Gcc used to create code to create trampolines (for nested functions) for
m68k without properly flushing the caches, leading to segmentation violations
on 68040/68060 systems.
Fixed by:
- importing the gcc 2.7.2.3 trampoline creation code into arch/m68k/m68k.h
- adding the OS-specific code for this into arch/m68k/netbsd.h
These changes have been reflected in gcc2netbsd.
- add /usr/include/objc to mtree
- set libobjc/Makefile INCSDIR=/usr/include/objc
- put to-be-installed includes to libobjc/Makefile into INCS=
- add the Objective-C source files to the appropriate lists.
- as the objc include files are installed now, no need to patch most
of the files.
- add /usr/include/obj/* to src/distrib/comp/mi
The changes have been applied to gcc2netbsd, and this script was then used
to create the new version of the files in libobjc/.
PIC support for GCC/arm.
This fix is untested on platforms other than the ARM and is only enabled
if GCC_27_ARM32_PIC_SUPPORT is defined.
These fixes are only required for GCC 2.7 and will not be needed in 2.8+
Norm Rubin and Richard Earnshaw.
PIC support for ARM will form part of the standard GCC tree from 2.8
onwards.
Also default to building soft-float as the resulting code will be a lot
faster then using emulated floating point instructions.
emit a call _cacheflush() to ensure Icache consistency
after writing trampoline code for lexical closures onto the stack
(i.e., into the Dcache).
GCC 2.x (x <= 7.2.2) emits a call cacehflush(), polluting the user
namespace.
memref_referenced_p(); otherwise there are cases where a later set that changes
the memory location may be moved too early, and thus the new value will be
(incorrectly) used where it should not be.
processing of /etc/ld.so.conf in ld.so itself (rather than in /etc/rc):
- added do_conf function and call to it in main to implement internal
processing of /etc/ld.so.conf
- moved maintenance of dir_list to dodir, with an added argument to
dodir to specify whether dir_list should be updated
- added option '-c' to suppress processing of /etc/ld.so.conf
- added option '-S' to suppress processing of std directories (but not
/etc/ld.so.conf)
- modified option -s to suppress processing of _both_ std directories
and /etc/ld.so.conf (i.e., it is equivalent to -cS)--this was done
so that users and scripts that interpret -s as meaning "process
only directories specified on the command line" would not be
startled
From PR #4031. One change was not committed, since it was unnecessary
(option variables are in the BSS, and do not need to be initialized to
0 explicitly).
Plus one change from me:
- getopt() returns -1 when completed, not EOF.
of the password file entry, not just first password file entry containing
$LOGNAME. (Fixes PR 3298 with a fix similar to that supplied in that PR
by Era Eriksson.)
* Ignore SIGINT and SIGQUIT (signals which are sent to the whole process
group) before editing sessions, and restore handler afterward, so that
SIGINTs and SIGQUITs delivered while editing (e.g. via Control-G in emacs)
don't cause send-pr to abort. (Fixes PRs 2530 and 2658.)
* When aborting because of a signal, if the PR has been modified from the
standard form, don't delete the edited copy (to prevent peoples' work
from being lost accidentally). (Fixes PR 2057.)
had not be implemented. It would cause an "adress space leak" and, if
the same object would opened multiple time, unwanted relocations.
Re: Comment from Chris:
"The a.out ld.so has some problems with dlclose. It doesn't properly
unmap objects which are dlclosed. That's a known problem (though a
serious one for programs which dlopen then dlclose lots of objects,
because it causes address space exhaustion), but it has a
previously-unknown side-effect.
If a single object is dlopened, then dlclosed, then dlopened _again_,
the relocations will be processed again. That causes obvious
problems."
Added a new patch mechanism to gcc2netbsd so changes from the FSF source
are seen in gcc2netbsd. Sync some Makefiles to output of current gcc2netbsd.
(Cosmetic changes only to the Makefiles.)
- separate out the common files used by rtld ldd ld
- move machine dependent files into arch
- move ld in its own directory
- factor out .PATH and CFLAGS common to all Makefiles
and hack to use mkstemp() instead of mktemp(). This *does* cause the
tempfile to be created at name generation time, but that's ok because
the code will fopen(tempname), use it, and the unlink it. Kinda cute
(``ugly but interesting'' :), but seems to work, and passes
'sh ./rcstest' as well as rudimentary tests by me.
<bsd.prog.mk> was being included before an .if !defined(NOPIC)
which added rtld to the SUBDIRs. This was so that the definition
of NOPIC would get picked up. Unfortunately, this is WRONG, because
after <bsd.prog.mk> is included you shouldn't be able to touch SUBDIR
as the magic has already happened.
However, a complementary bug in the .mk files kept anyone from
noticing until now.
Fixed by moving the include of <bsd.prog.mk> to the bottom and
including <bsd.own.mk> above the .if !defined, which is the correct
way to do this sort of thing.
support for shared libraries. Diffs from Manuel Bouyer, probably
derived from similar OpenBSD gcc diffs.
This change is in the source tree because it was used to build a NetBSD/pmax
shared-lib snapshot. The change isn't being integrated into the gcc2netbsd
script until the shared-lib diffs are finalized.
The NetbSD libc header files use GCC attributes to emit link-time warning
messages (e.g,. for gets()). SO, add a definition of the GCC back-end
macro ASM_OUTPUT_SECTION_NAME() to the mips back-end target-specific file.
This adds support for emitting warning attributes to binutils 2.6 or newer.
(Weak references may or may not work also).
This patch has been submitted to the FSF but hasn't made it into 2.7.2.1,
and seems to have got buried somewhere inside Cygnus.
the Alpha no longer needs -fno-for-scope. However, NetBSD/Alpha's
(cygnus snapshot) g++ doesn't allow virtual functions which aren't
declared inline to be later defined as inline without a warning (-> error).
Therefore, on the Alpha, define BROKEN_GXX_VIRTUAL_INLINE (which will
go away after the Alpha is upgraded to use a more recent toolchain), and
use it to make sure that declarations of virtual functions which will
be defined inline are actually marked as inline when they're declared.
the make to not fail even if the installations failed, which is wrong.
If the installations fail (e.g. because the target directory does not
exist), the whole install should fail, so that the user knows to
create the appropriate directories.
February 16th 1990, Message-ID 9002162329.AA11363@odin.think.com)
to our version of gdb. I had to do it by hand, but every thing
works fine now. I changed ns32k-pinsn.c to use the xxx_filtered
functions.
* Added functions to ns32k-nat.c to support gdb -k with crash dumps.
* Made stacktracebacks work better with execs not compiled with -g.
we're doing a dynamic link for, and change the page size used for offset
and size calculations if necessary. Allows the same ld.so to be used with
mixed m68k4k and m68k8k executables and libraries.
Thanks a million to Gordon Ross for the help in making this work!
mycroft