directly for MIPS. Making these global to keep gcc4 happy broke gcc3.
Shaves a few instructions off this path as well.
Other arches to follow with similar changes.
With much help from Nick Hudson.
reorders things breaking _init() and _fini().
for i386, the {init,fini}_fallthru() functions were being output
_before_ the _init()/ _fini(), and instead of falling through to
the rest of the .init/.fini section, it would fall through to
_init()/_fini() again, and again, and again, until the stack was
exhausted.
now i386 works, but i have not checked that this works for
sparc64 as well.
for __do_global_ctors_aux() and __do_global_dtors_aux(),
to fix building with gcc3 -O3.
Discussed with Matt Thomas & Christos Zoulas, and tested on alpha & i386.
_eprol@GOTPCREL, it resolved it internally as a PC-relative instruction,
getting garbage, since the needed value is in the GOT. Add a CPP hook
to export it explicitly, this makes it work. Only active on amd64. XXX
empty for others. x86_64 (and s390, which we do not have) have .eh_frame
as a read-only section, so __EH_FRAME_BEGIN__ and the register/deregister
functions must take a const foo *.
This also fixes an ICE in gcc 3.3.1 (see gcc PR #9552)
Previously dlsym resolved to the version in crt0.o or libc which would
mean that the caller's shared object couldn't be determined correctly
using __builtin_return_address(0).
Mainly from FreeBSD, but adapted by me. Benefits of this solutions are:
- backward comptibility maintained
- existing broken binaries are fixed with a new ld.elf_so
- __mainprog_obj can be removed from crt0.o
- we do the same thing as FreeBSD
Fixes PR 22067.
OKed by Jason and Christos.
replaced mostly with .IMPSRC, so add .c wrappers for variant .o names
* any machine with a shared crt0.c that also uses common_elf (I didn't see
any) will now need a reachover wrapper in .CURDIR a la crtbeginS.c
this should fix the build breakage
and .fini sections, respectively (same as the {init,fini}_fallthru
stubs).
This is needed on PowerPC, where otherwise the linker would be unhappy
with the relocation records for the fallthru stubs when building a
shared library.
things like the .note.netbsd.ident section are provided by crti/crtn.
crti/crtn also provide the _init() and _fini() routines.
crtbegin/crtend now only provide support for ctors/dtors. This paves
the way to using the "crtstuff" provided with GCC (when we upgrade to
GCC 3.3), which provides, among other things, much better C++/Java
exception handling.
o Primary mod: alpha and i386 code for .init and .fini section fallthroughs
o Actual fallthru is not yet enabled, see csu/{alpha,i386_elf}/Makefile
o When enabled, should fix bin/12865 & should fix at least one WINE bug,
see http://www.winehq.com/hypermail/wine-devel/2001/05/0021.html
o Also, removed the function pointer hack put in to prevent gcc from
incorrectly attempting to do short calls over section boundaries,
on some architectures, rather, simply simply move ctors() and dtors()
into .init and .fini.
Other notes.
o Our crtbegin.c and crtend.c were completely in C, despite
the fact that .init and .fini section fallthroughs require machine code.
These modules implement the gnu ctors/dtors mechanism directly, eliminating
in part the need for the elf .init/.fini section fallthrough. However,
if we ever want to use the gnu-provided constructor support, we need to
do the fallthrough, at least one program (WINE) apparently uses the
feature independently, and other compilers may depend on the mechanism.
o It is amusing that the previous NetBSD _init() and _fini() were put
into the .init and .fini sections, because nothing is gained over
plain .text if they don't fall through. Cargo cult!
o Other elf ports will need to copy the alpha and i386 example mods.
yamt