Re: RFC: Add dynamic list to version script

[Michael Matz <matz@suse.de>]

Hi H.J.,

On Mon, 24 Jul 2006, H. J. Lu wrote:

> Here is a prototype, which added dynamic list to version script. I used
> 
> {
> dynamic:
>   extern "C++"
>   {
>     *typeinfo*;
>   };
> };
> 
> It seems to work correctly with execptions between shared C++ libraries.

Yep.  Though it slightly changes the output even without using a version 
script but using -Bsymbolic.  For instance a very simple library defining 
just one class 'A' with a ctor has this set of interesting relocs without 
the patch, without any flag:

000017a0  00001201 R_386_32          00000730   _ZTS1A
00001884  00001506 R_386_GLOB_DAT    0000179c   _ZTI1A
0000189c  00001407 R_386_JUMP_SLOT   000006ca   _ZN1AC1Ei

With -Bsymbolic (still not applied the patch) this changes as expected to:

000018a0  00000008 R_386_RELATIVE
000018a4  00000008 R_386_RELATIVE

(and no JUMP_SLOT entry for the ctor), i.e. all three references are bound 
locally.  When the patch is applied, no version script is used, but 
-Bsymbolic I have this set of relocs:

000018a4  00000008 R_386_RELATIVE
00001790  00001201 R_386_32          00000720   _ZTS1A

So, no reference to the typeinfo node or the ctor anymore, locally bound.  
But still a reference to the typeinfo name symbol.  As said, with 
-Bsymbolic and no version script.  I think some of the (info->dynamic && 
!h->dynamic) conditions are wrong.  With a version script I get this set 
of relocs:

00001790  00001201 R_386_32          00000720   _ZTS1A
00001874  00001506 R_386_GLOB_DAT    0000178c   _ZTI1A

(no JUMP_SLOT, as expected that's now bound locally).  With -Bsymbolic the 
_ZTI1A reloc goes away, but the _ZTS1A remains.

I find the overloading of semantics a bit confusing.  I mean that using a 
version script with the "dynamic" tag implicitely has the effect of 
-Bsymbolic without actually mentioning that flag.  This means that an 
empty version script using the "dynamic" tag is the opposite of using no 
version script.  Perhaps such version script should only have an effect if 
the user also specifies -Bsymbolic (i.e. creating exceptions to the 
-Bsymbolic symbol set).  OTOH it's documented so it's not that important.

Oh, and yes, throwing exceptions does work, while without the version 
script it does not :-)  Nice.


Ciao,
Michael.