TODO for MacOS X:
- Add the vm_region tags as per Robert Bowdidge email of 11/6/2000.
- Determine which existing API's should call link_library_module() with the
new bind_fully parameter set to TRUE or link_object_module() with the new
bind_fully parameter set to TRUE.
- Figure out a way to make the __TEXT segment of dyld read-only when the kernel
bug to map it read only is fixed. For ppc the problem is the instructions
after __dyld_start in dyld_start.s which loads the value of the lable L1.
These create relocation entries in a read only section but are not need to
be relocated. Until the kernel bug is fixed that honors the protection
a vm_protect() call on dyld's __TEXT segment could be done.
- Deal with mig not having a dyld_event_MsgError() call back for mig errors
when user_dyld_event_server_callback() is called.
- Figure out if dyld should use malloc_error() to get "dyld: " in front of
malloc error messages.
TODO:
- For hppa cache flushing use FDC and FIC on the written instruction address.
Changes for the 5.8 release (the cctools-358 release):
- Fixed a bug in link_object_module() in symbols.c where when loading a module
privately with coalesced symbols the loop does not check the value of the
indirect symbol table entry against INDIRECT_SYMBOL_LOCAL and
INDIRECT_SYMBOL_ABS before using the value as in index into the symbol table.
Radar bug #2631405.
Changes for the 5.8 release (the cctools-357 release):
- Changed the Makefile from using:
LIBC = -L$(NEXT_ROOT)$(SYSTEMDIR) -lc -lcc_dynamic
to:
LIBC = $(NEXT_ROOT)$(SYSTEMDIR)/libc.a -lcc_dynamic
because some where between Cheetah1K16 and Cheetah1K19 a symbolic link
/usr/lib/libc.dylib -> libSystem.dylib
was created confusing the -lc reference.
Changes for the 5.8 release (the cctools-356 release):
- Moved the call to relocate_symbol_pointers_in_library_image() to the bottom
of the loop in the routine relocate_modules_being_linked() in reloc.c .
The thinking is that this could be the cause the failure of the the multi
threaded java testing program. Where one thread is in dyld relocating a
module and another thread gets in via a set lazy pointer then touches an
item that is not yet relocated and crashes. Radar bug #2584800.
Changes for the 5.8 release (the cctools-355 release):
- Fixed a bug in relocate_symbol_pointers() where an lazy_init library should
never have its symbol pointers relocated. This causes a dead lock in dyld as
the code in relocate_symbol_pointers() which is holding the lock then causes
an exception on the symbol pointer that is read-only. Then the exception code
trys to get the lock and spins. The fix is in relocate_symbol_pointers() in
symbols.c that checks to see if the image has the lazy_init bit set if so it
just returns. Radar bug #2582778.
- Fixed the fix to _dyld_NSLookupAndBindSymbolWithHint() in dyld_libfuncs.c
that was missing the line:
link_state = GET_LINK_STATE(*defined_module);
to set the value of link_state before testing it. Radar bug #2580180.
Changes for the 5.8 release (the cctools-354 release):
- Fixed a bug in _dyld_NSLookupAndBindSymbolWithHint() and
_dyld_lookup_and_bind_with_hint in dyld_libfuncs.c that did not check after
the symbol was found that is was linked.
Changes for the 5.8 release (the cctools-353 release):
- Added the use_header_dependencies parameter to call_dependent_init_routines()
in images.c to use the header dependencices instead of the symbol dependencies
when calling library init routines for the first time in a prebound app. The
goal of this is to avoid touching all the symbol tables of all the libraries
at launch time. Radar bug #2571628.
- Added lookup_symbol_in_hinted_library() to symbols.c and
_dyld_lookup_and_bind_with_hint(), _dyld_NSLookupAndBindSymbolWithHint() and
_dyld_NSIsSymbolNameDefinedWithHint() to dyld_libfuncs.c that use
lookup_symbol_in_hinted_library(). Radar bug #2571272.
- Moved the images.h header file to /usr/local/include when RC_OS is macos and
RC_RELEASE is not Beaker, Bunsen, Gonzo or Kodiak. Radar bug #2569880.
- Added check_libraries_for_overrides() to symbols.[ch]. Added the library_image
parameter to check_libraries_for_definition_and_refernce() which skips looking
at that library_image for symbol definitions and references.
Changed try_to_use_prebound_libraries() in images.c to call the new routine
check_libraries_for_overrides().
Radar bug #2564555.
Changes for the 5.8 release (the cctools-350 release):
- Added support for the new DYLD_NEW_LOCAL_SHARED_REGIONS environment variable
(Radar bug #2567132).
Changes for the 5.8 release (the cctools-349 release):
- Added inline versions of bsearch routines and strcmp. These are in the files
inline_bsearch.h and inline_strcmp.h. The inline strcmp() routine is in
gnu 2.95 syntax for PowerPC only. (Radar bug #2558068).
- Changed where the pointer to the mach_header field, mh, is in struct image to
match <mach-o/dyld_gdb.h>. As well as where the nimages field is in the
object_images and library_image structs to also match <mach-o/dyld_gdb.h>.
Also added a valid field to the image data structure for gdb to use intead of
exporting the info to have it look at the LINK_STATE for the UNUSED state.
This is set to TRUE in images.c after the structures are filled in and set
for FALSE before an image is to be removed. Added a calls to
gdb_dyld_state_changed() at the end of load_executable_image() in images.c
and at the end of _dyld_link_module(), _dyld_unlink_module() and
_dyld_NSAddLibrary() in dyld_libfuncs.c . Radar bug #2559324.
Changes for the 5.8 release (the cctools-347 release):
- Added the support for _dyld_register_func_for_remove_image() in dyld. This
effected register_funcs.{c,h} and images.c. Also added the missing calls
to send_event() for libraries that were removed. Radar bug #2553324.
- Added the support for NSAddLibraryWithSearching() in dyld. This was basicly
adding a enum bool force_searching parameter to load_library_image(). Radar
bug #2535633.
Changes for the 5.8 release (the cctools-346 release):
- Changed unload_remove_on_error_libraries() in images.c call a newly added
static routine unload_shared_file() which calls the new kernel api
reset_shared_file(). Radar bug #2549068.
Changes for the 5.8 release (the cctools-345 release):
- Changed unload_remove_on_error_libraries() in images.c to not vm_deallocate()
the memory of a split library. It currently prints a message saying it can't
unload the library. Radar bug #2544696.
- Added the ifdef __HERA__ for MacOS X Server in images.c as it was using
__GONZO_BUNSEN_BEAKER__ which was really not right. Also changed the comment
in the top level Makefile. This effected the files debug.c, mod_init_funcs.h,
mod_init_funcs.c, dyld_init.h, dyld_init.c and images.c .
Changes for the 5.8 release (the cctools-341 release):
- Changed link_library_module() and link_object_module() in symbols.c to also
not check for multiply defined symbols when the link_state of the module is
in the BEING_LINKED state as well as the LINKED state. This is done to fix
a problem with global coalesced symbols. It avoids having more than one
instance of two differen global coalesced symbol on the being_linked_list
which causes the symbol pointers to be set to the wrong global coalesced
symbol. Radar bug #2511358.
- Changed the tests in pickup_environment_variables() in dyld_init.c to allow
a getuid() of 0 (root) to let the environment variables have effect. Radar
bug #2497678.
- Added a call to free() in load_library_image() in images.c in the case that
is_library_loaded_by_name() returns TRUE and load_library_image() was called
by NSAddLibrary(), dl == NULL. Also did the same for the case that
is_library_loaded_by_stat() returns TRUE. Radar bug #2504987.
- Changed the value of NLIBRARY_IMAGES in images.h from 10 to 60 based the
Finder's uses with MacOS X Public Beta (Kodiak1G7). Also change the value
of NOBJECT_IMAGES from 40 to 5 based again on the Finder. Left the value of
MODULE_STATE_BLOCK_SIZE in images.c at 2000 since this is currently enough
for the Finder. The value of NSYMBOL_BLOCKS in symbols.c was changed to 45
from 35 as this is what the Finder needs to launch in its unprebound state.
Also note that 4 bytes got added to struct symbol_list in symbols.h without
the value allocated in section_order.s being updated. This means that the
value in section_order.s for symbol_blocks was changed from 84280 to 126360.
Radar bug #2504987.
- Changed the calling of module initialization in images marked with
MH_LAZY_INIT to also cause them to be run lazily. Radar bug #2502941.
Changes for the 5.7 release (the cctools-340 release):
- Picked up a change to debug.c that Klee Dienes made to make Radar bug #2487248
no longer happen. This involves adding a usleep(20000) to server_loop()
before it tries to reset the suspend count of the thread. Added an ifndef
__GONZO_BUNSEN_BEAKER__ around this as usleep() on MacOS X Server has
references to dyld functions.
Changes for the 5.7 release (the cctools-337 release):
- In dyld_start.s the stack alignment was changed from this:
clrrwi r1,r1,4 ; align to 16 bytes
was changed to this:
clrrwi r1,r1,5 ; align to 32 bytes
as it is in start.s in the Csu project. Radar bug #2495403.
Changes for the 5.7 release (the cctools-336 release):
- Fixed a bug in _dyld_image_containing_address() in dyld_libfuncs.c when a
plugin was unloaded. That code and many other places need to test the module
state of the object to see if it was unused before looking at the other
fields. Radar bug #2493911.
Changes for the 5.7 release (the cctools-335 release):
- Fixed a bug in construct_at_executable_path_library() in images.c where it was
doing a strncpy() followed by a strcat(). This doesn't work if a NULL is not
in the string after the strncpy() string as it does not copy the NULL.
Radar bug #2493400.
Changes for the 5.7 release (the cctools-334 release):
- Added support for lazy_inits using a second thread via the Mach 3.0 exception
API's. Most of these changes are in images.c . Radar bug #2490396.
- Fixed a bug in link_object_module() that could set a module back to the
BEING_LINKED state when it was in the LINKED or FULLY_LINKED state causing
it to be relocated twice (also part of Radar bug #2483733). Also changed
linked_library_module() to also protect against this case.
Changes for the 5.7 release (the cctools-333 release):
- Changed the call to load_shared_file() in images.c to match the changed
prototype for Kodiak1E (Radar bug #2484973).
Changes for the 5.7 release (the cctools-332 release):
- Fixed a bug in link_library_module() in symbols.c in the loop to resolve
private externs the call to link_library_module() should have passed the
prameters bind_now and bind_fully to the call to link_library_module() not
always pass them as TRUE. This caused things to get relocated twice in this
case a "FILE output_fp = stdout" (Radar bug #2483733).
- Fixed _dyld_image_containing_address() to deal with split dylibs.
Radar bug #2480801.
- Added setting the cksum field of the sf_mapping structs to zero for now. This
might be used for the digest sections.
Radar bug #2480801.
- Added a test for setting up the sf_mapping structs to only create a zerofill
map if the segment has write protection and the vmsize is greater than the
file size.
Radar bug #2480801.
- Changed the call to load_shared_file() to first call it with flags == 0 then
if that fails call it with flags == ALTERNATE_LOAD_SITE. Also changed the
error case to not return(FALSE) but to goto map_image_cleanup1 to clean up
the mapped memory and open file descriptor.
Radar bug #2480801.
- Changed the code around load_shared_file() to not subtract constants but
rather *seg1addr for setting the mapping_offset field. Also changed it so
that no address is subtracted from the base_address variable.
Radar bug #2480801.
- Changed to using SHARED_LIBRARY_SERVER_SUPPORTED instead of
_SHARED_MEMORY_SERVER_H_ in images.c and removed the code on the other side
of the ifdef NEW_LOAD_SHARED_FILE and made that the default.
Radar bug #2480801.
- Fixed a bug in map_image() in images.c with respect to fat files. The
file_offset of the mapping structure being set for an LC_SEGMENT can't be
set to just sg->fileoff but needs to be "sg->fileoff + library_offset".
Radar bug #2480801.
- To build dyld with -DNEW_LOAD_SHARED_FILE we need new headers from Chris and
a new libc.a to link against.
Note the prototype for load_shared_file() was in <mach/mach_traps.h> and now
is in <bsd/sys/mman.h>. Also note most of the headers are in BOTH Headers
and PrivateHeaders.
I picked this file up from
cdy1:/SandBox/kev2/75.2/xnu/BUILD/obj/EXPORT_HDRS/osfmk/mach/mach_traps.h
installed into:
/System/Library/Frameworks/System.framework/Headers/mach/mach_traps.h
and into:
/System/Library/Frameworks/System.framework/PrivateHeaders/mach/mach_traps.h
I picked this file up from:
cdy1:/SandBox/kev2/75.2/xnu/BUILD/obj/EXPORT_HDRS/bsd/sys/mman.h
installed into:
/System/Library/Frameworks/System.framework/Headers/bsd/sys/mman.h
I picked this file up from:
cdy1:/SandBox/kev2/75.2/xnu/BUILD/obj/EXPORT_HDRS/osfmk/mach/
shared_memory_server.h
installed into:
/System/Library/Frameworks/System.framework/Headers/mach/
shared_memory_server.h
and into:
/System/Library/Frameworks/System.framework/PrivateHeaders/mach/
shared_memory_server.h
We added "#define SHARED_LIBRARY_SERVER_SUPPORTED" to
/System/Library/Frameworks/System.framework/Headers/mach/
memory_object_types.h
and also to
/System/Library/Frameworks/System.framework/PrivateHeaders/mach/
memory_object_types.h
Then dyld will build with -DNEW_LOAD_SHARED_FILE.
Next we must link dyld with a libc that has the correct trap numbers. To do
this we need to add syscall stubs to Libc and recompile the Libc project
against these new headers:
I picked this file up from
cdy1:/SandBox/kev2/75.2/xnu/BUILD/obj/EXPORT_HDRS/osfmk/mach/mach_traps.h
installed into:
/System/Library/Frameworks/System.framework/Headers/mach/mach_traps.h
and into:
/System/Library/Frameworks/System.framework/PrivateHeaders/mach/mach_traps.h
I picked this file up from
cdy1:/SandBox/kev2/75.2/xnu/BUILD/obj/EXPORT_HDRS/bsd/sys/syscall.h
installed into:
/System/Library/Frameworks/System.framework/Headers/bsd/sys/syscall.h
and into:
/System/Library/Frameworks/System.framework/PrivateHeaders/bsd/sys/syscall.h
I picked this file up from
cdy1:/SandBox/kev2/75.2/xnu/BUILD/obj/EXPORT_HDRS/osfmk/mach/syscall_sw.h
installed into:
/System/Library/Frameworks/System.framework/Headers/mach/syscall_sw.h
and into:
/System/Library/Frameworks/System.framework/PrivateHeaders/mach/syscall_sw.h
The new syscall stubs were done by making a copy of Libc-139 and created the
files:
sys.subproj/ppc.subproj/load_shared_file.s
With this code in it:
#include "SYS.h"
SYSCALL(load_shared_file, 7)
blr
where 7 is nargs for the macro SYSCALL as defined in SYS.h.
sys.subproj/i386.subproj/load_shared_file.s
With this code in it:
#include "SYS.h"
UNIX_SYSCALL(load_shared_file, 7)
ret
And changed the sys.subproj/ppc.subproj/Makefile and
sys.subproj/i386.subproj/Makefile and added load_shared_file.s to the
OTHERLINKED macro and load_shared_file.o to the OTHERLINKEDOFILES macro.
And changed sys.subproj/ppc.subproj/PB.project and
sys.subproj/i386.subproj/PB.project to add load_shared_file.s after listen.s.
% otool -tV -p _load_shared_file "/usr/local/lib/system/libc.a(mach_traps.o)
/usr/local/lib/system/libc.a(mach_traps.o):
(__TEXT,__text) section
_load_shared_file:
000000d0 li r0,0xffcc
000000d4 sc
000000d8 blr
where 0xffcc is -52.
The recompiled libc will show
% otool -tV -p _load_shared_file "/SandBox/Libc-139.roots/Libc-139.dst/usr/
local/lib/system/libc.a(load_shared_file.o)"
(__TEXT,__text) section
__framesize:
00000000 li r0,0x12b
00000004 sc
00000008 b 0x10
0000000c b 0x30
00000010 mfspr r0,lr
00000014 bl 0x18
00000018 mfspr r12,lr
0000001c mtspr lr,r0
00000020 addis r12,r12,ha16(0x38-0x18)
00000024 lwz r12,lo16(0x38-0x18)(r12)
00000028 mtspr ctr,r12
0000002c bctr
00000030 nop
00000034 blr
where 0x12b is 299.
Changes for the 5.7 release (the cctools-330 release):
- Changed the ifdef's for the DEFAULT_FALLBACK_FRAMEWORK_PATH in dyld_init.h so
that moving forward /Library/Frameworks is part of the default paths
and /Local/Library/Frameworks is only used instead if __GONZO_BUNSEN_BEAKER__
is defined for older DP4, DP3 & PR2 builds. Radar bug #2477862.
- Fixed a bug in call_dependent_init_routines() in images.c that casued dyld to
crash when an undefined symbol was detected. This should not happen in the
normal case but can happen if an init routine were to unload a module another
init routine depended on (that unload would have to use RESET_LAZY_REFS as any
non-lazy reference would cause an error at the time of unloading). This did
happen because of the next bug listed below in dyld. The fix for this bug is
after the calls to lookup_symbol() where it returns NULL for defined_symbol is
now tested for an the loop is continued if defined_symbol is NULL. Radar bug
#2477856.
- Fixed a bug in dyld that did not correctly fully bind the references to a
shared library init routine if the executable is overridding a symbol that the
shared library init routine uses. Radar bug #2477856.
The fix starts in add_reference() in symbols.c when a reference to a symbol
being passed with fully_bind set to TRUE is looked up and found in a module
who's state is already being linked in. Then the symbol could be either be
found on the being_linked_list or not. In either case when bind_fully is TRUE
the FULLYBOUND_STATE needs to be checked. If the symbol was not found on the
being linked list then add_to_undefined_list() can simply be called with the
bind_fully parameter passed. If the symbol was found on the being linked list
it needs to be back on the undefined list and have its bind_fully field set to
TRUE.
The fix then moves on to resolve_undefineds(). The code that handles when a
symbol on the undefined_list is looked up and found with fully_bind being
TRUE was also changed. This code now just checks fully_bind for TRUE and
the FULLYBOUND_STATE (without checking if library_image is not NULL). Then
inside that "if" another test for library_image being NULL is maded. If
library_image is not NULL then the existing code calling link_library_module()
is done. Otherwise a new call to link_object_module() is made to fully_bind
the object module.
To allow this "else" branch to call link_object_module() the new routine
find_object_image() was added. So than it could get the object_image from
the image pointer to pass to link_object_module().
In this "else" branch link_object_module() is called with the new third
parameter "bind_fully" set to true to rebind the object module fully.
The code in for link_object_module() in symbols.c was change to take the new
third parameter "bind_fully". This parameter is then used in the "if" test
when adding its undefined references. The code that added the reference
to the undefined list or being linked list was removed and a call to
add_reference() was added in its place to clean things up and avoid
duplicating this logic. And lastly this code sets the FULLYBOUND_STATE at
the end of the routine is bind_fully is TRUE.
link_object_module() was called in two places in dyld_libfuncs.c (in
_dyld_lookup_and_bind_fully() and _dyld_link_module()) and those calls were
updated to add the third prameter. Also the prototype in symbols.h was
updated.
NOTE:
When return_on_error is set an executable can get fully linked and the
FULLYBOUND_STATE will get set in link_object_module().
The undefined symbols that got added when return_on_error was set will get
correctly removed by clear_undefined_list(TRUE). Note no non-lazy undefs
from the executable could be left because the executable is bound at launch
when return_on_error is FALSE.
The name of the routine clear_state_changes_to_the_library_modules() was
changed to clear_state_changes_to_the_modules().
clear_state_changes_to_the_modules() was change to clear the module
bit FULLYBOUND_STATE if needed. Also it clears module_states_saved field
back to FALSE so a second error will not restore old state.
The other state bits MODINIT_STATE, MODTERM_STATE or IMAGE_INIT_DEPEND_STATE
will not get set if an error happens so they don't need to be backed out.
This is also the case for prebinding so reset_module_states() does not need
to change these either.
clear_state_changes_to_the_modules() was changed to clears the link state
of object modules and the FULLYBOUND_STATE if needed when return_on_error
happens. Two new fields saved_module_state and module_state_saved were
added in the object_image struct. These get set in resolve_undefineds() just
like the saved module states for libraries.
- Changed dyld to not used its own processor_has_vec variable and use
_cpu_has_altivec from the System framework private header
<architecture/ppc/processor_facilities.h>. This effected the files Makefile,
dyld_init.h, dyld_init.c, images.c and mod_init_funcs.c (The implementation
was done for Radar 2442228 the header file was done for Radar 2470661).
Radar bug #2477859.
- Changed dyld to use the new processor_facilities_used() call to determine if
the thread has ever used floating-point and altvec registers and only save
them if they have been used. This effected the files images.c and
mod_init_funcs.c (The implementation was done for Radar 2442359 the header
file was done for Radar 2470661). Radar bug #2477859.
- Changed bind_lazy_symbol_reference() in symbols.c to call
link_library_module() when binding a lazy private extern since library
module that defines private externs can also now define extern symbols.
This makes sure that multiple definitions are caught. This is needed so
that if nmedit(1) is run on a dynamic library the right thing happens
Part of Radar bug #2420307.
Changes for the 5.7 release (the cctools-329 release):
- Added the routine call_dependent_init_routines() to images.c and a call to it
in call_image_init_routines() which causes dependent library initialization
routines to be called. Radar bug #2441683.
- Changed the uses of iinit and ninit fields of the struct dylib_module in
in mod_init_funcs.c to iinit_iterm and ninit_nterm. Radar bug #2469527.
- Change the module_state to use bits to keep track of if a module init routine
was called and if a module term routine was called. This will allow
_dyld_call_module_initializers_for_dylib() (used for __initialize_Cplusplus)
to call the module init routines and mark those that have been called so
call_module_initializers() in mod_init_funcs.c not to call them again.
Radar bug #2441683.
- Added the _dyld_mod_term_funcs() routine to mod_init_funcs.c added an entry
for it in the dyld_funcs table in dyld_libfuncs.c so it can be looked up.
It is looked up in crt1.o to register with atexit() to have module termination
routines called. Radar bug #2469527.
- Added the _dyld_call_module_initializers_for_dylib() routine to
mod_init_funcs.c added an entry for it in the dyld_funcs table in
dyld_libfuncs.c so it can be looked up. It is looked up in dylib1.o for
calls to __initialize_Cplusplus() to force module initialization routines
in a specified dylib to be called. Radar bug #2441683.
- Added a call to fork_mach_init() in start_debug_thread() in debug.c when
start_from_core is TRUE. Also added some code ifdef'ed CORE_DEBUG which
sets the 6th word in the (__DATA,__dyld) section used to tell where the
debug thread is from the task trying to start the debug thread. Radar bug
#2155790.
Changes for the 5.7 release (the cctools-328 release):
- Changed the Makefile so that it would use -L/usr/local/lib/system moving
forward and -L/Local/Developer/System when RC_RELEASE is Gonzo, Beaker or
Bunsen.
Changes for the 5.7 release (the cctools-327 release):
- Changed the Makefile so that it would use -L/AppleInternal/Developer/System
moving forward and -L/Local/Developer/System when RC_RELEASE is Gonzo, Beaker
or Bunsen.
- Changed the ifdef's for the DEFAULT_FALLBACK_FRAMEWORK_PATH in dyld_init.h so
that moving forward /MacOSX/Library/Frameworks is part of the default paths
and /Local/Library/Frameworks is only used instead if __GONZO_BUNSEN_BEAKER__
is defined for older DP4, DP3 & PR2 builds.
Changes for the 5.6 release (the cctools-324 release):
- Made the changes needed to make coalesced symbols work with using only
prebound libraries. The code in link_library_module() in symbols.c that
disables prebinding when launching_with_prebound_libraries is TRUE was moved
down after the check for the multiply defined symbols being coalesced symbols.
Also a call to setup_prebound_coalesed_symbols() was added near the end of
try_to_use_prebound_libraries() in images.c.
- Added a dummy _NSGetEnviron() routine in dyld_init.c because the Gonzo1I6
versions of malloc() needs it. If I don't define it then it uses the one
from libc which needs __dyld_lookup_and_bind which can't be used in dyld.
Changes for the 5.6 release (the cctools-323 release):
- Changed the ifdef __SPACE__ for the DEFAULT_FALLBACK_FRAMEWORK_PATH in
dyld_init.h as the directory changes have been scaled back and changed.
In preparation of this "Half Plan" /Local/Library/Frameworks was changed to
/MacOSX/Library/Frameworks and ifdef'ed __SPACE__. In the second half of the
plan (post-DP4) /System/Library/Frameworks will change to
/MacOSX/System/Frameworks and this is ifdef'ed __SPACE2__.
- Changed back where dyld picks up its the System Framework's private
headers back to /System/Library/Frameworks/System.framework/PrivateHeaders
from /MacOSX/System/Frameworks/System.framework/PrivateHeaders for the Space
release. This won't change till post-DP4.
- Changed _ppc_vec_save and _ppc_vec_restore in fp_save_restore.s to use the
VRsave register to know which registers to save and restore now that gcc
2.95 is the default and support VRsave. Radar bug #2450639.
Changes for the 5.6 release (the cctools-322 release):
- Added ifdef's for __SPACE__ to the DEFAULT_FALLBACK_FRAMEWORK_PATH in
dyld_init.h in preparation for the directory changes. /System/Library
changing to /MacOSX/System.
- Added the zero-fill support for load_shared_file() with the #ifdef
NEW_LOAD_SHARED_FILE and changed the casts in the call to load_shared_file()
with this ifdef also.
Changes for the 5.6 release (the cctools-321 release):
- Changed the test of the return value of load_shared_file() from checking for
not equal to 0 to testing for equal to -1 for the error case.
Changes for the 5.6 release (the cctools-320 release):
- Added an installhdrs target to the dyld Makefile to install images.h into
System frameworks the local include directory as an egregious hack for
sampler. Note the upper Makefile was changed to call this target and also
the include/Makefile installs "stuff/bool.h". (Radar bug #2376135).
Changes for the 5.6 release (the cctools-319 release):
- To support MH_SPLIT_SEGS formats where the r_address field of relocation
entries is to be relitive to the first read-write segment a new field,
segs_read_write_addr, was added to the image struct in images.h. This then
picked up in map_image() in images.c. Then all the reloc routines were
updated to use segs_read_write_addr in instead of seg1addr when calculating
the value of r_slide to locate the item being relocated.
Changes for the 5.6 release (the cctools-316 release):
- Fixed a bug where the shared library init routines were not bound correctly
when an attempt to use prebound libraries failed. The problem was the
FULLYBOUND_STATE of the modules and the library's init_bound was not reset in
reset_module_states() in image.c (Radar bug #2428317).
Changes for the 5.5 release (the cctools-315 release):
- Fixed a bug where the floating point volatile and vector volatile registers
were not being save when a shared library init or module init routine was
called. Radar bug #2424571.
- Fixed link_library_module() in symbols.c to correctly handle fully binding
the private extern references for the shared library initialization routines.
Radar bug #2423809.
Changes for the 5.5 release (the cctools-314 release):
- Added the DYLD_ABORT_MULTIPLE_INITS feature. Radar bug #2423143.
- Fixed yet another bug relating to fully binding init routines. This time
after an init routine bound a module then a signal call was made to do a
bind fully. The code in link_library_module() in symbols.c incorrectly
set the FULLY_LINKED module back to BEING_LINKED causing it to be relocated
twice. This showed up in Classic. Radar bug #2413950 (again).
- Added the calls to unload_bundle_image() in image.c to send a
DYLD_MODULE_REMOVED and a DYLD_IMAGE_REMOVED event. Radar bug #2421760.
- The DYLD_MEM_PROTECT feature is broken in MacOS X Server because the hacks
to override mig_get_reply_port() to not call mutex_try_lock() when
vm_protect() is called did not take into account that the libc code on
MacOS X Server did not have indr(l) properly run on it. So the extra
underbar '_' prefix needs to not be there for the MacOS X Server version in
mig_support.c (so now the names with the underbar '_' prefix is ifdef'ed
__OPENSTEP__ where libc is properly indr(l)'ed). Radar bug #2418378.
Changes for the 5.5 release (the cctools-313 release):
- Fixed a bug were fully binding the dependent modules for a shared library
init routine did not fully bind modules that were already in the BEING_LINKED
state. To fix this a bit in the module state to indicate if the module is
being fully bound was needed. So two macros were added, GET_LINK_STATE and
SET_LINK_STATE, to images.h. And everywhere the link_state was set or read
was changed to used these macros. Then the macros GET_FULLYBOUND_STATE and
SET_FULLYBOUND_STATE are used to get and set the bit to indicate fully
bound. The code to set this bit is in link_library_module() in symbols.c
near the end of the routine. And the code to test for this bit is in
resolve_undefineds() in symbols.c. Radar bug #2419089.
Changes for the 5.5 release (the cctools-312 release):
- Added the routine setup_prebound_coalesed_symbols() in symbols.c to deal with
prebinding and coalesced symbols.
Changes for the 5.5 release (the cctools-310 release):
- Updated the few places that did a binary search of the undefined symbols so
that they use a linear search when the object image has MH_BINDATLOAD set.
This is done because that the order of the undefined symbols is not sorted if
-bind_at_load is used so the cases when -bind_at_load was used and the dynamic
linker will resolved the undefined symbols in the same order as the static
linker fixing a bug where something succesfully staticly linked but failed to
dynamicly link. Radar bug 2378121.
Changes for the 5.5 release (the cctools-309 release):
- Fixed a bug in the binding of image initialization routines. The were not
fully bound in that they did not cause modules the reference to be fully
bound but just rather all references in the module that has the image
initialization routine to be bound. Radar bug #2413950.
Changes for the 5.5 release (the cctools-308 release):
- Added an #ifdef __OPENSTEP__ definition of getcwd() and a declaration of
realpath() so things would compile on Openstep 4.2.
Changes for the 5.5 release (the cctools-307 release):
- Added support for coalesced symbols. Radar bug #2411273.
- Changed other places in dyld that undoes the prebinding to use the routines
undo_prebinding_for_library_module(), make_image_writable() and
restore_image_vm_protections() to simplify the code. These were added in
cctools-306 but due to risk for MacOS X PR2 not used everywhere until
post-PR2.
- Really fixed Radar bug #2378121 where symbols other that the first one in the
list may be used. The new stratagy is that this must used with an executable
that has MH_BINDATLOAD set (or when DYLD_BIND_AT_LAUNCH is set). When the
static link editor detects this case and prints a warning when it is needed.
The change that was needed was to change lookup_symbol() when
executable_bind_at_load was set, to keep searching for the first definition
of a linked symbol before using a symbol not yet linked. There are still
cases because the undefined symbols are sorted this can still be different
from the static linker even with MH_BINDATLOAD set.
- Picked up Chris Kane's changes for the DYLD_IMAGE_SUFFIX feature and the
changes to eliminate an unneeded open()/close() pair or a stat()/fstat() pair
in the cases a library has already been loaded. Radar bug #2384450.
Changes for the 5.4 release (the cctools-306 release):
- Fixed a bug where a non-lazy pointer in a library module containing only
private externs were not bound. This showed up in C++ exception handling
code in libcc_dynamic.a when copied in to a dynamic library. The fix was
in link_library_module() in symbols.c . The new routine add_reference()
was added to share the code to add references to undefined symbols.
There also was a bug when bind_lazy_symbol_reference() is binding a symbol
that is a private extern where the module that defined the private extern
also has non-lazy references. So in bind_lazy_symbol_reference() code was
added to add the non-lazy references from the module before calling
link_in_need_modules() for the case where the lazy symbol being bound is
a private extern. Radar bug #2391709.
- Fixed some bugs where library modules that containing private externs and
external relocation entries which are prebound, do not have the prebinding
backed out when the prebinding can't be used. Two changes were made to
link_library_module(). The first is where the module has a reference to a
private extern, if the module that defines the private extern is in the
PREBOUND_UNLINKED state its prebinding is undone. The second is where the
library initialization routine is fully bound at the end of the routine,
if the module that defines the library initialization routine is in the
PREBOUND_UNLINKED state its prebinding is undone.
Also there was a change was made to bind_lazy_symbol_reference() where
it binds a symbol pointer for a private extern, if the module is in the
PREBOUND_UNLINKED state its prebinding is undone.
To support this the following routines were added to reloc.c:
undo_prebinding_for_library_module(), make_image_writable() and
restore_image_vm_protections().
Changes for the 5.4 release (the cctools-304 release):
- Ifdef out the change below in cctools-303. This caused the fully binding
case with signal handlers to be very slow. This showed up in launching
the Terminal app.
Changes for the 5.4 release (the cctools-303 release):
- Added code in link_library_module() in symbols.c to add defined symbols to
the being linked list. This is so that if a definition of a symbol is not
being used from the first library that has a definition a later lazy binding
will not use a different definition of the symbol. Radar bug #2378121. Also
see the comments at the end of lookup_symbol() for a case that is not yet
handled.
Changes for the 5.4 release (the cctools-300 release):
- Added support for library initialization routines.
Changes for the 5.4 release (the cctools-299 release):
- Made two changes the _stub_binding_helper_interface for PowerPC. The first
was to remove the saving and restoring of r2 (toc) into the linkage area.
This first change is needed to allow CFM vector libraries too work correctly
as it was a bug in _stub_binding_helper_interface as the value of r2 (toc) is
saved and set by the caller. The second change was to increase the frame
from 64 to 128 and change saving the registers that could contain parameters
to be saved in the _stub_binding_helper_interface's saved register area
rather than the caller's parameter save area. This second change is needed
as the caller many not have allocated a parameter save area large enough to
save all possible parameter registers if it only calls parameterless
functions. Also did some minor changes, restored the parameters registers
in low address to high for better cache behavior, and reschedule the getting
of old link register value before parameter restores. Radar bug #2358962.
Changes for the 5.4 release (the cctools-298 release):
- Added support for "@executable_path/" in library install names and picking
up the kernel's passing the exec_path above envp's trailing 0.
Changes for the 5.4 release (the cctools-297 release):
- Fixed a bug in cctools-296 where the size of a symbol_block changed with the
new remove_on_error field and that this was declared in assembly code which
was not changed.
Changes for the 5.4 release (the cctools-296 release):
- Added the NSLINKMODULE_OPTION_RETURN_ON_ERROR option to NSLinkModule() and the
new API NSLinkEditError() that returns the error information.
- Fixed a bug in NSAddLibrary() that did not load the dependent libraries.
Changes for the 5.4 release (the cctools-295 release):
- Fixed a bug in libdyld function _dyld_image_containing_address() which was
addring the vmslide to the mach header address that it should not have done.
Changes for the 5.3 release (the cctools-294 release):
- Added the libdyld function _dyld_image_containing_address() to
dyld_libfuncs.c. Radar bug #2325380.
Changes for the 5.3 release (the cctools-293 release):
- Cut over to using the zone protect stuff in libc_dynamic.a now that it has
suppose to appear in Beaker1K. The file malloc.c is #ifndef'ed __MACH30__ .
This was Tracker bug #48087 and Radar bug #2148087.
- Fixed the call to pthread_mutex_init() in malloc.c to have NULL as the second
argument. Also fixed a cast to call to malloc that needed a `*'.
Changes for the 5.3 release (the cctools-292 release):
- Ported malloc.c to used pthreads instead of cthreads.
- Added some braces to get rid of the warning message from the egcs compiler
"warning: suggest explicit braces to avoid ambiguous `else'". This was done
to mod_init_funcs.c
Changes for the 5.3 release (the cctools-288 release):
- Fixed a bug where on PowerPC if the RESET_LAZY_REFERENCES NSUnLinkModule()
option is used in a program that is not linked with crt1.0, bundle1.o or
dylib1.o that has the ".long dyld_stub_binding_helper" fix.
- Changed call_module_initializers() in mod_init_funcs.c to use the iinit and
ninit fields masked by 0xffff. This is done to get a dyld out there before
the high half of these feilds are used for module termination indexes and
counts. This will avoid programs crashing that have libraries that have
module termination routines but are running with an old dyld. With this dyld
the library module termination routines won't get called but the program won't
crash when the module initialization routines get called.
- Changed the printf() routine in errors.c to print to stderr not stdout and
not call fflush() (Radar bug #2274737).
Changes for the 5.3 release (the cctools-287 release):
- Support for private images has been added. Including support for the new API
NSLookupSymbolInModule().
- The support for NSUnLinkModule for only MH_BUNDLE file types has been added.
- The support for the private NSMakePrivateModulePublic() call need for dlopen()
has been added.
Changes for the 5.3 release (the cctools-286 release):
- Change the use of the i386 cpp macro to __i386__ and ppc to __ppc__
- Added the support for the module termination functions. In this version they
are only called for things loaded with NSLinkModule() when NSUnLinkModule()
is called. To test this the ifdef TESTING_MOD_TERM needs to be set.
This effected the files images.c, mod_init_funcs.c, dyld_libfuncs.c,
mod_init_funcs.h and images.h.
Changes for the 5.3 release (the cctools-285 release):
- Made a hack to errors.h to define the missing errno numbers for __MACH30__
like it is done for __MWERKS__.
- Added #define'ing NeXT if not defined to mig_support.c
- Also picked up a change to the server_loop() in debug.c which did not
correctly reset the thread suppend count of the thread_that_has_dyld_lock
when it's suppend count was not 1.
- Picked up fixes in dyld_debug.defs and debug.c to the Mach 3.0 ported code
made by David Peterson.
Changes for the 5.3 release, MacOS X bring up (the cctools-282 release):
- Changed host_self(), task_self(), thread_self() and port_deallocate() to
mach_host_self(), mach_task_self(), mach_thread_self() and
mach_port_deallocate() for MacOS X. Also included "stuff/openstep_mach.h" for
macros to allow it to still build on Openstep. Also changed ifdef's __SLICK__
to __OPENSTEP__.
- Changed the Makefile to allow for RC_OS=macos for MacOS X builds.
- Added an #ifdef __MACH30__ in _dyld_fork_child() and _dyld_fork_child_final()
in dyld_libfuncs.c for resetting mach_task_self_ .
- Added an #ifdef __MACH30__ #import <mach/thread_switch.h> to lock.c
- Ifndef'ed out __MACH30__ for all of mig_support.c .
- Ifdef'ed the use of vm_region() in images.c for __MACH30__ differences and
also changed port_t to mach_port_t.
- Changed port_t and thread_t to mach_port_t in debug.h and debug.c. Added
a typedef for msg_type_long_t for MacOS X bring up while the mig stuff is
ifdef'ed out to debug.c. Change PORT_NULL to MACH_PORT_NULL in debug.c.
Ifndef'ed out __MACH30__ all of start_debug_thread() and server_loop() in
debug.c for bring up.
- Added a typedef for msg_return_t to dyld_event_MsgError.h MacOS X bring up
while the mig stuff is ifdef'ed out.
- Changed thread_t to mach_port_t in lock.h and lock.c.
- Ifdef'ed out the use of the profileServer stuff in dyld_init.c and images.c
for __MACH30__ bring up.
- Ifdef'ed out most of the mig stuff in the .defs files for __MACH30__ for
bring up. And added some matching hacks in debug.c.
- Made a hack in the Makefile and if'ed mwcc the .s.o rule to use cc not
$(CC). As the MetroWerks compiler does not take .s files.
- Added #import <stuff/bool.h> to dyld_init.h so it would compile with the
MetroWerks compiler.
- Added the following:
#ifdef __MWERKS__
#define EBADEXEC 85 /* Bad executable */
#define EBADARCH 86 /* Bad CPU type in executable */
#define ESHLIBVERS 87 /* Shared library version mismatch */
#define EBADMACHO 88 /* Malformed Macho file */
#endif
to errors.h as the MetroWerks headers don't defined this.
- Removed the asm() in error.c to declare error_string into section_order.s to
make it compile with the MetroWerks compiler.
- Removed the asm() in symbos.c to declare symbol_blocks into section_order.s to
make it compile with the MetroWerks compiler.
- Changed the declaration of the address field dyld_funcs in dyld_libfuncs.c
from:
unsigned long address;
to:
void (*address)(void);
and changed the casts in the initialization to make it compile with the
MetroWerks compiler.
- Changed the #undef task_self() to #undef task_self in dyld_libfuncs.c to make
it compile with the MetroWerks compiler.
- Added an #import "stuff/bool.h" to lock.h to make it compile with the
MetroWerks compiler.
- Added an ifneq:
ifneq "mwcc" "$(notdir $(CC))"
MIG_FLAGS = $(RC_CFLAGS) $(PGFLAGS)
endif
around the MIG_FLAGS as a hack to get it to compile with the MetroWerks
compiler and not hang up on flags such as -relax_pointers -w off.
- Added #ifndef __MWERKS__ around the empty struct definitions in
inbound_message_reply in debug.c
- Changed 4 (void *) casts to (char *) casts in malloc.c to make it compile
with the MetroWerks compiler.
- Changed the Makefile to pass obscure ld flags via -Wl rather than directly
with the compiler driver.
Changes for the 5.2 release (the cctools-274 release):
- Fixed a bug in validate_NSSymbol() in symbols.c where it was return NULL
if the module is in the INITIALIZING state. This can happen if a module
initializer routine calls NSModuleForSymbol on a symbol in it's module.
The fix is to allow modules in the INITIALIZING as well as the LINKED or
FULLY_LINKED state to have info returned. Radar bug #2238301.
- Changed load_executable_image() in images.c to use the first segment in the
load commands as the seg1addr not the one with the lowest address. Radar bug
#2237029.
Changes for the 5.2 release (the cctools-273 release):
- Added a print() function to errors.c that calls printf() and then fflush() on
stdout. This hopefully will fix the problem launcher is having when piping
it's output. Radar bug #2227265. This bug was dropped from cctools-268.
Changes for the 5.2 release (the cctools-272 release):
- Changed the "bl L1" instructon to a "bcl 20,31 L1" in dyld_start.s for ppc.
Radar bug #2231996.
- Fixed two bugs in load_library_image() in images.c when many dynamic libraries
are used. The problem is that when is_library_loaded() returns TRUE the file
descriptor is not closed. This leads to the second problem where the wrong
errno is printed if the library can't be opened and then is searched for as
the errno gets over written. Radar bug #2230181.
Changes for the 5.2 release (the cctools-267 release):
- Added dyld_sample_debug for debugging shared pc sample buffer code. Radar
bug #2227592.
- Added a check in load_library_image() in images.c to make sure the file is a
plain file and prints an error for that case. This make it so the confusing
can't map error message is not printed for the directory case. Radar bug
#2227477.
- Changed search_for_name_in_path() in images.c to skip paths with no
characters in them. Radar bug #2227589.
Changes for the 5.1 release (the cctools-264 release):
- Added a definition of getenv() that always returns NULL so the reference to
_dyld_lookup_and_bind() is not pulled in by the libc_dynamic code.
This is needed for isspace() in strtod needed by vsnprintf() for locale stuff.
Changes for the 5.1 release (the cctools-262 release):
- In Titan1P the profil(2) system call got fixed. So the #ifdef SLICK code
that was in shared_pcsample_buffer() in images.c which had the divide by 2
was removed and now there is no #ifdef SLICK and the code that does not
divide by 2 is used for both Rhapsody and #ifdef SLICK. Radar bug #2215979.
Changes for the 5.1 release (the cctools-260 release):
- Added -c to all the install commands in the Makefile.
Changes for the 5.1 release (the cctools-259 release):
- Changed the Makefile to not use /usr/local/bin/install for Rhapsody but to
just call strip and then install to make it work.
Changes for the 5.1 release (the cctools-258 release):
- Fixed shared_pcsample_buffer() in images.c to use the 4.4BSD profile header.
Also delt with the bug in profil(2) in sizing the buffer. See the detailed
comments in the profileServer notes file. Radar bug #2204555.
Changes for the 5.1 release (the cctools-257 release):
- Changed dyld -seg1addr from from 0x12000000 to 0x41100000 in the Makefile.
- Changed the Makefile to use /usr/local/bin/install for Rhapsody for dyld
install to get at -sS option.
Changes for the 5.1 release (the cctools-253 release):
- Changed the default Framework fallback path to match the new directory
layout for Preimer.
Changes for the 5.1 release (the cctools-250 release):
- Changed the ppc version _try_to_get_lock in machdep_lock.s to a version that
lends itself to stepping over. The old code could not be run in single
step mode as it would alwasy loose the reservation. The new code can be
run to the next branch and not loose the reservation. This is to allow some
single step profiling tools.
Changes for the 5.1 release (the cctools-249 release):
- Changed the use of vsprintf() to vsnprintf() in errors.c in two places now
that it exists in the 4.4bsd. The code is in the else clause of an ifdef'ed
__SLICK__ . This was filed as Tracker bug #48089.
Changes for the 5.1 release (the cctools-248 release):
- Fixed a bug where dyld could not slide itself on ppc. The problem was caused
by calling vm_flush_cache() which should not have been called. The cache
flushing does not have to be done when the code in compiled -dynamic. See
dyld_reloc.c for changes and comments.
Changes for the 5.1 release (the cctools-247 release):
- Added MKDIRS if code to the Makefile to build native on Rhapsody.
- Changed the Makefile to use gnutar instead of tar because it works on Rhapsody
where tar does not.
- Changed the code ifdef __TEFLON__ to ifndef __SLICK__ (where __TEFLON__ will
nolonger be defined for Rhapsody builds) so the default builds will be native
Rhapsody builds. The changes were dyld_init.c, errors.c, images.c and
malloc.c
Changes for the 5.0 release (the cctools-235 release):
- Fixed the PPC_RELOC_BR24 to correctly check for overflow. It now checks
the top 7 (not 6) bits are the same.
Changes for the 5.0 release (the cctools-226 release):
- Fixed a bug when the $HOME environment variable was not set. The variable
home was initialized to NULL and then later a strlen was done on it. The
fix was to in initialized it to "/".
- Fixed a bug in ppc_reloc where it was not correctly calculating the immediate
of a HA16 reloc by checking the 0x8000 bit of the other half and and sign
extending it and adding it to form the correct immediate.
Changes for the 5.0 release (the cctools-224 release):
- Defined NO_INDR_LIBC if __TEFLON__ is defined in malloc.c and errors.c so it
will build with SDK2.
Changes for the 5.0 release (the cctools-222 release):
- Changed pickup_environment_variables() in dyld_init.c to not use atol() in
getting the value for DYLD_PREBIND_DEBUG because it pulls in stuff in the
TEFLON libc_dynamic.a that fails to link (locale stuff for "." which pulles
in dyld_lookup_and_bind("__environ")).
- Added some ifdef's for NO_INDR_LIBC because the libraries are not indr'ed
for the Teflon SDK2.
In mach_error() in error.c for not using mach_error_string to avoid pulling
in mach_error from -lc_dynamic and causing a multiply defined for
"_mach_error". In malloc.c changed _set_malloc_singlethreaded to
set_malloc_singlethreaded without the underbar.
- Added -lcc_dynamic to the link line because for PowerPC the compiler appears
to call libgcc functions unlike other architectures.
- Fixed a bug in stub_binding_helper.s for ppc which was incorrectly
referencing _bind_lazy_symbol_reference in the assembly code without the
leading underbar.
Changes for the 5.0 release (the cctools-221.4 and cctools-222 releases):
- Added including <mach-o/ppc/reloc.h> for ppc in images.c .
- Picked up changed entry_point.c to which changed struct ppc_thread_state_t to
ppc_thead_state_t.
Changes for the 5.0 release (the cctools-221 release):
- Picked up the changes for the correct Rhapsody thread state changes into
entry_point.c.
Changes for the 5.0 release (the cctools-219 release):
- Added searching fallback paths when dynamic libraries can't be found. The
Two paths are DYLD_FALLBACK_FRAMEWORK_PATH and DYLD_FALLBACK_LIBRARY_PATH.
If these are not set they default to $(HOME)/Library/Frameworks:
/LocalLibrary/Frameworks:/NextLibrary/Frameworks and $(HOME)/lib:
/usr/local/lib:/lib:/usr/lib (bug #69135).
Changes for the 5.0 release (the cctools-218 release):
- Updated the PowerPC locking code with suggestions from colleagues.
- Updated ppc_cache_flush in cache_flush.s with suggestions from colleagues.
Changes for the 5.0 release (the cctools-217 release):
- Moved sparc_cache_flush.s to cache_flush.s and added _ppc_cache_flush in it
that uses the "icbi" instruction cache block invalidate instruction. Changed
the relocation routines in ppc_reloc.c to use this.
- Wrote the machine dependent lock code for the PowerPC.
- Wrote __dyld_start for the PowerPC.
- Wrote _stub_binding_helper_interface for the PowerPC.
- Wrote and added ppc_reloc.c.
- Changed m98k ifdef to ppc in malloc.c.
- Added PowerPC support in for PPC_RELOC_PB_LA_PTR in images.c in 2 places.
- Added PowerPC support in get_entry_point() in entry_point.c.
Changes for the 4.2 release (the cctools-207 release):
- Fixed a bug in the DYLD_FRAMEWORK_PATH where if the framework has a suffix
like _profile then DYLD_FRAMEWORK_PATH does not work. The fix in in
get_framework_name() in images.c (bug #75900).
Changes for the 4.1 release (the cctools-203 release):
- Added DYLD_FRAMEWORK_PATH to search for frameworks. If set and a dynamic
library with a path of the form "Foo.framework/Versions/A/Foo" or
"Foo.framework/Foo" (where the last component Foo is the framework name and
the component A is any name) then "Foo.framework/Versions/A/Foo" or
"Foo.framework/Foo" would be appended to each element in the
DYLD_FRAMEWORK_PATH. This is tried before DYLD_LIBRARY_PATH if it is set.
This mostly effected images.c (bug #70403).
Changes for the 4.0 release (the cctools-200 release):
- The dyld debug interface can cause the program using it to crash if the target
task has memory mapped from a file that has been removed over NFS. The
problem was the code in libdyld for the dyld debug interface was crashing when
it used memory in response to a successful vm_read on this memory. Two
changes where made in libdyld. First was to vm_read from the target task at
the prefered address of dyld before going into the loop that scaned the
address space. Second in the loop that trys to start the debug thread only
100 trys are made in the loop that is doing the thread_switch waiting for the
debug thread to start. If 100 trys are made it is assumed the debug thread
can't be started and it returns failure. This causes gdb to not crash and not
hang and return continue without symbols from libraries. With this a back
trace is possible. And one can still force library symbols to be loaded with
the gdb add_module command (bug #65770).
Changes for the 4.0 release (the cctools-198 release):
- Added some code in the error case when a dynamic library was thin and for the
other byte sex so it reported the library as wrong architecture instead of
malformed (bug #65420).
Changes for the 4.0 release (the cctools-193 release):
- Fixed a bug in dyld's handling of library names for NSAddLibrary() where it
did not save away the name of the library. It simply used the pointer passed
to it which caused future loads of librarys to get confused it the same space
is used to load another library as dyld thinks it is already loaded (bug
#63579).
Changes for the 4.0 release (the cctools-191 release):
- Fixed a bug where the call backs for registered for adding images was not
being called immediately after NSAddLibrary() was finished. This delayed the
objc fixup until the next trip through dyld (bug #62882).
Changes for the 4.0 release (the cctools-190 release):
- Fixed a bug in the checking for symbols in the executable that are overriding
library symbols. The look in check_executable_for_overrides() was wrong in
that dyst->iextdefsym was not added to the terminating value the index. This
caused some to try to use prebound libraries when they shouldn't. This
happened with the program gdb and the symbol catch_exception_raise (bug
#62897).
Changes for the 4.0 release (the cctools-188 release):
- Fixed a bug in undoing the prebinding on RISC machines. RISC machine
instructions only hold part of a 32-bit expression that is being relocated.
The other half of the expression is in a relocation entry. When prebinding
has to be undone the process has two steps. First the prebound symbol value
is substracted from the relocated expression. Second the correct value of
the symbol is added back to the relocated expression. The bug in dyld is
that the relocated expression is split between the instruction and a
relocation entry and the part in the relocation entry is not updated.
So when the second step is done it does not reconstruct relocated expression
correctly. The fix was to update the part of the relocated expresion stored
in the relocation entry by first making the linkedit segment writeable and
then storing the update half of the immediate (bug #62117).
Changes for the 4.0 release (the cctools-186 release):
- Fixed a bug in _dyld_bind_fully_image_containing_address() in dyld_libfuncs.c
which did not deal with modules in the PREBOUND_UNLINKED state and call
undo_external_relocation() before calling link_library_module() (bug #61399).
Changes for the 4.0 release (the cctools-181 release):
- Fixed a bug that cause dyld to crash when undoing the external relocation
relocation entries in a prebound dynamic library. It fails to make the
segments which are read-only writeable before doing the relocation and gets
a bus error when trying to write on read-only memory. This had to be fixed
in three places before undo_external_relocation() was called. The three
places were in resolve_undefineds() in symbols.c, _dyld_bind_objc_module()
in dyld_libfuncs.c, and server_dyld_debug_bind_module() in debug.c
(bug #59553).
Changes for the 4.0 release (the cctools-180 release):
- Change the sparc_reloc() routine to use the sparc flush instruction for
sync'ing the icache (bug #59327).
Changes for the 4.0 release (the cctools-178 release):
- Fixed a bug where a if a non-prebound library gets loaded after the initial
launch undo_prebinding_for_library() mistakenly gets called (bug #58805).
- To allow dyld to be linked with libc_dynamic.dylib (the libc_static code
compiled -dynamic) two things were done. The first was to add a definition
of abort() to errors.c. abort() is used by malloc() when it fails, so the
abort() in errors.c just exits with DYLD_EXIT_FAILURE_BASE+DYLD_UNIX_RESOURCE.
This eliminates undefined symbol __dyld_bind_fully_image_containing_address
because signal.o gets pulled in with abort.o. Second the DYLD_HANG_TIME
feature was removed. This eliminates undefined symbol __dyld_lookup_and_bind
which used in getenv.o via timezone stuff and sleep (bug #58557 and #50228).
Changes for the 4.0 release (the cctools-177 release):
- Fixed a bug in external_relocation() and undo_external_relocaion() in
sparc_reloc.c where reloc types SPARC_RELOC_HI22 and SPARC_RELOC_LO10
had two statements reversed. The addition of the value to the immediate was
done before the extraction of the immeditate (bug #58729).
- Corrected the wording for compatibility version mismatch from "less than" to
"greater than" (bug #58644).
Changes for the 4.0 release (the cctools-176 release):
- Changed _dyld_fork_child() and _dyld_fork_child_final() from calling
mach_init() to calling task_self() and setting task_self_ which fixed the
DYLD_MEM_PROTECT feature and allowed the system to boot with fork stubs that
used _dyld_fork_child_final().
- Added _dyld_lookup_and_bind_objc() (bug #58442).
- Optimized undo_prebinding_for_library() by just setting the module_state to
PREBOUND_UNLINKED instead of undoing all the relocation for all the modules
(bug #58434).
- Fixed a bug where the time stamps of the libraries recorded in the executable
did not match the libraries' time stamp prebinding was not correctly
disabled. The check was added in map_library_image() to match the check that
was in is_library_loaded() (bug #58387).
- Fixed a bug in using just prebound libraries without prebound executables.
A check needed to be made that the executable does not defined any symbols
that the libraries defines and reference. check_executable_for_overrides()
in symbols was added and this is called in try_to_use_prebound_libraries()
in images.c (bug #58357).
- Fixed a bug in external_relocation() and undo_external_relocaion() in
sparc_reloc.c where reloc types SPARC_RELOC_WDISP22 and SPARC_RELOC_WDISP30
were incorrectly assumed to have a pair (bug #58436).
Changes for the 4.0 release (the cctools-175 release):
- Added support for using just prebound libraries without prebound executables.
Changes for the 4.0 release (the cctools-173 release):
- Implemented the stuff needed for the dyld library functions: NSNameOfModule,
NSLibraryNameForModule, NSNameOfSymbol, NSIsSymbolNameDefined,
NSLookupAndBindSymbol, NSAddressOfSymbol, NSModuleForSymbol, NSAddLibrary.
The first three are allowed to be called in a multiply defined handler.
- Removed the default link state from the image struct and the enum. This is
because for prebinding we are picking up which modules are bound from the
LC_PREBOUND_DYLIB commands.
- Added the support for using prebound images. This touched a bunch of stuff in
dyld_init.c and in images.c.
Changes for the 4.0 release (the cctools-171 release):
- Changed the check of the size of the shared pc sample buffer in
shared_pcsample_buffer() to round to sizeof(unsigned short) because gprof
complains when the number of samples in not a multiple that.
- Fixed a bug with shared_pcsample_buffer() which did not close the file
descriptor for the shared pc sample buffer. This caused getty to fail which
was expecting to get file descriptor 0.
Changes for the 4.0 release (the cctools-170 release):
- Picked up the changed from the sparc project to dyld_reloc.c which added
3 ifdef __DYNAMIC__ sequences so dyld could be compiled without -dynamic.
- Added the DYLD_INSERT_LIBRARIES feature which is set to a colon separated
list of libraries that are inserted as the first libraries to be searched.
- Changed the support for fork() so that DYLD_MEM_PROTECT works by adding a
_dyld_fork_child_final() call and changing _dyld_fork_child(). The change
to _dyld_fork_child() was to save away the value of dyld_mem_prot and set
it to FALSE. Then in _dyld_fork_child_final() mach_init() is again call and
dyld_mem_prot is restored. This is done as the code in the fork stub messes
with the ports and the vm_protect() call fails.
- Added in the real shared pcsample buffer interface called to dyld in
dyld_init.c and images.c.
Changes for the 4.0 and Argon releases (the cctools-166.10 and cctools-170
releases):
- Changed THREAD_NULL to MACH_PORT_NULL so cctools will build in Argon.
Changes for the 4.0 release (the cctools-169.1 and cctools-170 releases):
- Fixed a bug in allocate_with_no_error_check() in allocate.c where it did not
create the zone if it was NULL.
Changes for the 4.0 release (the cctools-169 release):
- Fixed a bug in map_library_image() that was saving the compatibility_version
incorrectly from the user of the library not from the LC_ID_DYLIB of the
library.
Changes for the 4.0 release (the cctools-168 release):
- Picked up sparc changes to stub_binding_helper.s, sparc_reloc.c, dyld_reloc.c
and dyld_start.s .
- Added shared pcsample buffer stuff. It is turned off right now with an ifdef
SHARED_PCSAMPLE_BUFFERS_TESTING in dyld_init.c and images.c waiting for the
server to be written.
- Added the functions _dyld_fork_prepare, _dyld_fork_parent and _dyld_fork_child
to fix bug #55484 and to allow DYLD_MEM_PROTECT to work.
- Change the version numbers of dynamic library to the form X[.Y[.Z]] to match
how next does versioning of it's projects. This effected libtool(1), ld(1),
otool(1), dyld, libstuff(version_number.c) and man pages for ld and libtool
bug #55627). In here the changes were in images.c .
- Fixed the i386 locking code in machdep_lock.s to use the xchgl instruction
and also added the lock prefix in the old ifdef'ed out code (bug #55469).
- Added back a few const qualifiers now that -dynamic does not put these into
the TEXT segment.
- Put symbol_blocks and error_string into their own section and ordered them
at the end of the data section (see asm()'s in symbols.c and errors.c). The
file section_order.s was added to define the section load order and is loaded
first. This gets the data segment to one page plus the page(s) needed for the
symbol blocks (which will not be needed for rocket launching).
- Tuned the data structure sizes so no malloc()'s are done by default.
- Changed allocation of symbol_list structs to come out of an array of
symbol_blocks so to dirty fewer pages.
- Added a pool to allocate module_state's out of for libraries.
- Changed what was "struct module" into module_state which is typedef'ed char
to save allocated space for it (as a struct it was subject to a minimum
alignment of a struct which was 2 or 4 bytes).
- Fixed a bug introduced in cctools-167 where the dylib_name was free()'ed and
it shouldn't be. Also added printing "dyld: " to DefaultMallocError() in
my copy of malloc so I can tell it is the malloc in dyld that got hosed.
- Added _dyld_debug_module_name() to the dyld debug interface to get image's
true names.
- Cleaned up a memory leak in server_dyld_debug_defining_module() in debug.c
where the out of line data for name was not vm_deallocated().
Changes for the 4.0 PR1 release (the cctools-166.7 and cctools-168 releases):
- Changed the RELINKED state to FULLY_LINKED and made it first class state.
This is so multiple calls that fully bind modules are fast after the first
call (bug #55192).
- Changed is_library_loaded() in images.c to compare device and inode numbers
if the name is not the same (bug #55085).
Changes for the 4.0 PR1 release (the cctools-166.4 and cctools-168 releases):
- Fixed a problem with bug fix #53397 in fully binding previously linked
modules. The problem was when a previously linked module was relinked it
got relocated a second time. To avoid this a new link state, RELINKED is
used by link_library_module() and resolve_undefineds() in symbols.c.
Changes for the 4.0 release (the cctools-167 release):
- The name of the executable, argv[0] is saved in images.c and then used in
printing error messages in error.c (bug #54144).
- Added a string table in images.c to save object file images names. Also
changed it so dylib names are used from the dylib id command when they
are not changed do to DYLD_LIBRARY_PATH which avoids allocating space for
them in the default case.
- Fixed a reported bug if DYLD_LIBRARY_PATH is set to nothing then something
bad happens which I could not reproduce (I think a library gets loaded twice)
(bug #54449).
- Added the DYLD_HANG_TIME feature for supporting sampling of launch times.
Changes for the 4.0 PR1 release (the cctools-166.3 and cctools-167 releases):
- Picked up the changes made to malloc for Lantern1X in dyld's copy.
- Fixed a bug in load_library_image() when DYLD_LIBRARY_PATH is set where the
allocated name is not allocated with enough space when the search fails and
the original name is used (bug #54542).
- Changed the way module initializers are called. They are now started out
delayed (that is not called). Then when _dyld_make_delayed_module_initiali-
zer_calls() is called (from crt1.o) the delayed initializers are called and
future initializers are no longer delayed (bug #54334).
- Added the function _dyld_bind_fully_image_containing_address() to be used
to fully bind functions that are installed as signal handlers. This required
the addtion of a bind_now parameter to the following routines:
link_in_need_modules(), link_library_module() and resolve_undefineds().
The link_library_module() bind_now does the same thing as link_object_module()
in that it binds lazy symbols in when TRUE. Also both of these functions
now have an extra check when checking for multiply defined symbols to see
if the symbol that is already defined is in the module being bound as a
module may be lazy bound and later be fully bound (bug #53397).
Changes for the 4.0 PR1 release (the cctools-166.2 and cctools-167 releases):
- Fixed a bug in new_object_image() when more than NOBJECT_IMAGES (10) were
loaded (bug #53456).
Changes for the 4.0 release (the cctools-165 release):
- Fixed the incorrect asumption that 0 was an invalid file descriptor and
changed this to -1. This was done in map_image() and the call to map_image()
for mapping in object file images.
- Also added the DYLD_EBADEXEC_ONLY and DYLD_BIND_AT_LAUNCH environment
variables.
Changes for the 4.0 release (the cctools-161 release):
- Fixed a bug in _dyld_moninit() which had the wrong increment expression for
steping through the list of libraries. This was:
for(i = 0; i < q->nimages; q++){
which should have been:
for(i = 0; i < q->nimages; i++){
This also appeared in two other places in dyld_libfuncs.c and in two places
in debug.c. In the _dyld_moninit() case this caused only the first dylib
to get profiled.
- Also added some ifdef MALLOC_DEBUG in dyld_init.c and allocate.c so that
libMallocDebug.a could be switched with malloc.o and used with malloc debug.
Changes for the 4.0 release (the cctools-160 release):
- Fixed the i386 machine dependent try_to_get_lock() assembly routine to use
a bts (bit test and set instruction) because the use of the xchgl was just
wrong. Some how the code would work but it would cause the threads to spin
in trying to get the lock.
- Fixed a problem with monaddition being called from dyld with the lock being
set and then later trying to dynamicly bind things resulting in a dyld
dead lock error. This is not thread safe and could crash if objects are being
unloaded while monaddition is being called.
Changes for the 4.0 release (the cctools-158 release):
- Added _dyld_fork_mach_init() which recalls mach_init() for use in the child
side of fork so that the fork code can be in a dynamic library. This is
needed as the value of task_self_ has changed and is an invalid port in the
child and if used the in the dyld code it will fail.
- Changed it so that check_and_report_undefineds() gets called with the lock
still taken and it releases the lock before calling the user's handler. Then
after the handler is called the lock is set again. Made the same basic
changes to the register functions and module initialization function calls.
- Added saving registers r20, r27 and r29 for the hppa in the
_stub_binding_helper_interface which are assumed to be saved by a millicode
call. Also changed the hppa interface to add 64 bytes to the stack on entry
because in the case of binding a millicode call in a leaf function it can't
assume the caller will allocate space. These two things were causing the
hppa version of Lantern1P not to work.
- Fixed a few problems in load_library_image() with the code that deals with
constructing library names with dyld_library_path.
Changes for the 4.0 release (the cctools-154 release):
- Fixed a bug that caused non-lazy pointers for N_INDR symbols not to get set.
The problem was that the N_INDR symbol did not end up on the being_linked list
and thus the non-lazy pointer for it was never set.
The first part of the fix was in resolve_undefineds(), link_library_module()
and link_object_module() to have the call to lookup_symbol() not to chase down
the indirection loop but just find the N_INDR symbol.
This was done by adding the special NO_INDR_LOOP argument as the indr_loop
argument to lookup_symbol().
The second part of the fix was in relocate_symbol_pointers_in_object_image()
and relocate_symbol_pointers_in_library_image() if the symbol on the
being_linked list was an N_INDR symbol a lookup_symbol() is called to get
the correct value of the symbol before calling relocate_symbol_pointers().
- Added support for stripping defined externals that appear as indirect symbol
table entries in non-lazy pointer sections. Strip uses the values
INDIRECT_SYMBOL_LOCAL and INDIRECT_SYMBOL_ABS to represent the symbols that
are defined that were stripped. relocate_symbol_pointers_for_defined_externs()
uses these values to know wether or not to adjust the value of the non-lazy
pointer.
- Fixed uses of defined absolute symbols that was adding the vmaddr_slide to
them. This occured in a number of places. Now testing is done if the type
is not N_ABS then the vmaddr_slide is added.
Changes for the 4.0 release (the cctools-150 release):
- Added code in bind_lazy_symbol_reference() to handle private extern symbols
being lazy bound.
- Changed with the the event message that this image was added is set to after
the local relocation is done in map_library_image().
Changes for the 4.0 release (the cctools-149 release):
- Fixed the problem with the dynamic linker relocating itself when compiled with
-dyanmic as introduced in cctools-148. Three changes to _dyld_reloc() in
dyld_reloc.c were made. First was the reference to _mh_dylinker_header,
second was the reference to SEG_LINKEDIT and third was adding the missing call
to relocate_symbol_pointers_for_defined_externs() (bug #50582).
- Changed the way cache_sync_on_reloc gets set. It is now based on
S_ATTR_SOME_INSTRUCTIONS instead of S_ATTR_PURE_INSTRUCTIONS which is set by
the assembler for all sections that it assembles instructions into.
- Changed the way set uid (and set gid) programs and DYLD_LIBRARY_PATH are
handled. If the programs is setuid or setgid then dyld_library_path is not
set and the environment variable DYLD_LIBRARY_PATH is written on so that
programs that are exec'ed do not have it set.
- Added support for executables built that should be bound at launch time. If
the executable has the MH_BINDATLOAD flag in the mach header (built with the
-bind_at_load to ld(1)) then dyld_init() sets executable_bind_at_load and when
a module is to be bound all undefined references (lazy and non-lazy) are added
to the list of needed symbols. This is also supported in the mach header of
bundle files.
- Finished the multiply defined error handler so that a new symbol can be
specified to be used. This changed effected multiply_defined_error() in
error.c to mark the symbol as discarded (N_DESC_DESCARDED in the n_desc
field of the nlist struct). To do this the linkedit segment is made writable
to do this. This mark is then looked for in lookup_symbol() in symbols.c .
Also added was the routine change_symbol_pointers_in_images() in symbols.c
which is called from multiply_defined_error() to reset the values of symbol
pointers if the symbol changes.
- Added the arguments new_image and prev_image to multiply_defined_error()
which are a pointers to the images for the new and previous definitions of
the symbols.
- Changed the first call to vm_protect() (for libraries not objects) in
relocate_modules_being_linked() in reloc.c to use (VM_PROT_WRITE |
VM_PROT_READ | VM_PROT_EXECUTE) instead of maxprot.
- Added a check in load_executable_image() in images.c for segments that have
external relocation (S_ATTR_EXT_RELOC) to make sure the maxprot has
VM_PROT_WRITE and VM_PROT_READ.
- Removed the unused first argument symbol_name from relocate_symbol_pointers()
in symbols.c.
Changes for the 4.0 release (the cctools-148 release):
- Changed how dyld gets built. Now it gets built with -dynamic. Had to add
a fake dyld_stub_binding_helper to dyld_init.s which traps so it will link.
Added compiling with -O and not building dyld.o. Removed most of the hack in
the Makefile because this will only be built with 4.0 tools now.
- Now installs dyld with strip -S.
Changes for the 4.0 release (the cctools-147 release):
- Added support for profiling dyld loaded code.
- Added _dyld_get_image_name().
Changes for the 4.0 release (the cctools-145 release):
- Fixed references to private externs symbols in libraries so that the library
module that defines them is forced to be linked in. This is need so that
module init functions, register functions for linking and debug events get
triggered for these modules. The code is in symbols.c at the end of
link_library_module(). This now closes the issue about putting the private
extern symbols in a dylibs table of contents. These modules may never be
unlinked or replaced (just a the library can't be unloaded).
- Fixed a bug in the m68k _stub_binding_helper_interface where the calling
convention has a1 as a pointer to an area to return structures and needs to
be saved. Since this wasn't accounted for we are forced to save it and
shuffle the arguments in the assembly interface.
- Fixed three bugs in the hppa _stub_binding_helper_interface where registers
that should have been saved weren't. The first was %r28 which is a pointer
to an area to return structures. The second was %r31 which is the millicode
return address register. And third were the argument registers 2 and 3 which
should also have been saved as they are treated as caller save registers.
- Added dyld allocated memory protection via the environment variable
DYLD_MEM_PROTECT. This involved picking up malloc.c and adding the
zoneprotect stuff to it. This could have problems when used with the debugger
interface and should not be used.
- Changed the time outs for the debug mach messages from 30 seconds to 4
miniutes. The debugger was timing out in reading large dylibs. It shouldn't
process the messages when it gets them but rather just create a list of things
to do and then do them.
- Fixed a bug in the locking of the dyld data structures. What was happening
is that lock and thread_that_has_lock could be inconsistant as the lock can
get taken by a thread and it switch out of before thread_that_has_lock is
set to thread_self(). Then an other thread can find the lock set and in
checking for a dead lock will test thread_that_has_lock against thread_self()
and think it is a dead lock. The fix was somewhat extensive. Lock was
changed to global_lock and used to protect dyld_lock and
thread_that_has_dyld_lock. Now global_lock is lock to inspect or set either
of dyld_lock or thread_that_has_dyld_lock. Then to make the debugger it
depends order of setting and clearing of dyld_lock and
thread_that_has_dyld_lock are such that if dyld_lock is true
thread_that_has_dyld_lock will be vaild or THREAD_NULL. The later case can
only happen when a tread is unlocking. So in set_lock and release_lock if
it is the debug thread nothing is done because if there was a thread that had
the lock it was resumed or it's trying to release the lock and it it fine.
Changes for the 4.0 release (the cctools-143 release):
- Change it so that registered functions get called when the dyld lock is not
out. This added a new link state.
- Changed the order dependent libraries of dymanic libraries get loaded. To
get all of the executables libraries loaded before libraries dependent
libraries. The change was in map_library_image() the call to
load_images_libraries() was removed. Then a new flag,
dependent_libraries_loaded, was added to the library_image structure. Then
the routine load_dependent_libraries() was added to call
load_images_libraries() for library images where dependent_libraries_loaded
was FALSE. load_dependent_libraries() is called from link_in_need_modules().
- Added dealing with N_INDR symbols in the dynamic linker and doing checking of
N_INDR symbol loop checking. This is in lookup_symbol() in symbols.c.
- Fixed a bug that did not have the correct address of the dynamic link editor's
mach header for DYLD_IMAGE_ADDED events when the dynamic linker was not moved
from it's staticly linked address. The fix was to initialize
dyld_image_header to the value of &_mh_dylinker_header in dyld_reloc.c .
Changes for the 4.0 release (the cctools-141 release):
- Fixed a problem with module bound events for the exectable not being reported.
The was a problem in that the link_state was not set to LINKED by
call_module_initializers(). This was also a problem for objects loaded that
had no section marked with S_MOD_INIT_FUNC_POINTERS.
- Added support for runtime loading. Only loading of MH_BUNDLE files is
currently supported (no unloading, replacing). This is the minimum support
to get programs using bundles up. This includes the routines to register
functions for loading and binding for the Objective-C runtime.
Changes for the 4.0 release (the cctools-138 release):
- Changed relocate_symbol_pointers_for_private_externs() to
relocate_symbol_pointers_for_defined_externs() which adds the vmaddr_slide
to all non-lazy pointers for symbols defined in the library (private or not).
This can happen with ld -r modules where a non-lazy pointer can end up with
a defined symbol.
- Added the calling of module initializaton routines.
- Changed the locking for _dyld_lookup_and_bind() in dyld_libfuncs.c and moved
it into bind_symbol_by_name() in symbols.c because it can cause the user's
undefined symbol handler to be called and the lock must be released before it
is called.
Changes for the 4.0 release (the cctools-136 release):
- Changed dyld_start for the i386 to push a zero as a end of frame marker for
the debugger and reset %ebp to zero because crt1.o pushes it's value as the
end frame marker.
Changes for the 4.0 release (the cctools-135 release):
- Finished the dyld debug interface stuff.
Changes for the 4.0 release (the cctools-134 release):
- Started adding the dyld debug interface stuff. This stuff is incomplete at
this release but will compile cleanly.
- Picked up the sparc fix to entry_point.c which was using the wrong struct
and field to get the entry point.
- Fixed a problem with the machine dependent locks on the m68k and i386 where
the address of the lock was passed to the assembly routines and there was a
missing indirection.
Changes for the 4.0 release (the cctools-133 release):
- Stub'ed out all the machine dependent stuff for sparc so at least it can
build without problems.
- Started in on the debug interface.
- Added the setting and calling of the user's error handlers.
- Added the lock and the locking for the dyld data structures.
- Fixed a bug in dyld_start.s for the hppa where the it was calling _dyld_init
not __dyld_init as it should have been.
- Fixed a bug in dyld_start.s for the m68k where the argument for __dyld_reloc
had a sp@+ instead of sp@- for the move of the argument on to the stack.
Changes for the 4.0 release (the cctools-132 release):
- Got hppa support done (relocation, dyld_start, entry_point).
Changes for the 3.3 release (the cctools-127 release):
- Got the kernel interface support debugged (under the i386). A few fixes, in
__dyld_start for the i386 the mh parameter was put at 4 off the esp which was
changed to zero, also in __dyld_start for both the i386 and m68k the mh
argument was not removed before calling the entry point, in generic_reloc.c
in local_relocation reloc[i].field was being used were field was broken out
depending on if the entry was scattered so it was changed to use the broken
out fields and in dyld_reloc() SEG_LINKEDIT needed vmaddr_slide added to it
to allow it to run before the code was relocated.
Changes for the 3.3 release (the cctools-122 release):
- Fixed a bug in the relocate_symbol_pointers_in_library_image() that did not
calculate the symbol_index right (the two values were reversed) this bug was
already fixed in relocate_symbol_pointers_in_object_image().
- Fixed a bug in relocate_symbol_pointers() that did not add the
image->vmaddr_slide value to the address to be set with the symbol value.
Changes for the 3.3 release (the cctools-120 release):
- Figured out what is needed in the way of compiler changes need at this point.
They are:
* Change the i386 symbol stub code to fix bugs in the code (yes I gave you code
that does not work, I've tested this new stuff).
For non-pic stubs, from this:
.symbol_stub
Lfoo$stub:
.indirect_symbol _foo
ljmp L_foo$lazy_ptr
Lfoo$stub_binder:
pushl $L_foo$lazy_ptr
jmp stub_binding_helper
to this:
.symbol_stub
Lfoo$stub:
.indirect_symbol _foo
jmp *L_foo$lazy_ptr
Lfoo$stub_binder:
pushl $L_foo$lazy_ptr
jmp dyld_stub_binding_helper
For pic code, from this:
.picsymbol_stub
<< I never wrote this case down >>
to this:
.picsymbol_stub
Lfoo$stub:
.indirect_symbol _foo
call L1foo$stub
L1foo$stub:
popl %eax
movl L_foo$lazy_ptr-L1foo$stub(%eax),%ebx
jmp %ebx
Lfoo$stub_binder:
lea L_foo$lazy_ptr-L1foo$stub(%eax),%eax
pushl %eax
jmp dyld_stub_binding_helper
* Change the m68k symbol stub code to fix bugs in the code.
For non-pic stubs, from this:
.symbol_stub
Lfoo$stub:
.indirect_symbol _foo
jmp @(L_foo$lazy_ptr)
Lfoo$stub_binder:
movel #L_foo$lazy_ptr,sp@-
bra stub_binding_helper
.lazy_symbol_pointer
to this:
.symbol_stub
Lfoo$stub:
.indirect_symbol _foo
movel L_foo$lazy_ptr,a0
jmp a0@
Lfoo$stub_binder:
movel #L_foo$lazy_ptr,sp@-
bra stub_binding_helper
.lazy_symbol_pointer
For pic code, from this:
.picsymbol_stub
Lfoo$stub:
.indirect_symbol _foo
jmp pc@(L_foo$lazy_ptr-.)
Lfoo$stub_binder:
movel pc@(L_foo$lazy_ptr-.),sp@-
bra stub_binding_helper
to this:
.picsymbol_stub
Lfoo$stub:
.indirect_symbol _foo
movel pc@(L_foo$lazy_ptr-.),a0
jmp a0@
Lfoo$stub_binder:
pea pc@(L_foo$lazy_ptr-.)
bra dyld_stub_binding_helper
* Change the name of the symbol stub_binding_helper to dyld_stub_binding_helper.
I want all names associated with the dynamic linker to start with "dyld".
* Change the compiler driver to pass "-NEXTSTEP-deployment-target XXX" to as(1)
and ld(1). I see no reason to pass it to lipo(1) which is the other tool I
own which cc(1) runs.
* Change the compiler driver when "-NEXTSTEP-deployment-target 3.3" is seen to
use -lcrt1.o instead of -lcrt0.o . I would suggest we leave the default as
3.2 until we have this all working (which will be still quite some time).
* Also if the -object or -preload ld(1) options are seen always link with
-lcrt0.o (reguardless of the -NEXTSTEP-deployment-target flag).
* Add the new ld flags to the compiler driver to pass them through to ld(1) if
specified:
-dylinker
-dylinker_install_name XXX
-output_for_dyld
-keep_private_externs
- Agreed on a kernel interface for the 3.3 kernel. This was the email I sent
to David Bohman after a discussion on this:
The kernel when it is executing an a.out with an LC_LOAD_DYLINKER command will
do the following:
* Map the segments Mach-O file specified by the LC_LOAD_DYLINKER into memory.
If the vmaddress of the segments are not available it will "slide" it to the
any available address. The segment of this Mach-O will be contigious so the
kernel can determine the size of the needed memory by adding up the vmsize's
of the LC_SEGMENT commands.
* Then the kernel will use the entry point in the LC_LOAD_DYLINKER (plus any
slide amount) to start the task instead of the entry point in the a.out.
* One additional argument will be passed to the entry point in the
LC_LOAD_DYLINKER in addition to the normal argc, argv, envp. This argument is
the value of the address where the a.out's file offset 0 appears in memory.
That is the vmaddr of the LC_SEGMENT command of the a.out who's fileoff is 0
(which is the address of the a.out's mach_header in memory). If there is no
such segment in an a.out with an LC_LOAD_DYLINKER the exec(2) should fail with
and EBADEXEC or some such thing. This extra argument will be first on stack
based architectures like the m68k and i386 such that if it is removed from the
stack the interface is the same as it is to the a.out's entry point. For
example the m68k interface would be:
/*
* Runtime startup for NeXT m68k interface to the dynamic linker.
* This is the same as crt0.s with the addition of the address of
* the a.out's mach header passed as the an extra first argument.
*
* Kernel sets up stack frame to look like:
*
* | STRING AREA |
* +-------------+
* | 0 |
* +-------------+
* | env[n] |
* +-------------+
* :
* :
* +-------------+
* | env[0] |
* +-------------+
* | 0 |
* +-------------+
* | arg[argc-1] |
* +-------------+
* :
* :
* +-------------+
* | arg[0] |
* +-------------+
* | argc |
* +-------------+
* sp-> | mh | Address of where the a.out's file offset 0 is
* +-------------+ in memory.
*
* Where arg[i] and env[i] point into the STRING AREA
*/
For the RISC architectures we will needed to work out the interface as to
where this extra argument will be.
* You also want to think about what you want the dynamic linker is to do in the
case of an EBADEXEC like error (like a dynamic shared file is not found). I
could send a mach message on a specific port or something to tell the parent
that even though the exec(2) did not fail the task did not run successfully.
* We also wanted to allow the user to control where and how big his stack is.
The kernel would check the stack pointer of the LC_UNIXTHREAD and if it was
non-zero would use that as the stack pointer. The a.out file must also have
an LC_SEGMENT that allocates the stack's memory which contains the stack
pointer's value. This would be implemented in the static link editor, ld(1),
by specifying "-stack_size hex_number" and possibly additionally
"-stack_pointer hex_number". The static link editor will create a zerofill
segment named __UNIXSTACK of the indicated size and set the stack pointer to
the top or bottom depending on the way the architecture grows it's stack.
* Lastly we wanted to change the project ownership of <mach-o/loader.h> and
<mach-o/nlist.h>. The first is installed by mk project and the later is
installed by the libc project. The ownership should be the cctools project.
Here's the current copy of <mach-o/loader.h> as in the cctools project:
Just tell me when you want me to change the cctools project to start
installing <mach-o/loader.h> and the mk project can just stop installing it.
Also who do I need to talk to with reguard to the libc project and the
ownership of <mach-o/nlist.h>?
Changes for the 3.3 release (the cctools-115 release):
- Started writing the dynamic link editor.