[Bug ada/65524] New: gnatbind generates decrementing the unexisting elab-counter into finalize_library

[Bug ada/65524] New: gnatbind generates decrementing the unexisting elab-counter into finalize_library

[demoonlit at panathenaia dot halfmoon.jp]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65524

            Bug ID: 65524
           Summary: gnatbind generates decrementing the unexisting
                    elab-counter into finalize_library
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: ada
          Assignee: unassigned at gcc dot gnu.org
          Reporter: demoonlit at panathenaia dot halfmoon.jp

I found some cases that gnatbind generates decrementing the unexisting
elaboration counter of each package Exxx.

   E142 : Short_Integer; pragma Import (Ada, E142, "system__exn_lli_E");
   ... some Exxx are declared ...

   procedure finalize_library is
   begin
      ...
      E132 := E132 - 1; -- * E132 is not declared in above *
      ...
   end finalize_library;

A cause is a package having _finalize_spec/_finalize_body without
_elabs/_elabb.
I still have not been able to make the minimal example, but have made a patch.

In bindgen.adb, Gen_Elab_Externals refers U.Set_Elab_Entity to generate Exxx.
However, Gen_Finalize_Library does not refer it.

--- a/gcc/ada/bindgen.adb
+++ b/gcc/ada/bindgen.adb
@@ -1434,7 +1434,9 @@ package body Bindgen is
             --  has a finalizer. In that case, this is where we decrement
             --  the elaboration entity.

-            if U.Utype = Is_Body and then Uspec.Has_Finalizer then
+            if U.Utype = Is_Body and then Uspec.Has_Finalizer
+               and then Uspec.Set_Elab_Entity
+            then
                if not Lib_Final_Built then
                   Gen_Header;
                   Lib_Final_Built := True;
@@ -1548,7 +1550,9 @@ package body Bindgen is

             WBI ("      begin");

-            if U.Utype /= Is_Spec then
+            if U.Utype /= Is_Spec
+               and then Uspec.Set_Elab_Entity
+            then
                Set_String ("         E");
                Set_Unit_Number (Unum);
                Set_String (" := E");

[Bug ada/65524] gnatbind generates decrementing the unexisting elab-counter into finalize_library

[charlet at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65524

Arnaud Charlet <charlet at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2015-03-23
                 CC|                            |charlet at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Arnaud Charlet <charlet at gcc dot gnu.org> ---
Sounds interesting/puzzling. But we really need a reproducer/test case to work
from.

[Bug ada/65524] gnatbind generates decrementing the unexisting elab-counter into finalize_library

[demoonlit at panathenaia dot halfmoon.jp]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65524

--- Comment #2 from yuta tomino <demoonlit at panathenaia dot halfmoon.jp> ---
Created attachment 35115
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35115&action=edit
example

I found the way of reproducing.

A tiny change of a-tags.ads is necessary.
Insert "is null" into Ada.Tags.Register_Tag to suppress the elaboration.
(Is there a pragma or restriction for this purpose?)

Note, Unregister_Tag is unrelated.

And, make a package as Pure or Preelaborate. Add a tagged type.
So the package would have _finalize_spec without _elabs.

(The direct cause of this case is the asymmetry that _elabs is not generated if
Register_Tag is null, against that,
_fialize_spec is always generated even if Unregister_Tag is null.
Register/Unregister_XXX are sometimes suppressed in parts of customized
runtimes.
My customized runtime satisfies the conditions about this time.
However, the way is only example for reproducing.
I think another approach is possibility because Build_Finalizer in exp_ch7.adb
is very complex...)

And then, prepare a main subprogram using the package. Compile and gnatbind.
See b~main.adb.

The log of compiling the attached example:

 % gnatmake -a -gnatp -g main
 gnatbind -x main.ali
 gnatlink main.ali -g
 b~main.adb:31:10: "E81" is undefined (more references follow)
 b~main.adb:34:07: "E05" is undefined (more references follow)
 gnatmake: *** link failed.

By the way, sorry, I realized that my first patch is not proper fix.
For multiple-elaboration, an elaboration-counter should be generated and
incremented in adainit when a _finalize_spec/body is existing, even if _elabs
is not existing.

Thanks.

[Bug ada/65524] gnatbind generates decrementing the unexisting elab-counter into finalize_library

[charlet at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65524

Arnaud Charlet <charlet at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
         Resolution|---                         |INVALID

--- Comment #3 from Arnaud Charlet <charlet at gcc dot gnu.org> ---
Using a custom runtime change isn't supported by GCC FSF.
I suspect the fix if any would be to geenrate the _elabs routine rather than
trying to patch the binder, but as you are using a customized runtime, there
might
be other issues at play and the simplest is for you to remove the use of a 'is
null' subprogram.