Re: [HEADSUP] Please try to build your packages for 64 bit

[Corinna Vinschen <corinna-cygwin at cygwin dot com>]

On Apr 20 18:49, Corinna Vinschen wrote:
> On Apr 20 11:24, Bob Heckel wrote:
> > I'm having trouble packaging 64-bit gc-7.2d "libgc" (upon which w3m
> > depends).  There are extensive 32-bit Cygwin adaptations to the upstream
> > libgc code. After much trial and error it seems I lack the experience in
> > Windows memory internals required to build a 64-bit port.
> > 
> > I'll continue trying but in the interest of speed, at this point I would
> > gladly turn over this 64-bit package to any volunteer.
> 
> I skimmed through the os_dep.c file and at first glance I only see
> one problem:
> 
>   # else /* CYGWIN32 */
>     /* An alternate version for Cygwin (adapted from Dave Korn's        */
>     /* gcc version of boehm-gc).                                        */
>     GC_API int GC_CALL GC_get_stack_base(struct GC_stack_base *sb)
>     {
>       extern void * _tlsbase __asm__ ("%fs:4");
>       sb -> mem_base = _tlsbase;
>       return GC_SUCCESS;
>     }
>   # endif /* CYGWIN32 */
> [...]
> suggest to change the above code to this target-independent expression:
> 
>       extern void * _tlsbase = NtCurrentTeb()->Tib.StackBase;
> 
> If you have more strange problems, feel free to explain them here.

I just found two more.  include/gc.h uses __CYGWIN32__:

  #ifdef __CYGWIN32__
    /* Similarly gnu-win32 DLLs need explicit initialization from the     */
    /* main program, as does AIX.                                         */
    extern int _data_start__[], _data_end__[], _bss_start__[], _bss_end__[];
  # define GC_DATASTART (_data_start__ < _bss_start__ ? \
			 (void *)_data_start__ : (void *)_bss_start__)
  # define GC_DATAEND (_data_end__ > _bss_end__ ? \
		       (void *)_data_end__ : (void *)_bss_end__)
  # define GC_INIT_CONF_ROOTS GC_add_roots(GC_DATASTART, GC_DATAEND); \
				   GC_gcollect() /* For blacklisting. */
	  /* Required at least if GC is in a DLL.  And doesn't hurt. */
  #elif defined(_AIX)

__CYGWIN32__ is extremly outdated, only kept for backward compatibility
on i686, and it's not defined anymore for 64 bit.  Always use __CYGWIN__
instead.

However, the code in include/gc.h guarded by __CYGWIN32__ wouldn't work
on 64 bit anyway.  All symbols need an additional underscore when
building on 64 bit Cygwin, due to the fact that under x86_64, symbols
are not prepended with an underscore, as on i686.

So the code should be changed to something along the lines of

  #ifdef __CYGWIN__
    /* Similarly gnu-win32 DLLs need explicit initialization from the     */
    /* main program, as does AIX.                                         */
  #ifdef __x86_64__
  # define GC_DATASTART (__data_start__ < __bss_start__ ? \
			 (void *)__data_start__ : (void *)__bss_start__)
  # define GC_DATAEND (__data_end__ > __bss_end__ ? \
		       (void *)__data_end__ : (void *)__bss_end__)
  #else
  # define GC_DATASTART (_data_start__ < _bss_start__ ? \
			 (void *)_data_start__ : (void *)_bss_start__)
  # define GC_DATAEND (_data_end__ > _bss_end__ ? \
		       (void *)_data_end__ : (void *)_bss_end__)
  #endif
  # define GC_INIT_CONF_ROOTS GC_add_roots(GC_DATASTART, GC_DATAEND); \
				   GC_gcollect() /* For blacklisting. */
	  /* Required at least if GC is in a DLL.  And doesn't hurt. */
  #elif defined(_AIX)

Also, there's this block in include/private/gcconfig.h:

  # if defined(__CYGWIN32__) || defined(__CYGWIN__)
  #   define I386
  #   define CYGWIN32
  #   define mach_type_known
  # endif

This needs extending to handle the new CPU type:

  # if defined(__CYGWIN__)
  #   if defined(__LP64__)
  #     define X86_64
  #   else
  #     define I386
  #   endif
  #   define CYGWIN32
  #   define mach_type_known
  # endif

And ask upstream if it's really necessary...

  #   ifdef CYGWIN32
  #       define OS_TYPE "CYGWIN32"

... to call the OS "CYGWIN32", despite the fact that it's not called
Cygwin32 but Cygwin since 1998 (15 years!).  "CYGWIN" would make more
sense.


Corinna

-- 
Corinna Vinschen                  Please, send mails regarding Cygwin to
Cygwin Maintainer                 cygwin AT cygwin DOT com
Red Hat