Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE> writes:

> While investigating PR bootstrap/84017, it turned out that it *is*
> possible to enable COMDAT group support on Solaris 10 with Solaris ld in
> some circumstances.  I'm posting my findings and the resulting patch
> here for reference only; this is certainly not GCC 8 material.
>
> Besides, there's a abi_check failure in libstdc++ which makes it unclear
> if this desirable at all.
>
> So here's the deal: I remembered that Solaris ld has a -z relaxreloc
> option, documented as
>
>      -z relaxreloc
>
>          ld normally issues a fatal  error  upon  encountering  a
>          relocation  using a symbol that references an eliminated
>          COMDAT section. If -z relaxreloc is enabled, ld  instead
>          redirects  such  relocations to the equivalent symbol in
>          the COMDAT section that was kept.  -z  relaxreloc  is  a
>          specialized  option,  mainly  of  interest  to  compiler
>          authors, and is not intended for general use.
>
> It was only introduced in Solaris 10 Update 10, so won't help on earlier
> versions.
>
> When ld supports that option, I always pass it when linking and enable
> comdat_group (and hidden_linkonce) in configure.ac.
>
> Test results are astonishingly good:
>
> * On Solaris 10/x86, I get three tests where ld SEGVs with -flto
>   (skipped/xfailed in the present patch).
>
> * On both Solaris 10/SPARC and x86, libstdc++-abi/abi_check FAILs:
>
> # of added symbols:              2
> # of missing symbols:            0
> # of undesignated symbols:       0
> # of incompatible symbols:       2
>
> 2 added symbols
> 0
> _ZNSt7__cxx1115basic_stringbufIwSt11char_traitsIwESaIwEED2Ev
> std::__cxx11::basic_stringbuf<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::~basic_stringbuf()
> version status: incompatible
> GLIBCXX_3.4.21
> type: function
> status: added
>
> 1
> _ZNSt7__cxx1115basic_stringbufIcSt11char_traitsIcESaIcEED2Ev
> std::__cxx11::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >::~basic_stringbuf()
> version status: incompatible
> GLIBCXX_3.4.21
> type: function
> status: added
>
>
> 2 incompatible symbols
> 0
> _ZNSt7__cxx1115basic_stringbufIwSt11char_traitsIwESaIwEED2Ev
> std::__cxx11::basic_stringbuf<wchar_t, std::char_traits<wchar_t>, std::allocator<wchar_t> >::~basic_stringbuf()
> version status: incompatible
> GLIBCXX_3.4.21
> type: function
> status: added
>
>
> 1
> _ZNSt7__cxx1115basic_stringbufIcSt11char_traitsIcESaIcEED2Ev
> std::__cxx11::basic_stringbuf<char, std::char_traits<char>, std::allocator<char> >::~basic_stringbuf()
> version status: incompatible
> GLIBCXX_3.4.21
> type: function
> status: added
>
>   Those symbols are present in the Solaris 11 baselines (where COMDAT
>   group always works), which is also used with Solaris 10 and GNU ld.
>
>   I'm not yet sure how to deal with this: we'd have to introduce 3
>   baselines:
>
>   * Solaris 10 without COMDAT group support
>   * Solaris 10 with COMDAT group support
>   * Solaris 11
>
>   Seems like a lot of trouble to me actually.  I've not done anything
>   about this yet, so as is the patch shows the abi_check failures.

I've now revised the patch a bit, especially for the abi_check failure:

* Since Solaris 10 with gld (which already did support COMDAT group)
  already used the *-solaris2.11 baselines, I've extended that to any
  Solaris 10 configuration with COMDAT group support.  This needed a
  configure test in libstdc++-v3/acinclude.m4 matching what the
  comdat_group effective-target keyword does.  However, I couldn't use
  a simple AC_TRY_COMPILE with -S added to CXXFLAGS: that will always
  fail since the macro checks for the existence of conftest.o as first
  indication of success, which of course doesn't exist with -S.

* I've also tried to enable USE_HIDDEN_LINKONCE on Solaris 10/x86 with
  COMDAT group, but that leads to four more testcases where ld SEGVs, so
  I've left it disabled.

Otherwise, nothing remarkable: bootstrapped without regressions on
i386-pc-solaris2.10 and sparc-sun-solaris2.10 (both as/ld and gas/ld)
without and with --disable-comdat; also bootstrapped on some
*-*-solaris2.11 configurations, no regressions.

I'm just posting this for comment here; I may or may not commit this in
the GCC 9 timeframe.

	Rainer

-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University


2018-01-28  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

	libstdc++-v3:
	* acinclude.m4 (GLIBCXX_CONFIGURE): Check for COMDAT group
	support.
	* configure: Regenerate.
	* configure.host (*-*-solaris2.1[0-9])): Check $comdat_group to
	select baselines.

	gcc/testsuite:
	* g++.dg/lto/pr65475c_0.C: Skip on Solaris 10/x86 with ld/comdat.
	* g++.dg/lto/pr68057_0.C: Likewise.
	* gcc.dg/debug/pr41893-1.c: Xfail on Solaris 10/x86 with ld/comdat
	for -gdwarf-2 -g3.

	gcc:
	* configure.ac (gcc_cv_ld_relaxreloc): New test.
	(comdat_group): Enable on Solaris 10 if ld -z relaxreloc is present.
	* configure: Regenerate.
	* config.in: Regenerate.
	* config/sol2.h (LINK_RELAXRELOC_SPEC): Define.
	(LINK_SPEC): Use it.