Re: PR middle-end/51411: handle transaction_safe virtual inlined methods

[Patrick Marlier <patrick dot marlier at gmail dot com>]

On 12/14/2011 07:56 AM, Aldy Hernandez wrote:
> > Without the patch, it is ok.
> > If I remove completely this part:
> > if (DECL_DECLARED_INLINE_P (new_decl) && DECL_EXTERNAL (new_decl))
> > {
> > DECL_EXTERNAL (new_decl) = 0;
> > TREE_PUBLIC (new_decl) = 0;
> > }
> > It ends with *only* one undefined symbol:
> > src/UnitsSkins.o: In function
> > `_ZGTtStltIcSt11char_traitsIcESaIcEEbRKSbIT_T0_T1_ES8_':
> > ./include/c++/4.7.0/bits/basic_string.h:2568: undefined reference to
> > `_ZGTtNKSs7compareERKSs'
> >
> > $ c++filt _ZNKSs7compareERKSs
> > std::basic_string<char, std::char_traits<char>, std::allocator<char>
> > >::compare(std::basic_string<char, std::char_traits<char>,
> > std::allocator<char> > const&) const
> >
> > I am not sure the way to give you useful information.
>
> This is actually good, because we were missing a testcase for the
> original patch that zapped DECL_EXTERNAL and TREE_PUBLIC. Do you mind
> reducing a testcase for me so we don't regress with this patch? I know
> it'll be a bit challenging...

See attached. Probably not completely reduced...
A call to _ZGTtNKSs7compareERKSs is done but the clone is not created 
(the original function _ZNKSs7compareERKSs is defined in libstdc++). I 
am not sure how to add the rule for the testsuite.

> Perhaps we can try unsetting (hee hee hee) DECL_WEAK as originally
> proposed. Does this fix your problem?

Yep it does! ;)

Patrick.

> Index: trans-mem.c
> ===================================================================
> --- trans-mem.c (revision 182290)
> +++ trans-mem.c (working copy)
> @@ -4259,6 +4259,7 @@ ipa_tm_create_version (struct cgraph_nod
> {
> DECL_EXTERNAL (new_decl) = 0;
> TREE_PUBLIC (new_decl) = 0;
> + DECL_WEAK (new_decl) = 0;
> }
>
> tree_function_versioning (old_decl, new_decl, NULL, false, NULL,

Attachment: _ZGTtNKSs7compareERKSs.i
Description: Text document