NetBSD

History

castor 82bc6ffe05 Avoid confusing various C preprocessors by avoiding # [0-9]		1999-01-18 06:49:57 +00:00
..
arch	Avoid confusing various C preprocessors by avoiding # [0-9]	1999-01-18 06:49:57 +00:00
debug.c
debug.h	Add __attribute__ for printf like commands.	1999-01-10 17:19:01 +00:00
headers.c
link.h	gdb's shared-lib mapping code in solib.c has #ifdef dependencies on	1998-10-30 05:43:40 +00:00
load.c
Makefile	Add i386 bits	1999-01-10 18:18:56 +00:00
malloc.c
map_object.c
paths.c
README	Add a few comments about the i386 and others.	1999-01-11 11:05:29 +00:00
reloc.c	Fixup some printf format argFixup some printf format argFixup some printf format argFixup some printf format argFixup some printf format argFixup some printf format argFixup some printf format argFixup some printf format argFixup some printf format args.	1999-01-11 23:12:16 +00:00
rtld.c	Add i386 bits	1999-01-10 18:18:56 +00:00
rtld.h	Add powerpc support.	1998-11-24 11:34:30 +00:00
rtldenv.h
search.c
symbol.c
sysident.h	At least make the comments accurate.	1998-09-13 05:53:34 +00:00
xmalloc.c
xprintf.c

README

$NetBSD: README,v 1.3 1999/01/11 11:05:29 christos Exp $

BUGS/PROBLEMS:

* rtld.c around line 147:

	#iffdef __alpha__
	/* XXX XXX XXX */
	_rtld_objself.pltgot = NULL;
	#endif

  It's a workaround for a linker bug, and will go away when the linker is
  fixed.

* Formerly, _DYNAMIC was _GLOBAL_OFFSET_TABLE_[0], but that isn't true with
  newer snapshots of gas/binutils (at least on the Alpha), which support
  multiple GOTs.  That's what the OLD_GOT #ifdefs are about.
  
* The method used to relocate ld.so is shoddy and fragile.  Currently it's
  something like:

	(1) assembly code relocates RELATIVE relocations in the GOT
	    (on some ports, all relative relocations; this will probably
	    have to be #ifdef'd in reloc.c), assuming that everything
	    in the GOT is a RELATIVE relocation.

	(2) C code relocates the rest of the loader, excluding those
	    RELATIVE relocations already done (via an ... iffy heuristic).

  Something better (like that the glibc ELF ld.so does) should be done
  eventually, where a single set of C code is compiled twice, once for
  normal use, and once to be used to relocate the shared linker.  No
  relocations should be processed in assembly language.

* On the Alpha, this code Really needs to be checked to be sure that
  it doesn't make calls that could use FP, while doing lazy binding.

* R_*_COPY relocations don't really work. On the i386 at least:
  call printf, generates an R_I386_COPY relocation and the
  linker generates a direct call to the bss.rel section.
  The copy relocation is supposed to put a jmp to printf
  there, but code is broken and does not do anything. I.e.
  it does not know the size of the relocation (it is 0) or
  put a jmp table there. For now we can create working binaries
  with -fPIC or -fpic because they don't use R_*_COPY relocations.

* On the i386, we cannot relocated the dynamic linker from 0, because
  in pic mode globals use the plt, so it makes it difficult to do things
  before we relocate ourselves. Maybe half-pic will help?

TO DO:

* Support for other platforms

* Support for coexistance of 32-bit and 64-bit ELF on platforms that can
  do that.

* KNF

* Possible support for ldd (standalone operation) and program loading
  a la the glibc shared loader.

* Support for relocating ld.elf_so on mips