From mboxrd@z Thu Jan  1 00:00:00 1970
From: Manfred Hollstein <manfred@s-direktnet.de>
To: law@cygnus.com
Cc: ram@netcom.com, egcs@cygnus.com
Subject: Re: SunOS shared libs ? 
Date: Thu, 09 Jul 1998 09:13:00 -0000
Message-id: <13732.58164.680856.974625@slsvhmt>
References: <199806101546.IAA05334@netcom13.netcom.com> <14465.899961057@hurl.cygnus.com>
X-SW-Source: 1998-07/msg00351.html

On Wed, 8 July 1998, 23:10:57, law@cygnus.com wrote:

 > 
 >   In message < 199806101546.IAA05334@netcom13.netcom.com >you write:
 >   > write_c_file - output name is ../../../lib/libtype1.so, prefix is libtype1_
 >   > so
 >   > static int count;
 >   > typedef void entry_pt();
 >   > void _GLOBAL__FI_libtype1_so() {
 >   >         ++count;
 >   > }
 > Here's the reason why you still need the -fPIC option when building
 > a shared library out of just pure C code.  The reference to "count"
 > must be PIC.  I don't know why we need to increment that counter.
 > Someone familiar with C++ and collect2 would have to comment.
 > 
 > jeff

I've made a small example which actually does not contain any C++
code, CTOR/DTORs etc. This is even committed by collect2:

   /u/b60/manfred/gnu/sparc-sun-sunos4/sparc-sun-sunos4.1.4/bin/nm -p libtest.so.1.0

  nm output with constructors/destructors.

  dynamic dependencies.

  0 constructor(s) found
  0 destructor(s)  found
  [Leaving libtest.so.1.0]

It's still generating the C file:

  write_c_file - output name is libtest.so.1.0, prefix is libtest_so
  
  ========== output_file = libtest.so.1.0, c_file = /tmp/cca02345.c
  #ifdef __cplusplus
  extern "C" {
  #endif
  
  write_c_file - output name is libtest.so.1.0, prefix is libtest_so
  static int count;
  typedef void entry_pt();
  void _GLOBAL__FI_libtest_so() {
          ++count;
  }
  void _GLOBAL__FD_libtest_so() {
  }
  void _GLOBAL__DI() {
          _GLOBAL__FI_libtest_so();
  }
  void _GLOBAL__DD() {
          _GLOBAL__FD_libtest_so();
  }
  #ifdef __cplusplus
  }
  #endif
  ========== end of c_file

Around lines  1452-1460  collect2's main  decides  whether or  not  to
generate the C stubs:

  if (constructors.number == 0 && destructors.number == 0
      && frame_tables.number == 0
#if defined (SCAN_LIBRARIES) || defined (COLLECT_EXPORT_LIST)
      /* If we will be running these functions ourselves, we want to emit
         stubs into the shared library so that we do not have to relink
         dependent programs when we add static objects.  */
      && ! shared_obj
#endif
      )

Obviously, the *.number  == 0 checks are all  true, and shared_obj  is
also true, hence it's going  the write_c_file way. Does anybody  know,
why  the  stubs are still  required,  even  if no  CTOR/DTORs or frame
tables are available? I think, this is plain wrong.

manfred