Crash during build on UltraSPARC, 32-bit glibc

[Nix <nix at esperi dot org dot uk>]

This crash happens with both glibc-2.3.2 and glibc-2.3.2+debian's
2.3.2-8.diff (roughly CVS as of 2003-07-15 with a few extra patches,
none affecting black magic areas like ld.so.)

Building glibc with gcc-3.3.x (I've seen it with 3.3, 3.3.1, and 3.3.2),
and binutils-2.14.90.0.6 (I've seen it with earlier versions too),
I get a segfault when rpcgen runs (i.e., the first time anything is
run with the new glibc).

No NPTL is used: configuration and building done via this (staid and
pedestrian) script:

,----
| #!/bin/sh
| 
| # Set up the kernel headers. asm/ is *synthetic*; an amalgam of 64-bit and
| # 32-bit headers, with autoswitching based upon compilation universe.
| 
| (cd /lib/modules/`uname -r`/build && ./generate-asm.sh include)
| 
| unset CONFIG_SITE
| export CFLAGS="-O2 -g -mcpu=ultrasparc -mtune=ultrasparc -pipe"
| export ASFLAGS="-O2 -g -mcpu=ultrasparc -mtune=ultrasparc -pipe"
| sparc32 /usr/packages/glibc/core/configure --prefix=/usr --with-fp --with-elf --enable-shared --disable-omitfp --enable-profile --disable-bounded --enable-add-ons --enable-kernel=2.4.22
| sparc32 make ASFLAGS-.os="-O2 -g -mcpu=ultrasparc -mtune=ultrasparc -pipe"
`----

(It also segfaults with -march=v8, -mcpu=ultrasparc, and may segfault
with other lesser flags too; not checked yet.)

(generate-asm.sh is nicked from Debian, RH, and everyone else. It does
the expected trickery to make a dual-universe kernel header tree.)

Debugging shows

,----
| amaterasu 458 /usr/packages/glibc/sparc-amaterasu% gdb --args elf/ld.so --library-path elf sunrpc/rpcgen
| GNU gdb 6.0
[snip warranty info]
| This GDB was configured as "sparc-unknown-linux-gnu"...
| (gdb) break _dl_sysdep_read_whole_file
| Breakpoint 1 at 0xdc88: file dl-misc.c, line 58.
| (gdb) run
| Starting program: /usr/packages/glibc/sparc-amaterasu/elf/ld.so --library-path elf sunrpc/rpcgen
| Breakpoint 1 at 0x800dc88: file dl-misc.c, line 58.
| 
| Breakpoint 1, _dl_sysdep_read_whole_file (file=0x8015fa0 "/etc/ld.so.preload", sizep=0xefffef74, prot=3) at dl-misc.c:58
| 58        int fd = __open (file, O_RDONLY);
| (gdb) bt
| #0  _dl_sysdep_read_whole_file (file=0x8015fa0 "/etc/ld.so.preload", sizep=0xefffef74, prot=3) at dl-misc.c:58
| #1  0x08002cb4 in dl_main (phdr=0x10034, phnum=65780, user_entry=0x0) at rtld.c:980
| #2  0x0801061c in _dl_sysdep_start (start_argptr=0x10034, dl_main=0x100f4) at dl-sysdep.c:233

(Things look reasonable here, in the first call to
_dl_sysdep_read_whole_file() in ld.so, but...)

| (gdb) step
| 
| Program received signal SIGSEGV, Segmentation fault.
| 0x0801234c in __syscall_error_handler ()
| (gdb) bt
| #0  0x0801234c in __syscall_error_handler ()
| #1  0x0800dc90 in _dl_sysdep_read_whole_file (file=0x2 <Address 0x2 out of bounds>, sizep=0xefffef74, prot=3) at dl-misc.c:58
| #2  0x08002cb4 in dl_main (phdr=0x8015fa0, phnum=4026527604, user_entry=0x3) at rtld.c:980
`----

(boom. Something smashed the stack. Hard. Everything above the `file'
parameter to _dl_sysdep_read_whole_file() is garbaged, including the
args to dl_main. Small wonder it crashed.)

rtld and friends are slightly black-magic to me, I'm afraid: I guess
that something's gone fubared with relocation, since it looks like this
is the first inter-translation-unit function call in rtld; but I don't
know enough about the relocation dance to guess what that might be.

Am I on the right track? Has someone else squashed this already?
Or is my cold causing me to hallucinate problems where there are
none?

-- 
`If you want a vision of the future, it is a wireless broadband network
 feeding requests for foreign money-laundering assistance into a human
 temporal lobe, forever. With banner ads.' --- John M. Ford