This is a discussion on Re: gcc on 5.0.7 - SCO ; Jean-Pierre Radley wrote: > Kean, this is from the newsgroup. Sum it up for Roger, would you please? > > Roger Cornelius typed (on Mon, Jul 14, 2003 at 02:38:15PM -0700): > | Ever since installing Openserver 5.0.7, I am ...
Jean-Pierre Radley wrote:
> Kean, this is from the newsgroup. Sum it up for Roger, would you please?
> Roger Cornelius typed (on Mon, Jul 14, 2003 at 02:38:15PM -0700):
> | Ever since installing Openserver 5.0.7, I am unable to use gcc 3.2.1.
> | It returns an Undefined symbol _fini error for any compilation
> | attempt.
> | I tried installing the Gnu Development Tools from the OS installation
> | CD and that gcc works so far as I've tested, but it's unable to
> | build/bootstrap gcc 3.3 and fails at some point with the _fini error.
> | Using the SCO provided cc to build gcc, make fails with an "undefined
> | symbol: DEFAULT_PCC_STRUCT_RETURN" error in i386.c for 3.2.2, and
> | configure fails with an "cannot determine size for long long" error
> | for gcc 3.3.
> | I've watched the c.u.s.* and gcc newsgroups and checked the gcc web
> | site but have not found any answers or significant discussion. Am I
> | missing something?
Can you mirror this reply to the newgroup so that others that watch it
can learn about this too?
Roger, heres the scoop. When you create an executable with a compiler,
you not only link in your objects and libraries, but behind the scenes,
the compiler driver (gcc, cc, whatevercc) will prepend special system
object files, and append other special ones. For C programs, these are
known as the CRT (C run-time) files. These typically do things like set
up the float point environment, set up any profiling counters and calls,
call any global initialization code and then finally call main(). When
the program terminates via any mechanism that uses exit(), then some
special code that is appended by the CRT files does global destruction
and other cleanups, calculates profiling totals, and then actually exits.
In ELF objects, code and data are put into special named "sections".
These are controlled by the ".section" assembler directive, and its
abbreviations like ".text" which is basically ".section .text" and
".data" which is essentially ".section .data". Code goes into the
section called .text, and data goes into the sections .data, .rodata or
..bss, depending on how the data is declared. There are several sections
which are very special to ELF. For your actual problem as experienced,
these are the .init section and the .fini section. These are designed to
have initialization and finalization code, respectively. When you link
edit an ELF program, the ELF ABI states that all code in a .init section
is executed before main() and that call code in a .fini section is
executed before exiting. This is how C++ global constructors and
destructors are implemented.
Although you are not experiencing problems with this yet, I want to
document this for the sake of completion, since this message will
hopefully make its way to the newsgroup and other people may or may not
be experiencing other problems.
Modern ELF extended the notion of the .init and .fini sections with
three special sections: .preinit_array, .init_array and .fini_array.
These arrays store pointers to functions, not arbitrary code as .init
and .fini do. The gABI specifies that each entry in the .preinit_array
section is executed before the program entry point. This means it gets
executed before the CRT stuff gets invoked, which means it gets executed
before any code in the .init section. Likewise, the functions in
..init_array are also called before the entry point. Shared libraries
have .init_array called every time they are loaded. The system also sets
things up such that functions in .fini_array are called when a shared
library is unloaded and before functions in the .fini section when the
program exits normally.
So thats what the code needs to do. Here is why you were having a
problem. Typically, the CRT is divided into three files. crt1.o, crti.o
and crtn.o. crt1.o contains the actual C startup stuff, and work some
magic to interact with teh run-time link editor (RTLD). crti.o contails
a single function, called _init, that is inserted at the very beginning
of the .init section. The same applies to a function called _fini and
starts the .fini section. These function do nothing, they are empty.
crtn.o, which is one of the last objects link edited into the program
or shared library, terminates the .init and .fini sections with a simple
return instruction. Thus, when you link edit a program, you get the
deifintion of the _init function, then the code from all of your program
objects, in order, with their .init and .fini sections, then the
termination from crtn.o. This results in a single contiguous function
such that when _init is called, all code from all .init sections is
executed linearly in object order, and all code from all .fini sections
is executed in object order.
When you link edit a program, you use all three startup files: crt1.o,
crti.o yourobjects.o libraries.so crtn.o. When you link edit a shared
library, you typically omit crt1.o, as the code in it has no meaning to
a shared library. There is no need to set up calls to main or reset the
floating point environment or set up profiling counters. In fact, doing
so can completely screw your program up. Consider, for example, if you
had carefully set up your floating point environment to be in IEEE mode,
and the code in crt1.o sets it to some other mode. Or consider the fact
that profiling counters may get reset erroneously. All the shared
libraries need is the stuff in crti.o and crtn.o, to start and finish
the .init and .fini sections (and correspondingly the _init and _fini
functions for that shared libary).
In releases prior to 5.0.7, OpenServer did not have a crti.o. It had the
_init stub in crt1.o. This caused untold grief, especially when we
started using G++ to create shared libraries. So, starting in release
5.0.7, I split out the stub stuff from crt1.o and put it in crti.o where
it always should have been. The native compiler driver (cc) was changed
to link with crti.o and so was the officially supported version of GCC
(2.95.3). At the time that I made this change, there was no mor work
being done on the 2.95 branch of GCC, they were in the middle of doing
GCC 3.2. We didnt fully trust the GCC 3 branch yet which is why we never
did any work on it to update it to use crti.
However, things are now to the point where we will be adopting GCC 3 as
the officially supported version of GCC. This will happen for release
5.0.8 of SCO OpenServer. To that end, I did all the required work to get
this stuff working in GCC 3. At the time that I did the work, they were
at the tail end of getting ready to release 3.3, and the GCC Steering
Committee decided that these changes were not appropriate for that
release. Thus, these changes were put in at the head of the tree, and
are in (or will be in, when it is releaed) GCC 3.4. If you really want a
GCC 3.x release, you need to check out the head of the tree to get this
stuff. I was intending to backport the changes to 3.3.1 but I have not
had the time. If there is ever a 3.3.2, it may go into that but I
suspect that the steering committee will veto that notion. For for GCC
3.x, the first version that will be known to work propperly on OSR5 is 3.4.
If you are feeling brave, when 3.3.1 is released in a few days/weeks,
you can get the stuff from the head from CVS, look at the changelogs for
what I had to do for OSR5 and back-port it. I may do that myself if I
have the time, and then get Boyd Lynn-Gerber to post the required patch
on his website where he collects such things. In two short weeks we will
be releasing a significan update to the GNU Development Tools package,
which provides the current officially supported GCC and a plethora of
other GNU tools that software developers frequently use. This will rely
on Support Level Supplement OSS646B, which is the execution environment
update. This update fixes the RTLD, libc and crt*.o files such that they
meet the gABI requirements, as outlined above. As far as I am concerned
every OpenServer system from 5.0.4 on upwards should install OSS646B,
but some administrators are loathsome to install such big changes and
thats a whole other issue.
Hope this helps.