.vuarand@gmail.com> wrote:

> 2009/7/27 Juris Kalnins <juris@mt.lv>:

> For example lua_newuserdata is guaranteed to return a non-NULL
> pointer. Therefore C code that use that API is unlikely to check the
> validity of that pointer. Another example is lua_touserdata calls in a
> userdata metamethod: the metamethod wouldn't be called if the first
> argument was not a valid userdata, so the return pointer of
> lua_touserdata won't be checked.

Thank you for pointing this out. If I can know likely maximum userdata  
allocation size, it might be possible to add a "yellow zone" to the memory  
allocator, when it both returns allocated memory and raises out-of-memory  
if there is less than specified amount of memory left. (and keep an  
emergency piece of continuous memory for this case).

But C resource handling and Lua exceptions are a tricky thing anyway. It  
is really easy to write resource leaks.
Just an example, from quickly looking at a random library (lposix.c)

static int Pdir(lua_State *L)                   /** dir([path]) */
{
        const char *path = luaL_optstring(L, 1, ".");
        DIR *d = opendir(path);
        if (d == NULL)
                return pusherror(L, path);
        else
        {
                int i;
                struct dirent *entry;
                lua_newtable(L);
                for (i=1; (entry = readdir(d)) != NULL; i++)
                {
                        lua_pushstring(L, entry->d_name);
->                      ^^^^^^ Who is going to close DIR *d, if this call  
runs out of memory?
                        lua_rawseti(L, -2, i);
                }
                closedir(d);
                lua_pushinteger(L, i-1);
                return 2;
        }
}

Having longjmp-free errors would actually _fix_ this. :)

> Lua error mechanism only require setjmp/longjmp, not C++ exceptions.
> Are you sure -fexceptions is needed for setjmp/longjmp to work ? IIRC
> -fexceptions generate data needed for destructors to be called, which
> setjmp/longjmp and thus Lua error mechanism do not need.

But c++ does need it. It runs destructors on function exit. You cannot  
simply discard stack frames in C++. (And see above example for why it is  
bad for C, too).