The stubs mechanism encapsulates the interface to a library in a stubs table to provide portable dynamic linking. Instead of linking directly against a shared object, the client links against a smaller static archive containing the stubs table and an initialization function, and then calls the initialization function when it loads the shared object. Tcl, Tk, and other extensions and libraries provide stubs tables.
A stubs table is a C structure whose members are the functions that comprise the interface to a shared library. This stubs table is statically linked into the client along with an initialization function which the client calls to fill in the table after it loads the library. The client then calls the functions through this table.
This process bypasses the system linker and is therefore not subject to any constraints it imposes. For example, name and version conflicts between shared objects and difficulties surrounding versioned symbols might be avoided. This also makes it possible for a Tcl extension to call functions in the library that loaded it even if the system linker doesn't provide this capability.
To use the stubs mechanism the client defines USE_TCL_STUBS, includes tcl.h which redefines all Tcl_* function calls as lookups into the Tcl stubs table, links against the static Tcl stubs library which includes the stubs table and Tcl_InitStubs(), and then calls Tcl_InitStubs() after loading the library.
When Tcl loads an extension it calls Extensionname_Init, passing it the the current interpreter. The extension can then use this interpreter to call Tcl_InitStubs.
When a program embeds Tcl, it can link against the Tcl stubs library and then call Tcl_InitStubs() to get access to the Tcl C interface. It must do some work in order to get a handle on an interpreter it can use to call Tcl_InitStubs. There are examples at How to embed Tcl in C applications.
The stubs mechanism has the following advantages:
A shared object that doesn't use stubs to call interface functions of other libraries such as Tcl, but instead is directly linked to those libraries when it is created, it is bound to those libraries by a particular name and version, and it will be necessary to make sure parties to whom the shared object is distributed provide compatible libraries. It will probably also be necessary recompile such a shared object when switching to newer version of the shared libraries it depends on.
A Tcl extension that uses stubs to call into the Tcl or Tk C API can be loaded by any version of Tcl that supports the API required by the extension. Many extensions use stubs in this way, and any extension that is included in starkit must use stubs.
If it provides a C API, an extension (or other library) may also provide a stubs table of its own. This allows other code to dynamically bind to it. For example, TclOO provides a stubs table that other Tcl extensions can use. TclOO, Tk, and memchan are all examples of libraries that both use stubs tables and provide stubs tables.
To see the differences take a look at the following which describes what happens when Tcl loads two different extensions.
Loading an extension which is linked directly to Tcl:
Loading an extension which uses Tcl's stub interface:
Perhaps Jean-Claude Wippler suggested it to Paul Duffin in 1999, and Paul and Jan Nijtmans, with whom Jean-Claude had also been discussing ideas, implemented it in 2000. Others involved in the first generation were ??? (backlinking details) ...
DKF: I certainly remember suggesting something like the mechanism we ended up with back before Paul produced his initial version, though his initial version was a lot more elegant...
Well, one place is that development against a Tclkit/Starkit/Starpack environment, extensions that are not built using Stubs result in more difficulty. (Or is it actually that they cannot be used at all?)
Is all this stubs stuff necessary? Why does the build system have to be so difficult. Believe us: this is far less difficult than the alternative. Yes, libtool and autoconf are maddening, but there are things outside the Tcl world that demonstrate we don't have it so bad. DKF briefly defends stubs in a follow-up [1 ] to a lengthy thread that itself touches several aspects of generation.
Helmut Giese notes that, even when compiler providers agree on object formats, they might still construct libraries in incompatible ways. In particular, under Windows, those who choose to work with gcc-based compilation might need to rebuild tclstub84.lib before Stubs-enabled extensions load correctly.
Georgios Petasis and Michael Schlenker observe that Stubs-less languages need "batteries included'', because reliance on version-specific extensions would otherwise be prohibitively onerous: "the C interface of these languages seems so primitive that you have to recompile everything each time a new version is out."
Phils take on stubs Stubs - Another explanation
KBK: "Even in version lockstep, stubs can be handy if only to keep all of your libraries following the same linkage conventions."
hat0 Please note that, in building a cross-platform stub-enabled extension, that Windows may require this modification to your C code's declaration:
- int Extension_Init(Tcl_Interp *interp) + int DLLEXPORT Extension_Init(Tcl_Interp *interp)
Without this, linking may fail in curious ways, with errors such like:
Cannot export [email protected][email protected][email protected]: symbol not found Cannot export [email protected][email protected][email protected]: symbol not found Cannot export [email protected][email protected]@[email protected]: symbol not found