GCC Bugzilla Bug 26670

[forwarded from http://bugs.debian.org/356548]

Under some circumstances, non-PODs are not properly stripped of their
packedness:

struct nonpod {
    nonpod();
};

struct nonpod_pack {
    nonpod n;
} __attribute__ ((packed));

struct nonpod_pack2 {
    nonpod_pack p;
} __attribute__ ((packed));

gives

foo.cc:6: warning: ignoring packed attribute on unpacked non-POD field 'nonpod
nonpod_pack::n'

nonpod_pack2::p should also have the packedness ignored, since it is a non-POD.

Hmm, it does not warn about non-PODs which contain non-PODS which had been
warned about already.

It's not just a missing warning, it also changes behavior:

int& foo(nonpod_pack& n) {
    return n.n.i;               // this passes
}

int& foo(nonpod_pack2& p) {
    return p.p.n.i;             // this barfs
}

Hmm, related to PR 13983.

And to PR 17519.

This is intended behavior: a non-POD type which has been marked as packed can
be used as a packed field.  13983 and 17519 are bugs in that feature.

In this case, nonpod_pack::n is not packed, so it is safe to pack fields of
nonpod_pack type.

Hmm, actually comment #2 does seem like a bug, both foos should compile
cleanly.

Or perhaps we should ignore the __attribute__ ((packed)) on nonpod_pack, since
one of its fields isn't safe to pack.  Yes, that seems like the right way to
handle this testcase.

Subject: Bug 26670

Author: jason
Date: Tue Aug 29 06:55:35 2006
New Revision: 116555

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=116555
Log:
        PR c++/26670
        * class.c (check_field_decls): Unset TYPE_PACKED (t) if one of the
        fields can't be packed.

Added:
    trunk/gcc/testsuite/g++.dg/ext/packed11.C
Modified:
    trunk/gcc/cp/class.c

Subject: Bug 26670

Author: jason
Date: Tue Aug 29 07:10:38 2006
New Revision: 116556

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=116556
Log:
        PR c++/26670
        * class.c (check_field_decls): Unset TYPE_PACKED (t) if one of the
        fields can't be packed.

Modified:
    trunk/gcc/cp/ChangeLog

Fixed.

reported against 4.1; bootstrapped 4.1.2 CVS with this patch with no
regressions. ok to apply to the 4.1 branch?

  Matthias

Subject: Bug 26670

Author: jason
Date: Wed Aug 30 15:51:17 2006
New Revision: 116591

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=116591
Log:
        PR c++/26670
        * class.c (check_field_decls): Don't unset TYPE_PACKED until all
        the fields have been processed.

Modified:
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/class.c

Subject: Bug 26670

Author: jason
Date: Wed Aug 30 15:52:12 2006
New Revision: 116592

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=116592
Log:
        PR c++/26670
        * class.c (check_field_decls): Don't unset TYPE_PACKED until all
        the fields have been processed.

        PR c++/26670
        * class.c (check_field_decls): Unset TYPE_PACKED (t) if one of the
        fields can't be packed.

Modified:
    branches/gcc-4_1-branch/gcc/cp/ChangeLog
    branches/gcc-4_1-branch/gcc/cp/class.c

fixed on 4.1 branch too.