Hi,

One of the issues we had with libunwind for 32-on-64 unwinding was that
there are some parts that need access to local 64bit memory, but defined
an unw_word_t as wide as the target (32 bit). This patch makes it so
that for those parts we now use an address space that is indexed of a
64bit pointer (passed as arg to the accessor methods). And gets rid of
the MAP_32BIT hacks we had.

When we push this upstream, we should do the same for the libunwind
ptrace backend to show why it is useful. But I didn't want to spend too
much time on that now since we don't use it in frysk.

frysk-imports/libunwind/ChangeLog
2007-11-28  Mark Wielaard  <mwielaard@redhat.com>
 
    * src/mi/Gget_unwind_table.c (unw_get_unwind_table): Make addr a
    char *, pass addr as arg to dwarf_read functions, make
    local_access_mem use arg as base address for start indexed reads.

frysk-sys/lib/unwind/ChangeLog
2007-11-29  Mark Wielaard  <mwielaard@redhat.com>

    * cni/ElfImage.cxx: Don't define MAP_32BIT.
    (mapElfImage): Don't use MAP_32BIT.
    * cni/UnwindH.hxx: Don't define MAP_32BIT.
    (createElfImageFromVDSO):  Don't use MAP_32BIT.

Tested by also running all -arch 32 tests on a 64bit system (and also
straight make check on x86 and x86_64).

It would be nicer if we didn't need to feed an elfimage to libunwind and
libunwind didn't try to do the elf parsing itself here. But when we feed
it debug_frame info we will have the table come not from the inferior
memory directly anyway, so for now it is useful (although ideally
libunwind would just accept a pointer or addressspace reference to the
debug_frame or debug_info table and use that).

Cheers,

Mark