re: BFD patch to fix Mungw32 problem

re: BFD patch to fix Mungw32 problem

[Paul Sokolovsky]

Hello Bfd@cygnus.com,,

  Recently I proposed patch to workaround mingw32/win95 bug with
producing incorrect dll's. However, that patch broke handling of other
coff formats. New version explicitly applies only to PE images (format
pei).

Best regards,
 Paul                          mailto:paul-ml@is.lg.ua
coffcode.diff

Re: BFD patch to fix Mungw32 problem

[Mikey]

There is a better way to handle this that is PEI
specific.

in ld/scripttempl/pe.sc
just add
    LONG (0); LONG (0); LONG (0); LONG (0); LONG (0);
to whatever section isn't being terminated properly.

you can use the -T option to ld, and write a linker script
to test.

On Thu, 11 Mar 1999 19:15:18 +0200, you wrote:

>Hello Bfd@cygnus.com,,
>
>  Recently I proposed patch to workaround mingw32/win95 bug with
>producing incorrect dll's. However, that patch broke handling of other
>coff formats. New version explicitly applies only to PE images (format
>pei).
>
>Best regards,
> Paul                          mailto:paul-ml@is.lg.ua

Re: BFD patch to fix Mungw32 problem (Better Solution!!)

[Mikey]

Finally!!!

in ld/scripttempl/pe.sc

  .reloc ${RELOCATING+BLOCK(__section_alignment__)} :
  { 			
    reloc_start_addr__ = .;		
    *(.reloc)
    reloc_end_addr__ = .;
    . += ( reloc_end_addr__ - reloc_start_addr__ ) ? 5 : 0 ;
  } =0x0000

This works without adding a .reloc section to all images!!

On Thu, 11 Mar 1999 19:15:18 +0200, you wrote:

>Hello Bfd@cygnus.com,,
>
>  Recently I proposed patch to workaround mingw32/win95 bug with
>producing incorrect dll's. However, that patch broke handling of other
>coff formats. New version explicitly applies only to PE images (format
>pei).
>
>Best regards,
> Paul                          mailto:paul-ml@is.lg.ua

Re: BFD patch to fix Mungw32 problem (Better Solution!!)

[Mikey]

Woops, forgot the location counter moves in bytes not longs ;-)
should be

in ld/scripttempl/pe.sc

  .reloc ${RELOCATING+BLOCK(__section_alignment__)} :
  { 			
    reloc_start_addr__ = .;		
    *(.reloc)
    reloc_end_addr__ = .;
    . += ( reloc_end_addr__ - reloc_start_addr__ ) ? 5 * 4 : 0 ;
  } =0x0000

BFD patch to fix Mungw32 problem

[Paul Sokolovsky]

Hi!

   I was told that this is correct place to submit BFD patches. Here
it is with a bit of motivation.



This is a forwarded message
From: Paul Sokolovsky <paul-ml@is.lg.ua>
To: Mumit Khan , DJ Delorie <dj@delorie.com>, cygwin@sourceware.cygnus.com <khan@xraylith.wisc.edu>
Subject: DLLs built corrupted with mingw32 hunt

===8<==============Original message text===============
Hello Mumit,


      This describes problem concerning building DLLs with mingw32 on
win95, with solution, some kind of.

      Fact: dlls, linked with native mingw32 ld on win95 in most cases
are not relocatable. The problem is in behavior of underlying lseek()
(SetFilePointer() ultimately). It sure not there when cross-compiling,
and I don't believe that NT has it - how could they get C2 for 3.51
with it in?

      External symptom is that referred dlls has garbage at the
sections paddings. I send 2 samples to Mumit if he'd like to take a
look at them. Here's a utility to check relocatabilty by practise:

----
#include <windows.h>

char tempname[100];

void main(int argc,char *argv[])
{
  HANDLE h=0,h2=0;
  if (argc!=2) {printf("Usage: check-dll-reloc <dll>\n"); exit(0);}
  h=LoadLibrary(argv[1]);
  if (!h) {printf("Couldn't load image '%s' at all\n",argv[1]); exit(0);}
  strcpy(tempname,"dllXXXXXX");
  mktemp(tempname);
  strcat(tempname,".dll");
  printf("Copying to temp '%s'\n",tempname);
  CopyFile(argv[1],tempname,TRUE);
  h2=LoadLibrary(tempname);
  printf("Handles (load addrs) are: %x %x\n",h,h2);
  printf("Image is %srelocatable\n",h2?"":"NOT ");
  FreeLibrary(h);
  FreeLibrary(h2);
  DeleteFile(tempname);
}
----

    The problem itself is due win95 SetFilePointer(), when seeking
past end of file, leaving previous medium contents in areas not
explicitly written, despite POSIX saying they should be zeroed. BFD
likes such seeks, so it goes. In cygwin b20 dll was introduced workaround
for this, so its ld produces correct dll's. It's, however, would be
nice to have mingw32 self-sufficient, so I tried to fix it. I cast
aside making wrapper around lseek by analogy with cygwin because I
neither like that solution nor have idea how to implement it
correctly. Instead I found "bad" places in BFD and did workarounds there.

    Attached is patch as of bfd from cygwin b20. It's somewhat adhoc,
so changes wrapped in #ifdefs.

    With this patch dll's are ok. (If someone has this problems and
wants immediate solution, there's precompiled binary at
http://ftp.is.lg.ua/pub/gnuwin32/alpha/ld-mingw32-replacement.zip .
It's however an alpha effort featuring bfd as dll (required) at
http://ftp.is.lg.ua/pub/gnuwin32/alpha/bfd-dll.zip and another
underlying library). But strip built with patched bfd still wrecks dll
off relocatability, just as cygwin b20 strip does.


Best regards,
 Paul                          mailto:paul-ml@is.lg.ua

===8<===========End of original message text===========



Best regards,
 Paul                            mailto:paul-ml@is.lg.ua
coffcode.h.dif