Re: PATCH: PR ld/12730: regression] crash when allocating in a static constructor

[Alan Modra <amodra at gmail dot com>]

On Thu, May 05, 2011 at 06:26:15AM -0700, H.J. Lu wrote:
> On Thu, May 5, 2011 at 1:27 AM, Alan Modra <amodra@gmail.com> wrote:
> > On Wed, May 04, 2011 at 10:18:32PM -0700, H.J. Lu wrote:
> >> When we put .ctors into .init_array, we have to reverse copy .ctors secton.
> >> Otherwise, constructor function may not work with C++ run-time library
> >> correctly. ?OK for trunk?
> >
> > What about .dtors? ?You have the same problem there. ?I suspect, but
> 
> You are right.  Here is the updated patch to handle .dtors sections
> with the updated testcase.  OK for trunk?

No.

> > haven't verified, that .ctors.* and .dtors.* also need reversing. ?If
> > that is true then it would be better to do your reversing trick for
> > anything going to the .init_array output section that isn't named
> > .init_array* and similarly for .fini_array.

Are you sure there is no need to reverse .ctors.* and .dtors.*?

The reason I recommended testing the output section is that limits
section reversal to that particular output section.  I will not
approve a patch that ignores this recommendation.  You also should
remove reverse_copy_ctors.  That's plain wrong.  Consider people
linking using a custom (old) script that does not put .ctors into
.init_array.

> > You also need to reverse any dynamic relocations applying to the
> > sections you are reversing.
> 
> It isn't a problem since we apply relocations on the input sections
> first and copy relocated input sections to output where I reverse
> copy .ctors/.dtors sections if needed.

Try compiling your testcase as a PIE or shared lib on a target that
uses RELA.  I haven't tried it, but I think the dynamic RELATIVE
relocs you'll get in .ctors will undo your section reversing.

-- 
Alan Modra
Australia Development Lab, IBM