In the case of this proposal, then, I suggest the CU fields
(AT_{name,comp_dir}) retain their exact current definitions. Language
implementations, regardless of whether they might want to support embedding
source, currently use the filesystem. This extension is essentially just
cacheing source which may become unavailable to the consumer by the time the program is debugged. This means the producer can put standard values in each CU field, and also embed source in the line table. If in the future there is a need to add CU fields or modify existing ones to capture some other attribute,
that can be done in a different proposal.


On 2018-02-01 17:32, Michael Eager wrote:
On 02/01/2018 08:07 AM, sc...@scottlinder.com wrote:
Hi John,

In the case where the files are actually available on disk, and the source is simply being "cached", the attributes are exactly the same. In the case where sources are generated, and so have no true path on disk, I would suggest we might just leave the exact meaning to be implementation defined; the producer can still provide valuable information which will aid in locating where sources originate, such as indicating the OpenCL kernel name. Consumers which are unaware of this extension will simply fail to find the source (as before), while new consumers can at least provide an identifier to distinguish sources.

Implementation-defined generally means that different implementations
will be incompatible.  Incompatible implementations are the antithesis
of a standard.

As a general DWARF principle, there should be no secret understandings
between producer and consumer. There should be no "secret handshake"
such as the one you describe where a producer provides "valuable
information" in some undefined manner usable only by a consumer which
is "in on the secret".  It's not that a different consumer doesn't
implement the extension, it's that a different consumer cannot implement
the extension.

Attributes which have a defined meaning, such as AT_name or AT_comp_dir,
should have a well defined meaning in all circumstances.

The remaining attributes (DW_AT_language, DW_AT_producer, etc.) seem pretty naturally orthogonal.


On 2018-01-31 14:40, John DelSignore wrote:
Hi Scott,

Question: What does the DW_TAG_compile_unit look like for an embedded
source file? For example, what does the DW_AT_name and DW_AT_comp_dir
look like?

Cheers, John D.

On 01/31/18 17:05, sc...@scottlinder.com wrote:
Hello all,

I am a compiler engineer at AMD, working on tools for debugging online-compiled programs. The problem I am attempting to solve was brought up previously in the DWARF Standard issue 161018.1 titled "DWARF-embedded source for online-compiled programs", and is the result of runtimes like OpenCL doing online compilation in an environment where it is not desireable (or even feasible) to write sources to disk. In these cases, it would be useful to support embedding the source directly in the resulting DWARF. I would like to propose a similar solution to the one outlined in the above issue, but without structural changes
to the specification.


Add two new optional fields to the file_names prologue of the line table.

Add two bullets after "5. DW_LNCT_MD5"
6. DW_LNCT_has_source
    DW_LNCT_has_source indicates that the value is a boolean which affects the     interpretation of an accompanying DW_LNCT_source value. When present there     must be an accompanying DW_LNCT_source value. When true, consumers may use     the embedded source instead of attempting to discover the source on disk.     When false, consumers will ignore the DW_LNCT_source value. This code point
    is always paired with a flag form (e.g. DW_FORM_flag or
7. DW_LNCT_source
    DW_LNCT_source indicates that the value is a null-terminated string which     is the original source text of the file. When present there must be an     accompanying DW_LNCT_has_source value. The string will contain the UTF-8     encoded source text with '\n' line endings. When the accompanying     DW_LNCT_has_source value is false, the value of DW_LNCT_source will be the     empty string. This code point is always paired with a string form (e.g.
    DW_FORM_string, DW_FORM_line_strp, DW_FORM_strp).

New type codes can be allocated for them in a backwards-compatible way, or
codes for these new content types can be added in the range of
[DW_LNCT_lo_user, DW_LNCT_hi_user] to avoid changing the spec itself.

Table 7.27:
Add DW_LNCT_has_source  0x6
Add DW_LNCT_source      0x7

Any DWARFv5 consumer which is unaware of this extension would continue to operate as before, ignoring the new fields. Any consumer which is aware of the extension would know to check DW_LNCT_has_source for each file_name entry in order to determine whether the embedded source field (DW_LNCT_source) contains
the source text of the corresponding file.


My team and I believe this simplifies the design by removing the need for changes to the compile unit sections, and by avoiding the addition of multiple file_name_entry_formats in a single program, all without sacrificing any information. We have a preliminary implementation in LLVM/Clang, which supports embedding source (clang -gdwarf-5 -gembed-source) and inspecting it via llvm-dwarfdump and llvm-objdump (with the -source flag). The patches are
available at https://reviews.llvm.org/D42765 (LLVM) and
https://reviews.llvm.org/D42766 (Clang).

I would like any and all feedback on the design, and want to see about the possibility of adding the new content type codes outside of the "user" range (i.e. adding new entries for them in Table 7.27) in the next version of the

Scott Linder

Dwarf-Discuss mailing list

Dwarf-Discuss mailing list

Dwarf-Discuss mailing list

Reply via email to