Re: [PATCH] Handle weak symbols

[Mark Mitchell <mark at codesourcery dot com>]

OK, now I understand the problem, and I agree that there's a bug
here.  That's one step. :-)  By the way, you left out the rtlanal.c
patch in your original posting, but I can imagine it. :-)

I don't fully understand what you're proposing from this:

  - is_on_pragma_weak_list() walking the (usually short) #pragma weak list
  - add_pragma_weak() called from the #pragma handling
  - is_on_assembled_weak_list() called from the assemble_* functions and 
  weak_finish()
  - add_weak() called from the assemble_* functions

but, it's probably sensible enough, or we can come up with something
else that works.  I can certainly see how we can do this efficiently.

Thanks for the test-case.  Unfortunately, on x86-linux I seem to get
code with a test in it when I compile with -O2:

  atexit:
	  pushl	%ebp
	  movl	$__dso_handle, %eax
	  movl	%esp, %ebp
	  subl	$12, %esp
	  xorl	%edx, %edx
	  testl	%eax, %eax
	  je	.L4
	  movl	__dso_handle, %edx
  .L4:
	  pushl	%edx
	  movl	8(%ebp), %eax
	  pushl	$0
	  pushl	%eax
	  call	__cxa_atexit
	  movl	%ebp, %esp
	  popl	%ebp
	  ret

I'd really like a test-case that actually manifests the bug, even
though I can understand why it would exist.  Did I do something 
wrong?

In any case, you say:

  >   - Or, are development versions of glibc using some new construct
  >     that old ones didn't?  If so, why?

  It was introduced recently.

Is there another way that glibc could do its job without using this
construct?  That would allow us not to put these changes into GCC 3.0.
At this point, we're trying to avoid everything we can.  If this is
only a development version -- as opposed to released version -- of
glibc, I'd rather see glibc work around the bug.

And then I think we should fix the bug for 3.1.

Thanks again,

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com