Hello, FPC developers' list.

Is it known that gdb returns "void" type for the Variant variables when STABS debug info is used?

When DWARF is used Variant = TVarData which is also not perfect. I want to see only variant value, not the whole record when I debug the program in Lazarus. When I need the whole record I can always evaluate TVarData(V).

Is it possible to improve both stabs and dwarf info?

Paul Ishenin wrote

Is it known that gdb returns "void" type for the Variant variables when STABS debug info is used?

Not that I know. Stabs is pretty much legacy though.

In Stabs it is impossible to only show the actual variant value, as Stabs does not know a "variant" concept. DWARF does, but GDB doesn't support it yet.

Jonas

Jonas Maebe wrote

In Stabs it is impossible to only show the actual variant value, as Stabs does not know a "variant" concept.

And lazarus does not need that. I think it would be enough to mark the variant variables somehow (not "void").

DWARF does, but GDB doesn't support it yet.

We only need to disctinct Variant and TVarData. So when user want to watch Variable variable we will perform TVarData() cast and show the value ourself. When TVarData is used we will show as TVarData (iow, as record).

Best regards,
Paul Ishenin.

Would it be enough if the type name were changed to "variant"? It would not be 100% safe (since "variant" is not a reserved word, anyone can declare a variable/type/... with the name "variant"), but I don't immediately see another solution.

Jonas

Jonas Maebe wrote

Would it be enough if the type name were changed to "variant"? It would not be 100% safe (since "variant" is not a reserved word, anyone can declare a variable/type/... with the name "variant"), but I don't immediately see another solution.

I think yes.

Best regards,
Paul Ishenin.

Does the attached patch work for DWARF with Lazarus? "ptype variantvar" still shows tvardata, but "info variantvar" shows "static VARIANT V;", and I don't know what the GDB MI will return for the query that Lazarus uses.

Jonas

Index: dbgdwarf.pas
=================================================================== --- dbgdwarf.pas (revision 14293)
+++ dbgdwarf.pas (working copy)
@@ -3245,8 +3245,11 @@

      procedure TDebugInfoDwarf2.appenddef_variant(list:TAsmList;def:
tvariantdef);
        begin
-        { variants aren't known to dwarf2 but writting tvardata
should be enough }
-        appenddef_record(list,trecorddef(vardatadef));
+        append_entry(DW_TAG_typedef,false,[
+          DW_AT_name,DW_FORM_string,'Variant'#0
+        ]);
+        append_labelentry_ref(DW_AT_type,def_dwarf_lab(vardatadef));
+        finish_entry;
        end;
      function TDebugInfoDwarf2.dwarf_version: Word;

the mi version of info var.. is not implemented in gdb, so lazarus uses ptype var

So I guess for this case, when a TVarrec is returned we'll inspect further using info.

Thanks Marc

It's also possible to create another definition of tvarrec with "variant" as name. It's just slightly less clean to implement. Also note that "info var" takes a regular expression, so make sure to use "info var ^varname$", otherwise you'll get info about a lot of unrelated matches.

Jonas

In that case I think I prefer a full dwarf variant :)

It's just slightly less clean to implement. Also note that "info var" takes a regular expression, so make sure to use "info var ^varname$", otherwise you'll get info about a lot of unrelated matches.

:)

Thanks.

Marc

Ok, it's now called "Variant". As an extra advantage, I think that this cannot clash with user-defined types, because afaik those type names are always uppercased.

Jonas

Jonas Maebe wrote

Ok, it's now called "Variant". As an extra advantage, I think that this cannot clash with user-defined types, because afaik those type names are always uppercased.

Thanks. This is what I asked. Can it be similar for STABS?

Best regards,
Paul Ishenin.

Jonas, using your change I implemented variant variables showing in Lazarus. It works fine on my big test. For stabs it still shows 'void'.

Best regards,
Paul Ishenin.

Paul Ishenin wrote

Jonas, using your change I implemented variant variables showing in Lazarus. It works fine on my big test. For stabs it still shows 'void'.

That's because I did not change anything to Stabs. Stabs is old and legacy (not deprecated, but not actively developed anymore either). If someone else wants to fix it in Stabs, be my guest.

Jonas

Jonas Maebe wrote

That's because I did not change anything to Stabs. Stabs is old and legacy (not deprecated, but not actively developed anymore either). If someone else wants to fix it in Stabs, be my guest.

Can I be your guest?

The patch do the job. Althought it writes VARIANT to the debug info. But in Lazarus there is no problem to check for both 'Variant' and 'VARIANT'. If you apply it (or allow me to apply it) then I will fix lazarus too.

Best regards,
Paul Ishenin.

Paul Ishenin wrote

Jonas Maebe wrote: > That's because I did not change anything to Stabs. Stabs is old and legacy (not deprecated, but not actively developed anymore either). If someone else wants to fix it in Stabs, be my guest. Can I be your guest?

Sure!

The patch do the job. Althought it writes VARIANT to the debug info. But in Lazarus there is no problem to check for both 'Variant' and 'VARIANT'. If you apply it (or allow me to apply it) then I will fix lazarus too.

It looks fine (but please change the two empty lines after vardatadef:=trecorddef(..) to one), feel free to commit it.

Thanks,

Jonas