cmd/compile: consider using DWARF 5

[aclements]

DWARF 5 has several advantages over previous versions of DWARF. Notably,

1. It supports position-independent representations, which significantly reduces the number of relocations in object files and hence the size of object files and the load on the linker. In the go binary, 49% of the 503,361 total relocations are in the DWARF.

2. It supports much more compact location and range list formats. The location and range list sections are 6% of the 12MiB of the go binary, even when zlib compressed.

3. It has an official language code for Go. :)

DWARF 5 is quite new, and I don't think the rest of the ecosystem is ready yet, but I wanted to get the idea floating. It is supported by the GNU and LLVM toolchains and some debuggers. Support was added in GCC 7.1 (May 2017) and GDB 8.0 (June 2017). It appears to be in the latest LLVM, which covers most of the Xcode tools, though I can't find when it was added.

It is currently not supported by LLDB or the macOS linker. We could potentially get around the macOS linker by leaving out the Go DWARF from the objects we pass to the system linker and then merging it in to the final binary (we already do a merge step). This is more feasible with DWARF5 because it's mostly position-independent, so we wouldn't need dsymutil to relocate it for us.

/cc @cherrymui @heschik @dr2chase @randall77 @ianlancetaylor

[heschi]

Strongly in favor of doing this the moment we think we can get away with it. I'm happy to do the work for location lists.

[bradfitz]

Un-proposaling this per discussion with @ianlancetaylor.

[aarzilli]

For reference, the debug info for the linux_amd64/go1.11.1 version of cmd/compile is 15MB of which 8MB are in debug_loc and debug_ranges, specifically:

total size 20752567

section             compressed    uncompressed
.zdebug_abbrev             274 B         467 B
.zdebug_line            739610 B     1931456 B
.zdebug_frame           158093 B      559020 B
.zdebug_pubnames         38451 B      251078 B
.zdebug_pubtypes         45270 B      211390 B
.debug_gdb_scripts                        42 B
.zdebug_info           1257828 B     3694829 B
.zdebug_loc             989048 B     5914085 B
.zdebug_ranges          442419 B     2461408 B

Rewriting debug_loc and debug_ranges in the new format specified by DWARF 5 will reduce the size of debug_ranges to 774460 B (31% of DWARF 4 size) and of debug_loc to 2032960 B (34% of DWARF 4 size).

Edit: debug_ranges and debug_loc were accidentally swapped in the last sentence.

[dr2chase]

Is compression still doing a better job at reducing binary size than DWARF 5, or can D5 also be compressed? I did just mostly finish a DWARF splitter for OSX that recreates the expected file in the expected place, and adds the expected UUID for matching the two files. I suspect I ought to be using a hash of the binary contents excluding debugging information, so that identical files will remain identical after UUIDs are added.

[aclements]

The compression isn't actually part of the DWARF spec, so it's orthogonal to the DWARF version and can be used with DWARF 5.

@aarzilli, thanks for doing that experiment. Can you get the numbers for DWARF 5 if it's also zlib compressed?

[aarzilli]

@aarzilli, thanks for doing that experiment. Can you get the numbers for DWARF 5 if it's also zlib compressed?

After compression the size is basically the same, 98% (of compressed DWARF4) for debug_ranges and 77% (of compressed DWARF4) for debug_loc. The compression time however is reduced by 40%.

[dr2chase]

Is that compression time+space comparing current low-effort (higher speed) versus future low-effort? Sorry to be so picky, it's just that we've made mistakes here before.

[aarzilli]

I'm comparing compression using zlib.BestSpeed for both, like the linker. Since measuring how much time it actually took inside the linker is hard I measured how much time it takes to recompress the compressed section.

Because I have send the whole section to the compressor and the linker doesn't, I get slightly better compression than the linker. But the difference is small (around 1%) and since I'm doing it for both DWARF 4 and DWARF 5 the result should be valid.

[joeyjiaojg]

Parsed a DWARF5 linux kernel module, the below statement of index 0 is not nil. It returns the real file path.

<html>
<body>
<!--StartFragment-->

// Files returns the file name table of this compilation unit as of
--
  | // the current position in the line table. The file name table may be
  | // referenced from attributes in this compilation unit such as
  | // AttrDeclFile.
  | //
  **| // Entry 0 is always nil, since file index 0 represents "no file".**
  | //
  | // The file name table of a compilation unit is not fixed. Files
  | // returns the file table as of the current position in the line
  | // table. This may contain more entries than the file table at an
  | // earlier position in the line table, though existing entries never
  | // change.
  | func (r *LineReader) Files() []*LineFile {
  | return r.fileEntries
  | }

<!--EndFragment-->
</body>
</html>

[mdempsky]

FWIW, GCC 11.1 was released on April 27, 2021. From https://gcc.gnu.org/gcc-11/changes.html:

For targets that produce DWARF debugging information GCC now defaults to DWARF version 5 (with the exception of VxWorks and Darwin/Mac OS X which default to version 2 and AIX which defaults to version 4).

Also, Clang 14 was released on March 25, 2022. From https://releases.llvm.org/14.0.0/tools/clang/docs/ReleaseNotes.html#dwarf-support-in-clang:

The default DWARF version has increased from DWARFv4 to DWARFv5. You can opt back in to the old behavior with -gdwarf-4 or -fdebug-default-version=4. Some platforms (Darwin, Android, and SCE for instance) already opt out of this version bump as is suitable for the platform