[Bug target/33704] AIX runs c++ constructors in incorrect order

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

--- Comment #23 from David Edelsohn <dje at gcc dot gnu.org> ---
Author: dje
Date: Sat Nov 23 15:38:07 2013
New Revision: 205309

URL: http://gcc.gnu.org/viewcvs?rev=205309&root=gcc&view=rev
Log:
libgcc:

    PR target/33704
        * config/rs6000/aixinitfini.c: New file.
        * config/rs6000/t-aix-cxa (LIB2ADD_ST): Add aixinitfini.c.
        * config/rs6000/libgcc-aix-cxa.ver (GCC_4.9): Add libgcc initfini
        symbols.

gcc:

    PR target/33704
        * config/rs6000/aix.h (COLLECT_SHARED_INIT_FUNC): Define.
        (COLLECT_SHARED_FINI_FUNC): Define.

        * collect2.c (aix_shared_initname): Declare.
        (aix_shared_fininame): Declare.
        (symkind): Add SYM_AIXI and SYM_AIXD.
        (scanfilter_masks): Add SCAN_AIXI and SCAN_AIXD.
        (struct names special): Add GLOBAL__AIXI_ and GLOBAL__AIXD_.
        (aixlazy_flag): Parse.
        (extract_init_priority): SYM_AIXI and SYM_AIXD have highest priority.
        (scan_prog_file, COFF): Handle SYM_AIXI and SYM_AIXD.

Added:
    trunk/libgcc/config/rs6000/aixinitfini.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/collect2.c
    trunk/gcc/config/rs6000/aix.h
    trunk/libgcc/ChangeLog
    trunk/libgcc/config/rs6000/libgcc-aix-cxa.ver
    trunk/libgcc/config/rs6000/t-aix-cxa

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #24 from David Edelsohn <dje at gcc dot gnu.org> ---
Fixed.

[Bug target/33704] AIX runs c++ constructors in incorrect order

[bkoz at gcc dot gnu dot org]

------- Comment #22 from bkoz at gcc dot gnu dot org  2007-10-19 02:00 -------

> I believe there are other functions in libstdc++ that depend on initialization
> order. I do not have specifics.

Perhaps:
FAIL: 27_io/objects/char/6.cc execution test
FAIL: 27_io/objects/wchar_t/6.cc execution test
FAIL: ext/mt_allocator/check_new.cc execution test
FAIL: ext/pool_allocator/check_new.cc execution test

from:
http://gcc.gnu.org/ml/gcc-testresults/2007-10/msg00820.html

These are the ones I would suspect might change.

FWIW, I agree 100% with Wolfgang's venting in #8, #10. Correct __cxa_atexit
behavior is essential.

-benjamin


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #21 from ajd at gentrack dot com  2007-10-19 01:44 -------
Created an attachment (id=14372)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14372&action=view)
support -blazy, add option to disable

I suggest exportinitfini_flag is enabled by default or atleast for the build of
libstdc++.a and libgcc_s.a.
It should have minimal impact/risk.

initdependents_flag has risk/impact attached. You may want to reverse the
default.

My case for having it on by default:
To use exceptions within global constructors a module needs to be linked with
initdependents_flag enabled.

I believe there are other functions in libstdc++ that depend on initialization
order. I do not have specifics.

Assuming AIX was the only OS that works this way, it allows removal of these
types of workarounds:
http://gcc.gnu.org/ml/libstdc++/2001-01/msg00369.html


-- 

ajd at gentrack dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #14327|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu dot org]

------- Comment #20 from dje at gcc dot gnu dot org  2007-10-18 00:48 -------
Yes, effectively -blazy, because of AIX's loader semantics.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #19 from ajd at gentrack dot com  2007-10-14 21:39 -------
(In reply to comment #18)
> Do you plan to support lazy binding as well?

Do you mean -blazy?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu dot org]

------- Comment #18 from dje at gcc dot gnu dot org  2007-10-12 12:56 -------
I think your patch would be okay as an option.  Do you plan to support lazy
binding as well?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #17 from ajd at gentrack dot com  2007-10-10 04:04 -------
(In reply to comment #16)
> lib1.a depends on lib2.a
> main depends on lib1.a, but does not explicitly link with lib2.a

main initializes what it is explicitly linked against.
lib1 initializes lib2 before it initializes itself.

--
See attached test_init_order.cpp:

$ # -- create lib2
$ g++ -shared -o lib2.a -D LIBNAME=LIB2 test_init_order.cpp
$ # -- create lib1 depends on lib2
$ g++ -shared -L. -l2 -o lib1.a -D LIBNAME=LIB1 -DCALLEE=LIB2
test_init_order.cpp
$ # -- hide lib2 from main
$ mv lib2.a /tmp/lib2.hide
$ # -- link main against lib1
$ g++ -L. -l1 -o m -D LIBNAME=main -DCALLEE=LIB1 test_init_order.cpp
$ # -- unhide lib2
$ mv /tmp/lib2.hide lib2.a
$ ./m
LIB2_init
LIB1_init
main_init
main()
LIB1()
LIB2()
main_fini
LIB1_fini
LIB2_fini
$ # -- replace lib1 with a library that doesn't depend on lib2
$ # -- leave main alone
$ g++ -shared -o lib1.a -D LIBNAME=LIB1 test_init_order.cpp
$ /usr/local/bin/sudo slibclean
$ ./m
LIB1_init
main_init
main()
LIB1()
main_fini
LIB1_fini


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu dot org]

------- Comment #16 from dje at gcc dot gnu dot org  2007-10-10 03:17 -------
lib1.a depends on lib2.a
main depends on lib1.a, but does not explicitly link with lib2.a

Your proposed solution requires that the link of main see all dependent
libraries.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #15 from ajd at gentrack dot com  2007-10-09 21:44 -------
(In reply to comment #14)
> You seem to want to have the main program run the share library constructors. 
> That's a no-no.  And decide at link-time.  Double no-no.

Do you mind explaining a situation that won't work because of this?


The the following work should work as -binitfini is still used:
- LDR_PRELOAD 
- libraries loaded by dlopen at runtime
- executables linked by xlc loading libraries compiled by gcc

If main was directly linked against lib1.a:
- lib1.a can add/remove dependencies against other libraries.
- lib1.a can change it's constructors.
- lib1.a can be resolved to a completely different library as long as the
replacement contains __GLOBAL__FI_lib1.

The two situations I can identify that won't work are if the replacement lib1.a
was renamed or was compiled by xlc. I can look at resolving these if they're
the only problems.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu dot org]

------- Comment #14 from dje at gcc dot gnu dot org  2007-10-09 21:10 -------
You seem to want to have the main program run the share library constructors. 
That's a no-no.  And decide at link-time.  Double no-no.  You essentially want
to behave like a static link.

If you look at the discussion from years ago, to solve this correctly one needs
to implement a solution at runtime, not link-edit time.  At runtime, the
constructors need to register themselves instead of running, then the init
processing in the main program performs a global sort, and then actually runs
the constructors depth-first.

Please open the throw failure as a separate PR if it is a separate issue.  I
cannot tell if you think it is a byproduct of AIX behavior or a separate issue.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #13 from ajd at gentrack dot com  2007-10-09 20:06 -------
(In reply to comment #9)
> A solution within GCC is very complicated.  

Have you looked at the suggested patch?
If you could point out any concerns/faults, I could have another try at it.

XLC has it's own solution implemented within the compiler / support libraries.
So it seems the loader behaviour only impacts gcc. It could be seen as a
deficiency in the gcc compiler / support libraries.

(In reply to comment #11)
> The problem *could* be solved in GCC, so it probably is worth retaining the PR.

Should I raise a separate PR for throw_failure.cpp?
I lumped them together because I figured it was the same root cause.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[bangerth at dealii dot org]

------- Comment #12 from bangerth at dealii dot org  2007-10-09 15:18 -------
(In reply to comment #11)
> The problem *could* be solved in GCC, so it probably is worth retaining the PR.

OK.


> AIX Brand is much more likely to respond to a customer requirement than a
> request from within IBM.  When the problem affects a sale, IBM will address it.

Fair enough. Sorry to (ab)use you in this regard.

W.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu dot org]

------- Comment #11 from dje at gcc dot gnu dot org  2007-10-09 15:07 -------
The problem *could* be solved in GCC, so it probably is worth retaining the PR.

AIX Brand is much more likely to respond to a customer requirement than a
request from within IBM.  When the problem affects a sale, IBM will address it.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[bangerth at dealii dot org]

------- Comment #10 from bangerth at dealii dot org  2007-10-09 14:53 -------
If it's a request to improve the AIX linker, should we even keep this PR
open?

David, as a sidenote (because you're at IBM), this is the sort of thing that
makes AIX such an unpopular target to do development on. It just surprises
me that 10+ years after everyone else had their linker fixed to support proper
C++ templates and initialization order, IBM's linker still can't do it. I
found porting our 350 kloc template-heavy code to AIX one of the more painful
exercises I ever did. That's not your fault, but maybe you can pass it on to
someone in charge. Sigh, and sorry for the digression...

W.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[dje at gcc dot gnu dot org]

------- Comment #9 from dje at gcc dot gnu dot org  2007-10-09 12:50 -------
This is a long-standing issue about the way AIX loader behaves.  AIX loader
implements breadth-first, not depth first.  This is a difference from
SVR4-based systems.  I believe that SVR4 does specify the ordeer.  Look about
10 years ago on the mailinglist for a similar discussion.

A solution within GCC is very complicated.  This really is an enhancement
request for AIX itself to make it more compatible with Linux and other
SVR4-like operating systems.


-- 

dje at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2007-10-09 12:50:46
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[bangerth at dealii dot org]

------- Comment #8 from bangerth at dealii dot org  2007-10-09 03:43 -------
(In reply to comment #6)
> And the order is still undefined between them, does not matter if it is a
> shared library or otherwise.

It is true that the standard only talks about the order of initialization
within each translation unit. That doesn't make the current PR any less
of a valid enhancement request. It's not a coincidence that all the other
systems allow this to work -- this is a very useful feature and it would
be nice if we could get it to work on AIX as well.

W.


-- 

bangerth at dealii dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dje at watson dot ibm dot
                   |                            |com, bangerth at dealii dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #7 from ajd at gentrack dot com  2007-10-09 02:43 -------
Introduced here:
http://gcc.gnu.org/ml/gcc-patches/2004-03/msg00616.html

My suggested fix is co-incidentally reversing some of that patch.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[pinskia at gcc dot gnu dot org]

------- Comment #6 from pinskia at gcc dot gnu dot org  2007-10-09 01:59 -------
And the order is still undefined between them, does not matter if it is a
shared library or otherwise.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #5 from ajd at gentrack dot com  2007-10-09 01:55 -------
Please reconsider.

This is about initialization order between shared libraries. Not betweeen
objects within a library.

This works on atleast Linux, Windows, Solaris, HPUX, Tru64.

The testcase throw_failure.cpp shows that libgcc/libstdc++ depend on this
initialization order.

So if the initialization order is really unspecified, then libgcc/libstdc++
have a bug.


-- 

ajd at gentrack dot com changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #4 from ajd at gentrack dot com  2007-10-09 01:50 -------
Created an attachment (id=14327)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14327&action=view)
suggested solution

The attached patch is a suggested solution to this problem.
It calls the __GLOBAL__FI_... functions (which call the C++ constructors) of
linked libraries before calling the constructors of the current module.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #3 from ajd at gentrack dot com  2007-10-09 01:49 -------
Created an attachment (id=14326)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14326&action=view)
shows fallout from this (throw does not work in a global constructor as
libstdc++ and libgcc_s constructors have not been run).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[ajd at gentrack dot com]

------- Comment #2 from ajd at gentrack dot com  2007-10-09 01:48 -------
Created an attachment (id=14325)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14325&action=view)
testcase shows constructors running in incorrect order

Compile like:
g++ -shared -o liba1.a -D LIBNAME=LIB1 test_init_order.cpp
g++ -shared -L. -la1 -o liba2.a -D LIBNAME=LIB2 -DCALLEE=LIB1
test_init_order.cpp
g++ -shared -L. -la2 -o liba3.a -D LIBNAME=LIB3 -DCALLEE=LIB2
test_init_order.cpp
g++ -L. -la3 -o m -D LIBNAME=main -DCALLEE=LIB3 test_init_order.cpp

Creates libraries dependencies like:
m -> LIB3 -> LIB2 -> LIB1

Expected output is:
LIB1_init
LIB2_init
LIB3_init
main_init
main()
LIB3()
LIB2()
LIB1()
main_fini
LIB3_fini
LIB2_fini
LIB1_fini

Actual output is:
main_init
LIB3_init
LIB2_init
LIB1_init
main()
LIB3()
LIB2()
LIB1()
LIB1_fini
LIB2_fini
LIB3_fini
main_fini


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704

[Bug target/33704] AIX runs c++ constructors in incorrect order

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2007-10-09 01:48 -------
It is not specificed which order of constructors are run between two different
Translation units.


-- 

pinskia at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33704