[Dwarf-Discuss] line table dir/file

[Robinson, Paul]

> -----Original Message-----
> From: Philip Craig <philipjcraig at gmail.com>
> Sent: Tuesday, February 25, 2020 3:44 AM
> To: Robinson, Paul <paul.robinson at sony.com>
> Cc: David Anderson <davea42 at linuxmail.org>; dwarf-
> discuss at lists.dwarfstd.org
> Subject: Re: [Dwarf-Discuss] line table dir/file
> 
> On Tue, 25 Feb 2020 at 05:34, Robinson, Paul via Dwarf-Discuss
> <dwarf-discuss at lists.dwarfstd.org> wrote:
> > When the committee was reworking the file/dir tables for DWARF v5,
> > one thing that came up was that the root file/directory were *not*
> > in the .debug_line section, and therefore the line table could not
> > be interpreted fully without a .debug_info section, because that
> > was where the root file and compilation directory were kept--not
> > in the line table itself.  Am I imagining this discussion?  We
> > added text to the document specifically describing file/dir 0 and
> > that they were to be explicitly kept (duplicated from .debug_info)
> > so that the .debug_line section could be fully dumped even in the
> > absence of a corresponding .debug_info section.  This was the
> > "strip all but the line table" scenario.
> 
> That matches the reasoning given in the informative paragraphs in
> Section 6.2.4. As I've mentioned before, it seems partially flawed,
> since I think it was only the compilation directory that was required
> from the .debug_info section. The root file change seems unnecessary,
> and I am unsure if it is a good change (it's an unnecessary change
> which increases chances of implementation bugs, particularly since it
> is not isolated to .debug_line, but it makes the file and directory
> indexing consistent).
> 
> > Offhand it looks like there are three possibilities:
> > a) Clang got it right and everyone else got it wrong.
> > b) The spec is ambiguous and we're observing the consequences.
> > c) Clang got it wrong and everyone else got it right.
> 
> I don't think it's "Clang" vs "everyone else" because the only v5
> .debug_line producer that David looked at was Clang, and no consumers
> were looked at.
> 
> gdb appears to accept an index of 0 for v5 .debug_line, including for
> DW_AT_decl_file. I didn't test, but only looked at its source code.
> 
> libbacktrace handles v5 .debug_line, but uses 1-based indexing for
> both DW_LNS_set_file and DW_AT_call_file. This may be an
> implementation bug such as I mentioned above.
> https://github.com/ianlancetaylor/libbacktrace/blob/ca0de0517f3be44fedf5a2
> c01cfaf6437d4cae68/dwarf.c#L3171
> 
> > But in all three cases, we need to settle on (i) what the spec
> > actually says, (ii) does it say what the committee intended,
> > (iii) if not, what to do about it.
> 
> For (i), the spec says two things that seem conflicting to me:
> 
> Section 2.14, Page 50:
> The value of the DW_AT_decl_file attribute corresponds to a file
> number from the line number information table for the compilation unit
> containing the debugging information entry and represents the source
> file in which the declaration appeared (see Section 6.2 on page 148).
> The value 0 indicates that no source file has been specified
> 
> Section 6.2.4, Page 158:
> The line number program references file names in this sequence
> beginning with 0, and uses those numbers instead of file names in the
> line number program that follows.
> 
> I think there is no doubt that this intends the line number program to
> use 0-based indexing. I think it is reasonable to assume that
> DW_AT_decl_file uses the same numbering as the line number program.
> But doing so means DW_AT_decl_file cannot reference the first file
> name since a value of 0 indicates no source file has been specified.
> 
> One possibility is that Section 2.14 was missed when updating the
> standard for the new file name entry at index 0, and that a
> DW_AT_decl_line of 0 is sufficient to indicate no source file.

I think that (missing the section 2.14 reference) is quite likely.
I had always been under the impression that setting line=0 was
sufficient to indicate "no source," especially given that the
line-number program encoding makes it easier/shorter to encode a
line change than a file change.  It was something of a surprise
to me when the bit about DW_AT_decl_file came up.

So we should fix 2.14 to say DW_AT_decl_file = 0 means the root
file, but DW_AT_decl_line = 0 still means "no source."  We don't
have to fix DW_LNE_define_file because it's deprecated in v5.

I *think* that's enough to tidy up the spec.  If there are other
places that imply 1-based dir/file indexing we should fix those
too.

--paulr