[Bug ld/4538] New: static initialization ignored in static archive (.a)

[ejt at andrew dot cmu dot edu]

This is my $0.02 that I consider the ignoring of static initializers with side 
effects to be a bug, as per 
discussions at:
http://www.ecos.sourceware.org/ml/binutils/2003-04/msg00262.html
http://sources.redhat.com/ml/binutils/2003-02/msg00131.html

This recently cost me quite a bit of time to figure out why my code worked 
under OS X (shared libraries) 
but not linux (happened to be using static libraries), and it was eventually 
tracked down that since 
nothing explicitly referenced the translation unit, it was never linked in from 
the archive, and thus its 
static initialization was ignored.

Users writing auto-registration patterns will expect their code to register, 
and are going to be surprised 
when it fails to do so, particularly if it varies based on linking style.  
There is a nice list of workarounds 
here:
http://sources.redhat.com/ml/binutils/2003-02/msg00158.html
but each has shortcomings, mostly requiring end users of a library to have 
internal knowledge as to 
which symbols or files need to be explicitly imported, or requiring 
--whole-archive, which invalidates 
the point of using an archive.

The only argument in the threads referenced above *against* importing all 
static initialization sections 
was that it would load too many symbols from system libraries.  Hal Black 
points out this is not the 
case:
http://www.ecos.sourceware.org/ml/binutils/2003-04/msg00289.html
Although there are further hand-wavy counter-arguments regarding .ctor/.init 
sections, I'm not sure 
those would need to be imported to give the functionality we're looking for.

I would like to see some numbers there to prove that it is performance concern, 
since that appears to 
be the *only* reason to defend the current behavior, and is a demonstrated 
source of confusion and 
error.  If performance is indeed a valid case for leaving this issue open, 
there should at least be a better 
workaround provided...  for example:

OPTION A) Some kind of __attribute__ flag (i.e. 'used' or 'constructor') which 
marks that specific symbols 
be imported 'greedily' regardless of whether they are referenced.  This would 
keep knowledge about 
auto-registration in the code that's being registered, instead of requiring 
library end users to maintain 
a list of auto-registered classes in those libraries, which defeats the entire 
purpose of the pattern.

OPTION B) --all_init flag, similar to --whole_archive, but only the 
initalization sections (and their 
dependencies of course).  Some of us want our standards compliance in the 
spirit of the law, without 
having to resort to importing *everything*.

        From the C++ draft spec 2 December 1996:

        Section 3.7.1:
        2 If an object of static storage duration has initialization or a
        destructor with side effects, it shall not be eliminated even if it
        appears to be unused, except that a class object or its copy may be
        eliminated as specified in 12.8. [hwb: see ** below]

-- 
           Summary: static initialization ignored in static archive (.a)
           Product: binutils
           Version: unspecified
            Status: NEW
          Severity: normal
          Priority: P2
         Component: ld
        AssignedTo: unassigned at sources dot redhat dot com
        ReportedBy: ejt at andrew dot cmu dot edu
                CC: bug-binutils at gnu dot org


http://sourceware.org/bugzilla/show_bug.cgi?id=4538

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.