gcc13 defaults to the latest version of dwarf

[nexus6-haiku]

Unlike gcc11, gcc13 seems to default to dwarf 5 that is incompatible with Debugger.
-gdwarf-3 must be passed manually and works for me.
From the recipe I can see the patch forces gcc to dwarf 2, do we know if version 3 has any issue?

[davidkaroly]

default dwarf version seems to be 4 which is specified here:

void
haiku_override_options (void)
{
  if (!global_options_set.x_dwarf_strict)
    dwarf_strict = 1;
  if (!global_options_set.x_dwarf_version)
    dwarf_version = 4;

  /* PIE does not work on Haiku, but PIC does the right thing for position
   * independant executables. So if asked for PIE, do PIC instead.
   */
  if (flag_pie && !flag_pic)
    flag_pic = flag_pie;
  flag_pie = 0;
}

see:
https://github.com/haikuports/haikuports/blob/master/sys-devel/gcc/patches/gcc-13.2.0_2023_08_10.patchset#L5280
(the same code used to be there also in 8.3 and 11.2)

[davidkaroly]

should we try something like this?

diff --git a/gcc/config/haiku.c b/gcc/config/haiku.c
index af14c02..bdd6f08 100644
--- a/gcc/config/haiku.c
+++ b/gcc/config/haiku.c
@@ -28,7 +28,7 @@ haiku_override_options (void)
   if (!global_options_set.x_dwarf_strict)
     dwarf_strict = 1;
   if (!global_options_set.x_dwarf_version)
-    dwarf_version = 4;
+    dwarf_version = 3;

   /* PIE does not work on Haiku, but PIC does the right thing for position
    * independant executables. So if asked for PIE, do PIC instead.

before

~> gcc -g -c test.c
~> readelf -w test.o | grep -A2 Unit
  Compilation Unit @ offset 0:
   Length:        0xc4 (32-bit)
   Version:       4

after

~> gcc -g -c test.c
~> readelf -w test.o | grep -A2 Unit
  Compilation Unit @ offset 0:
   Length:        0xc6 (32-bit)
   Version:       3
~>

@nielx @waddlesplash

[nielx]

I think @waddlesplash identified that it is actually not the case that we previously set dwarf version 2, but we have been at dwarf version 4 for a while now. Only it seems that GCC is now using Dwarf version 4 output for lineinfo, and our Debugger does not support that.

I suggest the actual solution is not to downgrade the default dwarf output, but to patch Debugger to support the lineinfo.

[nielx]

This was on the haiku-commits list: https://www.freelists.org/post/haiku-commits/haiku-hrev57242-srckitsdebuggerdwarf,2

[davidkaroly]

I suggest the actual solution is not to downgrade the default dwarf output, but to patch Debugger to support the lineinfo.

yes you're right, that would be a better way forward

[pulkomandy]

Can this be closed thanks to https://cgit.haiku-os.org/haiku/commit/?h=hrev57248&id=3f2464f759bbb21f49e8abf5bad429ba7d344952 ?

[waddlesplash]

I don't know about that; last I was seeing, without explicitly specifying -gdwarf-4, GCC was still generating version 5 lineinfo.

[davidkaroly]

I don't know about that; last I was seeing, without explicitly specifying -gdwarf-4, GCC was still generating version 5 lineinfo.

yep that seems to be the case: gcc -g generates DWARFv4 debug info with v5 lineinfo, which cannot be parsed by Haiku Debugger.
gcc -gdwarf-4 generates DWARFv4 debug info with v4 lineinfo which can be processed by Debugger.

[nielx]

This smells like a GCC bug, but spending about an hour looking through the code and to be honest, nothing really stands out. There seems to be a global variable, dwarf_version which is set to the dwarf version that is either set explicitly, or implicitly. Then the meat seems to be in output_line_info(), specifically line 13092, where output_dwarf_version() seems to output 5, even though I assume that other uses of this function would output 4 (like the output_compilation_unit_header().

Am I looking at the right code? And if so, does that mean that somewhere during outputting dwarf info the global variable is changed, unless the option was set by command line?

[korli]

While at here, the code for other OSes was updated to use if (!OPTION_SET_P (dwarf_version)) and if (!OPTION_SET_P (dwarf_strict))
see gcc-mirror/gcc@00f3429

An alternative like vxworks would be to set DWARF_VERSION_DEFAULT in gcc/gcc/config/haiku.h

[waddlesplash]

DWARF_VERSION_DEFAULT sounds like the best solution.

[davidkaroly]

Meanwhile Debugger now should be able to parse v5 line number info, see: hrev57309

[nielx]

Arguably the better solution anyway ;-)

[korli]

Meanwhile Debugger now should be able to parse v5 line number info, see: hrev57309

Albeit only in nightlies... or do you intend to bring this into the beta4 branch?

[davidkaroly]

Meanwhile Debugger now should be able to parse v5 line number info, see: hrev57309

Albeit only in nightlies... or do you intend to bring this into the beta4 branch?

not sure.

is there a description of the backporting process/policy for the beta maintenance branches? i.e. what kind of backports can be accepted, who makes the decision etc.

[pulkomandy]

The process is to push some changes to the refs/for/r1beta4 branch on Gerrit and it gets reviewed and merged as usual (someone has to vote +2 and merge, anyone who has commit permissions can do it).

There isn't really a policy except for "don't break things". So, no ABI changes, mainly bugfixes, and usually changes that have been deployed on the nightlies for a few days/weeks to make sure there is no regressions.

And, "Debugger does not show any sourcecode" can be considered a bug, so, we can put it in the bugfix category?

[korli]

I see 7 commits in src/kits/debugger since r1beta4, they look well concentrated on src/kits/debugger/dwarf. Debugger code isn't affected it seems.

Gerrit lets the user cherry-pick changes for review in the UI.

[davidkaroly]

The patches for parsing lineinfo v4 have been cherry-picked to r1beta branch.
Two more patches are open in gerrit, for lineinfo v5. (update: these are also merged)

[korli]

Can we close this?

[nexus6-haiku]

Fine from my side