[Bug binutils/29250] New: readelf erases CIE initial register state

[amodra at gmail dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=29250

            Bug ID: 29250
           Summary: readelf erases CIE initial register state
           Product: binutils
           Version: 2.39 (HEAD)
            Status: NEW
          Severity: normal
          Priority: P2
         Component: binutils
          Assignee: unassigned at sourceware dot org
          Reporter: amodra at gmail dot com
  Target Milestone: ---

From: Vsevolod Alekseyev <sevaa@sprynet.com>

I'm debugging a DWARF parser library. We are testing it against GNU readelf,
and we've found a discrepancy on the dump of the interpreted .eh_frame section
of a particular x86_64 ELF binary.

The binary's first FDE in .eh_frame has initial_location 0x1060, and the
following instructions:
DW_CFA_advance_loc 4      # Move PC by 4
DW_CFA_undefined 16       # Change the rule for R16 to undefined

The linked CIE marks R16 as the return address, and has the following
instructions:
DW_CFA_def_cfa 7, 8       # CFA is at R7+8)
DW_CFA_offset 16, 1       # Set the rule for R16 to
[CFA+1*data_aligment_factor])


The GNU readelf, if executed with --debug-dump=frames-interp, dumps the FDE
as follows:

00000018 0000000000000014 0000001c FDE cie=00000000
pc=0000000000001060..0000000000001086
     LOC           CFA      ra
0000000000001060 rsp+8    u
0000000000001064 rsp+8    u


Meanwhile, the alternative parser thinks that at the range [0x1060-0x1064), the
rule for RA/R16 should be as inherited from the CIE, and it goes rsp+8.

I've debugged readelf (the latest master, as of 06/01/22), to that point. 
There are two passes over the FDE instructions: one starting on dwarf.c:9296,
the other starting at dwarf.c:9442. On the first pass, when DW_CFA_undefined is
encountered, there is the following case statement:

READ_ULEB (reg, start, block_end);
if (frame_need_space (fc, reg) >= 0)
  fc->col_type[reg] = DW_CFA_undefined;
break;

If I understand correctly, the intended purpose of the first pass is to
allocate enough memory in the fc->col_type and fc->col_offset arrays, and the
logic of this operator's handling was meant to be: if this register was not
mentioned before, allocate space for it, and reset its rule to undefined.
HOWEVER, if the register WAS mentioned before (e. g. in the CIE),
frame_need_space() returns 0, and the if() body executes anyway, and resets the
rule for the register to undefined, erasing the initial state as specified by
the CIE.

I think the if statement should go, instead, "if (frame_need_space (fc, reg) >
0)". Same for other register-rule-type operators on the first pass.

The binary can be seen at
https://github.com/eliben/pyelftools/issues/409#issuecomment-1136720254

-- 
You are receiving this mail because:
You are on the CC list for the bug.