The problem is trying to unwind from a function where %ebp is NOT
used as the frame pointer, and the size of the frame changes over
the lifetime of that function.
For instance, trying to unwind past the GNAT runtime function
called system.tasking.rendezvous.timed_selective_wait on x86-linux,
one can get:
(gdb) bt
[...]
#3 0x0805364b in system.tasking.rendezvous.timed_selective_wait ()
#4 0xb7fe5068 in ?? ()
Backtrace stopped: previous frame inner to this frame (corrupt stack?)
Looking at the CFI, we find the following initial instructions...
> DW_CFA_def_cfa: %esp+4 (r4 ofs 4)
> DW_CFA_offset: %eip at cfa-4 (r8 = %eip)
... and the associated FDE:
> 00001be4 00000054 00001be8 FDE cie=00000000 pc=08053310..08053951
[...]
> DW_CFA_advance_loc: 8 to 080534ad
> DW_CFA_def_cfa_offset: 112
> DW_CFA_advance_loc2: 414 to 0805364b
> DW_CFA_def_cfa_offset: 108
[...]
The problem is that the DWARF frame unwinder executed the FDE until
the row for PC == 0x0805364b. But in reality, our program hasn't
executed the instruction at that address yet (it is the return address).
So GDB executed a little too much of the FDE, giving us the wrong
offset for the frame base, and thus the wrong address where %eip
got saved.
This patch fixes the problem by using a more correct PC as the bound
for executing the FDE.
gdb/ChangeLog:
* dwarf2-frame.c (dwarf2_frame_cache): Use
get_frame_address_in_block instead of get_frame_pc as
the bound for executing the frame's FDE.
gdb/testsuite/ChangeLog:
* gdb.ada/rdv_wait: New testcase.
0c92d8c1c3