Re: [PATCH v2 00/32] Rewrite the DWARF "partial" reader

[Tom de Vries <tdevries@suse.de>]

On 11/22/21 8:59 PM, Tom Tromey wrote:
> Tom> I'll send v3 of the series once I do some more testing... I plan to
> Tom> regression test with a few different approaches (dwz, probably
> Tom> split DWARF too).
> 
> I think the branch is regression free in the normal case and when using
> the cc-with-dwz-m board.
> 
> I also tried the fission-dwp board, and I have a single regression.  I'm
> inclined to ignore this one, but I figured I'd describe it first in case
> anyone wants to lobby for it to be fixed.
> 
> First, the regression only affects one Ada test case, dgopt.exp, and
> only when using fission-dwp.  In this scenario, the executable has a
> stub CU that looks like:
> 
>      <0><3f>: Abbrev Number: 1 (DW_TAG_compile_unit)
>     [...]
>         <50>   DW_AT_stmt_list   : 0x24d
>     [...]
>         <58>   DW_AT_comp_dir    : (indirect string, offset: 0x0): /home/tromey/gdb/build/gdb/testsuite/outputs/gdb.ada/dgopt
> 
> Note there is no DW_AT_name.
> 
> The corresponding .dwo file says:
> 
>      <0><b>: Abbrev Number: 1 (DW_TAG_compile_unit)
>     [...]
>         <e>   DW_AT_name        : (indexed string: 0x1): /home/tromey/gdb/binutils-gdb/gdb/testsuite/gdb.ada/dgopt/x.adb
>         <f>   DW_AT_comp_dir    : (indexed string: 0x2): /home/tromey/gdb/build/gdb/testsuite/outputs/gdb.ada/dgopt
> 
> That is, this has the correct file name in DW_AT_name.  This is why the
> psymtab reader works -- the psymtab reader uses the line table, but also
> uses this name for file name matches.
> 
> The line table does not mention x.adb, but instead:
> 
>      The File Name Table (offset 0x26a):
>       Entry Dir     Time    Size    Name
>       1     0       0       0       x.adb.dg
> 
> That is, the file name table here holds the name of a file that's
> generated by the Ada compiler when run in a special mode.
> 
> So, my inclination is to say that this is probably a compiler bug, or in
> any case an obscure oddity that gdb shouldn't bother with.  To my
> knowledge, there aren't Ada users who also use Fission anyway.
> 

I've looked at the readelf -w output for native, and I see the same,
that is, a CU with:
...
    <14e2>   DW_AT_name        :
/home/vries/gdb_versions/devel/binutils-gdb.git/gdb/testsuite/gdb.ada/dgopt/x.adb
...
and a .debug_line entry with:
...
 The File Name Table (offset 0x361):
  Entry Dir     Time    Size    Name
  1     0       0       0       x.adb.dg
...

At the gnat docs we find:
...
-gnatD[=nn]
  ...
  The debugging information generated by the gcc -g switch will refer to
the generated xxx.dg file.
...

So, I'm inclined to say that it's a compiler bug that we have
"DW_AT_name x.adb" rather than "DW_AT_name x.adb.dg".  I'd file a gcc PR
for this.

Only, it's a compiler bug that the compiler doesn't stick to its own
spec, but AFAIU this is valid DWARF.

I tried a two-file test-case:
...
$ cat test.c
#include "test2.c"
$ cat test2.c
int
main (void)
{
  return 0;
}
$ gcc -g test.c
...
and got the same structure.

FWIW, it could be interesting to create a similar test-case with these
sources and see if the same problem pops up for your branch with
fission-dwp.

Thanks,
- Tom

> I do think this could be fixed.  It would require maybe storing a
> file_and_directory in the dwarf2_per_cu_data.  So if you think this is
> important, I'd probably refactor to take this approach.
> 
> TBH I had expected this scenario to make a .gdb_index -- that's what
> fission-dwp is supposed to do, I think, via gold -- but this didn't
> happen.  So possibly this is another oddity in the Ada toolchain, or
> maybe the gdb test suite.  I'm not certain.
> 
> Tom
>