[[!meta copyright="Copyright © 2011 Free Software Foundation, Inc."]]

[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
id="license" text="Permission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant
Sections, no Front-Cover Texts, and no Back-Cover Texts.  A copy of the license
is included in the section entitled [[GNU Free Documentation
License|/fdl]]."]]"""]]

[[!tag open_issue_glibc]]

IRC, freenode, #hurd, 2011-07-22

    [additional init-first.c patch]
    < tschwinge> civodul: The only thing I wonder about: Roland also once had
      done similar changes, which I then found they didn'T work with GCC 4.1,
      and backed them out in 08f53ee9d265ffdc7e0affd6acf346cceeb65559 and fixed
      the issue differently in d8d27e633a7860b37fd2e3142822b640a066cc0f (and
      e17cef66140d4c97710ea88bd8d12612799e1e0f).  Have you reviewed this?
    < tschwinge> That's in the Savannah glibc repository.
    < tschwinge> And this has been in 2007, four years ago.  I don't remember
      all the details.
    < tschwinge> And here is quite a good summary of this stuff, from
      init-first.c:
    < tschwinge> /* XXX This is all a crock and I am not happy with it.
    < tschwinge>    This poorly-named function is called by static-start.S,
    < civodul> braunr: thanks; i must admit it took me a while to figure it out
      ;-)
    < tschwinge>    which should not exist at all.  */
    < tschwinge> civodul: I can imagine...  :-/
    < civodul> tschwinge: re Roland's changes, that's weird; i plan to try to
      reinstate his change and see if it works
    < civodul> now, i won't test with GCC 4.1...
    < tschwinge> Yeah...
    < tschwinge> I'm happy if it works with 4.4 onwards.
    < tschwinge> civodul: And it's safe (in GCC terms) to write to ``* ((void
      **) __builtin_frame_address (0) + 1)'', and similar?
    < tschwinge> Or should we be coding this few stuff in assembly?
    < civodul> tschwinge: well, we should add a compile-time assertion for
      __builtin_return_address (0) == *((void**)__builtin_frame_address (0) +
      1)
    < civodul> (i think GCC can figure it out at compile-time)
    < civodul> but on IA32 it should always be true
    < civodul> what's the name of glibc's compile-time assert macro already?
    < tschwinge> I wonder whether that might interfere with some of GCC's
      optimizations?
    < civodul> what?
    < tschwinge> Well, it seems unclean for me to be modifying a function's
      return address from within C code.
    < tschwinge> civodul: I added a verify.h in the t/verify.h branch.  But
      people didn't really like it too much.  They rather wanted to directly
      inline the array[(cond)?1:-1] code.
    < civodul> ok
    < civodul> i remember a debate about Gnulib's verify.h
    < civodul> i thought something comparable had landed eventually
    < tschwinge> civodul: Oh, maybe I missed it.
    < tschwinge> civodul: In init-first.c:init, what about the usage of
      data[-1] in the else path (not using cthreads) -- is that good as-is?
    < civodul> tschwinge: oooh, it probably needs to fixed too
    < civodul> but i haven't reached that point yet ;-)
     * civodul tries to cross-bootstrap GNU from scratch
    < tschwinge> civodul: I'd be happy to learn what was wrong with Roland's
      original idea of fixing this.  Or perhaps this was a GCC 4.1 bug?  Or
      perhaps GCC was inlining to much, and then got confused with frames and
      return addresses?
    < civodul> tschwinge: Roland's change looks good to me, so it could have
      been a GCC bug
    < civodul> tschwinge: OK to commit the patch to t/init-first.c (with both
      data[-1] replaced)?
    < tschwinge> civodul: OK, if you are confident that it works with GCC 4.4
      onwards.  If yes, please add your changelog snippet to .topmsg, and also
      add a not that Roland's original code may in fact have been fine, and we
      may have hit a compiler bug.
    < civodul> tschwinge: OK, will do
    < civodul> tschwinge: though regarding Roland's change, i'd prefer to
      actually test and see
    < tschwinge> civodul: Thanks!