RTLD_GLOBAL flag on linux (was Lua 5.1.2-rc5 now available)

["Jose Luis Hidalgo" <joseluis.hidalgo@...>]

Hi All,
 Sorry for the annoyance, but had anybody tested this? I would like
to know if this results are reproducible, I'm really concerned about
Lua and C++ (+RTTI) integration.

BTW I've been told to give a reason to use RTLD_GLOBAL (and here it
is, basically this can lead to invalid C++ library behaviors when
loaded into lua, *only in linux* not in windows) but I can not
understand why it shouldn't be used, reading the documentation about
RTLD_GLOBAL flag I can not see anything dangerous.

I swear not talk about this topic again :)

Cheers,
  Jose L.

*only in linux* == in fact is a GCC > 3.0 "problem"

2007/3/30, Jose Luis Hidalgo <joseluis.hidalgo@gmail.com>:
> > Not now, because this is not a bug. Also, I'm still not convinced that
> > using RTLD_GLOBAL is a good thing, on the contrary, and others seemed
> > to agree. Your application has quite specific needs.
>
> Well, from the mail:
> http://lua-users.org/lists/lua-l/2006-11/msg00210.html
>
> Let's remember that gcc since 3.0 version uses address comparisons to
> determine type equality, so a very big part of the C++ RTTI mechanism
> ( typeid, dynamic_casts, ...) relies on this. I've made a test, you
> can download here:
>
>  http://www.pplux.com/files/lua_RTTI_test.tgz
>
> Basically there is a type "Common<class T>", and as a template it is
> implemented in the header of  "common.hpp", two fictitious libraries
> "m1" and "m2" use the "common" library and instances the template with
> a float. Actually, Common<float> should be the SAME type in both
> dll... but it is not because how libraries where loaded.
>
> This will not be a very big deal, but in Windows lua has the correct
> behavior, (in windows the type check is not done based on the typeid
> pointers). So, for me, different non portable behaviors between
> platforms is some kind of bug...
>
> This is the Example (lua code) :
>
> require("m1") -- loads m1
> require("m2") -- loads m2
> print(m1.info()) -- info about common<float> type (at m1)
> print(m2.info()) -- info about common<float> type (at m2)
>
> obj1 = m1.get() -- get the typeid(common<float>) from m1
> obj2 = m2.get() -- get the typeid(common<float>) from m2
> m2.check(obj1) -- check if typeid are the same (at m2 lib), should be
> "1" for correct RTTI
> m1.check(obj2) -- check if tyepid are the same (at m1 lib)
>
> This is the output :
>
> ************************************************************
> On Windows ( VS8/lua/lua5.1.exe )
> ************************************************************
> Lua 5.1.1  Copyright (C) 1994-2006 Lua.org, PUC-Rio
> > dofile("../../test.lua")
> Creating an object...
> Creating an object...
> m1:typeid = 003AC04C typename = class Common<float> * Base = 003AC06C
> m2:typeid = 0043C04C typename = class Common<float> * Base = 0043C06C
> Check = 1
> Check = 1
> >
> ************************************************************
> On Linux standard lua
> ************************************************************
> Creating an object...
> Creating an object...
> m1:typeid = 0xb7eeb79c typename = P6CommonIfE Base = 0xb7eeb794
> m2:typeid = 0xb7cef79c typename = P6CommonIfE Base = 0xb7cef794
> Check = 0
> Check = 0
> ************************************************************
> On Linux modified lua (RTLD_GLOBAL)
> ************************************************************
> Creating an object...
> Creating an object...
> m1:typeid = 0xb7fa879c typename = P6CommonIfE Base = 0xb7fa8794
> m1:typeid = 0xb7fa879c typename = P6CommonIfE Base = 0xb7fa8794
> Check = 1
> Check = 1
>
> The "Checks" are comparisons between the same types from m1.so ->
> m2.so and from m2.so -> m1.so ... "0" means they are not the same
> type, "1" means they are.  As you can see, the vanilla lua interpreter
> in windows has not the same behavior as Linux, but adding the
> RTLD_GLOBAL everything is correct at both platforms.


--
 Jose L. Hidalgo Valiño (PpluX)
 ---- http://www.pplux.com ----