GCC testing on FreeBSD

GCC testing on FreeBSD

[Jonathan Wakely]

Hi Gerald, Andreas,

How are you testing on FreeBSD?

When I build GCC trunk on FreeBSD 14.0 and try to run the libstdc++
testsuite it fails due to lots of these errors:

Excess errors:
/usr/local/bin/ld: /tmp//ccev946q.o: relocation R_X86_64_32 against
symbol `_ZTIN10__cxxabiv115__forced_unwindE@@CXXABI_1.3.2' can not be
used when making a PDE object; recompile with -fPIE
/usr/local/bin/ld: failed to set dynamic section sizes: bad value

Which suggests that -fPIE is missing from the default test flags.

Have you seen this? Do you do something locally to work around it?

Thanks,
Jonathan

Re: GCC testing on FreeBSD

[Gerald Pfeifer]

Hi Jonathan,

On Fri, 26 Apr 2024, Jonathan Wakely wrote:
> How are you testing on FreeBSD?
> 
> When I build GCC trunk on FreeBSD 14.0 and try to run the libstdc++
> testsuite it fails due to lots of these errors:
> 
> Excess errors:
> /usr/local/bin/ld: /tmp//ccev946q.o: relocation R_X86_64_32 against
> symbol `_ZTIN10__cxxabiv115__forced_unwindE@@CXXABI_1.3.2' can not be
> used when making a PDE object; recompile with -fPIE
> /usr/local/bin/ld: failed to set dynamic section sizes: bad value

my first reaction was to recommend using binutils instead of /usr/bin/ld 
which is LLD 16.0.6 or similar (since a while ago FreeBSD switched to that
toolchain as part of the base system).

My nightly tester has been using GNU ld since

  # 2012-03-11  Configure using --with-as=$localbase/bin/as and
  #             --with-ld=$localbase/bin/ld on *.freebsd.org.

in the script invoked by cron, even before FreeBSD made that switch.

Seeing /usr/local/bin/ld in the error message it appears you are doing 
that already, though?

> Which suggests that -fPIE is missing from the default test flags.
> 
> Have you seen this? Do you do something locally to work around it?

All I have in terms of adjustments to the FreeBSD systems I build on via 
that script are the following

  CONFIGUREFLAGS="--with-gmp=$LOCALBASE --with-as=$LOCALBASE/bin/as $CONFIGUREFLAGS"
  LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$LOCALBASE/lib
  PATH=$LOCALBASE/bin:$PATH

where LOCALBASE looks like it would be /usr/local in your case. Nothing 
explicit around PIE there.

Looking at the logs of a serialized build I triggered manually where I 
remove the --with-ld configure option, I see 

  checking linker PIE support with copy reloc... no
  checking for -fno-PIE option... yes
  checking for -no-pie option... yes

and then build invocations like

  c++ -std=c++11  -fno-PIE -c ...

during all-stage1-target-libgcc which ultimately fails with - mystery?! -

  ld: error: unable to find library -lc
  collect2: error: ld returned 1 exit status
  gmake[3]: *** [Makefile:1005: libgcc_s.so] Error 1

(Disclaimer: all my tests on FreeBSD 13.2, not FreeBSD 14 as in your 
case.)


Looking at the lang/gcc* ports that I maintained for two decades until 
Lorenzo (copied now) kindly took them over two years ago I see the 
following change among others for newer versions:

  % git show b6a5871a0cf40
  commit b6a5871a0cf40dfc194217704e2dc03e2e91fb62
  Author: Lorenzo Salvadore <salvadore@FreeBSD.org>
  Date:   Fri Feb 3 20:12:49 2023 +0100

    lang/gcc10: Mark PIE_UNSAFE
    
    Building the port with WITH_PIE fails if the BOOTSTRAP option is
    enabled. Mark PIE_UNSAFE when this option is enabled until a better
    solution is found.
    
    PR:             268901

where https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=268901 is the PR 
referenced.

PIE_UNSAFE=yes in a port's Makefile (the equivalent to a spec file in RPM 
land) is used as follows in ports/Mk/Features/pie.mk:

  .  if !defined(PIE_UNSAFE)
  PIE_CFLAGS?=    -fPIE -fPIC
  CFLAGS+=        ${PIE_CFLAGS}
  CXXFLAGS+=      ${PIE_CFLAGS}
  LDFLAGS+=       -pie
  STATIC_PIE_ARGS+=       -static-pie
  .  endif


You are not using the Ports Infrastructure, I believe, so the above does 
not directly apply, may provide some additional background, though?


Hope this helps - and please chime in Lorenzo and Andreas!

Gerald

Re: GCC testing on FreeBSD

[Rainer Orth]

Hi Gerald,

> On Fri, 26 Apr 2024, Jonathan Wakely wrote:
>> How are you testing on FreeBSD?
>> 
>> When I build GCC trunk on FreeBSD 14.0 and try to run the libstdc++
>> testsuite it fails due to lots of these errors:
>> 
>> Excess errors:
>> /usr/local/bin/ld: /tmp//ccev946q.o: relocation R_X86_64_32 against
>> symbol `_ZTIN10__cxxabiv115__forced_unwindE@@CXXABI_1.3.2' can not be
>> used when making a PDE object; recompile with -fPIE
>> /usr/local/bin/ld: failed to set dynamic section sizes: bad value
>
> my first reaction was to recommend using binutils instead of /usr/bin/ld 
> which is LLD 16.0.6 or similar (since a while ago FreeBSD switched to that
> toolchain as part of the base system).

right: some time ago I tried bootstrapping trunk on FreeBSD 14.0 myself,
which was a very rough journey unfortunately.  I've filed a number of
PRs for the issues found.  They should all be referenced in

PR target/112959	install.tex needs updates on FreeBSD

which is primarily a call for documention so the next unwary non-FreeBSD
developer won't run into all the same known issues again.

	Rainer

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

Re: GCC testing on FreeBSD

[Lorenzo Salvadore]

On Sunday, April 28th, 2024 at 12:24, Gerald Pfeifer <gerald@pfeifer.com> wrote:

> On Fri, 26 Apr 2024, Jonathan Wakely wrote:
> 
> > How are you testing on FreeBSD?
> > 
> > When I build GCC trunk on FreeBSD 14.0 and try to run the libstdc++
> > testsuite it fails due to lots of these errors:
> > 
> > Excess errors:
> > /usr/local/bin/ld: /tmp//ccev946q.o: relocation R_X86_64_32 against
> > symbol `_ZTIN10__cxxabiv115__forced_unwindE@@CXXABI_1.3.2' can not be
> > used when making a PDE object; recompile with -fPIE
> > /usr/local/bin/ld: failed to set dynamic section sizes: bad value

Hi Gerald and Jonathan!

I normally test every weekly GCC snapshots through the FreeBSD ports
framework on Cirrus, so that all my tests are publicly accessible:
http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc11-devel
http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc12-devel
http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc13-devel
http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc14-devel

And of course the cirrus configuration is public as well:
https://github.com/lsalvadore/freebsd-ports/blob/lang/gcc11-devel/.cirrus.yml

By the way, is it possible for FreeBSD to enter GCC's CI pipeline? Where
would one start? I am willing to help of course.

Cheers,

Lorenzo Salvadore

Re: GCC testing on FreeBSD

[Rainer Orth]

Hi Lorenzo,

> On Sunday, April 28th, 2024 at 12:24, Gerald Pfeifer <gerald@pfeifer.com> wrote:
>
>> On Fri, 26 Apr 2024, Jonathan Wakely wrote:
>> 
>> > How are you testing on FreeBSD?
>> > 
>> > When I build GCC trunk on FreeBSD 14.0 and try to run the libstdc++
>> > testsuite it fails due to lots of these errors:
>> > 
>> > Excess errors:
>> > /usr/local/bin/ld: /tmp//ccev946q.o: relocation R_X86_64_32 against
>> > symbol `_ZTIN10__cxxabiv115__forced_unwindE@@CXXABI_1.3.2' can not be
>> > used when making a PDE object; recompile with -fPIE
>> > /usr/local/bin/ld: failed to set dynamic section sizes: bad value
>
> Hi Gerald and Jonathan!
>
> I normally test every weekly GCC snapshots through the FreeBSD ports
> framework on Cirrus, so that all my tests are publicly accessible:
> http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc11-devel
> http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc12-devel
> http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc13-devel
> http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc14-devel
>
> And of course the cirrus configuration is public as well:
> https://github.com/lsalvadore/freebsd-ports/blob/lang/gcc11-devel/.cirrus.yml

this isn't particularly helpful if you just try to build upstream GCC
for comparision with your own targets or to verify a patch of yours.
Having to go hunting for configs like this if you're not a regular
FreeBSD user is a no-no IMO.  GCC trunk should either build out of the
box or the quirks be documented in install.texi.  Otherwise, non-FreeBSD
developers will get frustrated and give up on the target, to the
detriment both of their patches and the platform.

Unfortunately, it's pretty common that targets keep necessary patches in
some ports collection of their own (usually a different one per target)
and neglect to submit them upstream.

	Rainer

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

Re: GCC testing on FreeBSD

[Lorenzo Salvadore]

On Sunday, April 28th, 2024 at 19:19, Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE> wrote:

> 
> 
> Hi Lorenzo,
> 
> > On Sunday, April 28th, 2024 at 12:24, Gerald Pfeifer gerald@pfeifer.com wrote:
> > 
> > > On Fri, 26 Apr 2024, Jonathan Wakely wrote:
> > > 
> > > > How are you testing on FreeBSD?
> > > > 
> > > > When I build GCC trunk on FreeBSD 14.0 and try to run the libstdc++
> > > > testsuite it fails due to lots of these errors:
> > > > 
> > > > Excess errors:
> > > > /usr/local/bin/ld: /tmp//ccev946q.o: relocation R_X86_64_32 against
> > > > symbol `_ZTIN10__cxxabiv115__forced_unwindE@@CXXABI_1.3.2' can not be
> > > > used when making a PDE object; recompile with -fPIE
> > > > /usr/local/bin/ld: failed to set dynamic section sizes: bad value
> > 
> > Hi Gerald and Jonathan!
> > 
> > I normally test every weekly GCC snapshots through the FreeBSD ports
> > framework on Cirrus, so that all my tests are publicly accessible:
> > http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc11-devel
> > http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc12-devel
> > http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc13-devel
> > http://cirrus-ci.com/github/lsalvadore/freebsd-ports/lang/gcc14-devel
> > 
> > And of course the cirrus configuration is public as well:
> > https://github.com/lsalvadore/freebsd-ports/blob/lang/gcc11-devel/.cirrus.yml
> 
> 
> this isn't particularly helpful if you just try to build upstream GCC
> for comparision with your own targets or to verify a patch of yours.
> Having to go hunting for configs like this if you're not a regular
> FreeBSD user is a no-no IMO. GCC trunk should either build out of the
> box or the quirks be documented in install.texi. Otherwise, non-FreeBSD
> developers will get frustrated and give up on the target, to the
> detriment both of their patches and the platform.
> 
> Unfortunately, it's pretty common that targets keep necessary patches in
> some ports collection of their own (usually a different one per target)
> and neglect to submit them upstream.

I totally agree with you: upstreaming patches is important! It is not
only for the upstream project itself, but for the target as well: having
patches sitting in a ports collection also requires more maintainance,
they require to be kept up to date with the upstream progresses.

Unfortunately, it is not always possible to upstream a patch. Sometimes,
patches are too specific to a target to be suitable for upstream.
For example, smaller projects might be interested in supporting only
very popular platforms and would not accept (or could not accept)
the complexity of supporting a less known target.
Hopefully this is not the case for GCC.

Other times, developers do try to upstream a patch, but fail. This
happened to myself in GCC too, when I was unable to get any
attention to a patch I submitted, and could not do any better
than keep the patch into the FreeBSD ports collection:
https://gcc.gnu.org/pipermail/jit/2023q3/001642.html
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101491#c5

If you are able to help with this upstreaming, I would appreciate
it a lot. Thanks.

Cheers,

Lorenzo Salvadore

Re: GCC testing on FreeBSD

[Rainer Orth]

Hi Lorenzo,

> I totally agree with you: upstreaming patches is important! It is not
> only for the upstream project itself, but for the target as well: having
> patches sitting in a ports collection also requires more maintainance,
> they require to be kept up to date with the upstream progresses.

indeed: e.g. some mechanical changes that affect your port can be dealt
with easily when the developer in question finds the code in tree.
Otherwise, this is only noticed during your next merge, possibly months
later.

> Unfortunately, it is not always possible to upstream a patch. Sometimes,
> patches are too specific to a target to be suitable for upstream.

Those are called hacks ;-)  True: the code needs to be general to avoid
the codebase slowly becoming unmaintainable.  Maintainers keep the
general code structure in mind, which may create more work for the
submitter to develop a cleaner implementation.  However, this ensures
that the hack isn't broken subsequently and the code is more
understandable to everyone.

> For example, smaller projects might be interested in supporting only
> very popular platforms and would not accept (or could not accept)
> the complexity of supporting a less known target.
> Hopefully this is not the case for GCC.

I don't think so: the primary question is if the code is actively tested
and maintained.  And even so, we do have several ports in tree that have
seen barely any maintenance in a long time.  If they become a burden,
this may lead to them being obsoleted and removed.

> Other times, developers do try to upstream a patch, but fail. This
> happened to myself in GCC too, when I was unable to get any
> attention to a patch I submitted, and could not do any better
> than keep the patch into the FreeBSD ports collection:
> https://gcc.gnu.org/pipermail/jit/2023q3/001642.html
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101491#c5

It happens to all of us.  It usually helps to include the maintainers in
question in the Cc:  I've often also had better success by pinging not
in the patch submission thread (where the pings can become lost in the
noise), but by posting a separate message to gcc-patches with subject
and URL to the latest version and perhaps a short indication who's
required to review the patch, also Cc'ing the relevant maintainers.

> If you are able to help with this upstreaming, I would appreciate
> it a lot. Thanks.

Probably not much: for one I'm pretty busy with my own Solaris work, and
I've also stopped testing on FreeBSD even with the issues I've found and
developed workarounds for.

	Rainer

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

Re: GCC testing on FreeBSD

[Jonathan Wakely]

[-- Attachment #1: Type: text/plain, Size: 3721 bytes --]

On Sun, 28 Apr 2024, 11:24 Gerald Pfeifer, <gerald@pfeifer.com> wrote:

> Hi Jonathan,
>
> On Fri, 26 Apr 2024, Jonathan Wakely wrote:
> > How are you testing on FreeBSD?
> >
> > When I build GCC trunk on FreeBSD 14.0 and try to run the libstdc++
> > testsuite it fails due to lots of these errors:
> >
> > Excess errors:
> > /usr/local/bin/ld: /tmp//ccev946q.o: relocation R_X86_64_32 against
> > symbol `_ZTIN10__cxxabiv115__forced_unwindE@@CXXABI_1.3.2' can not be
> > used when making a PDE object; recompile with -fPIE
> > /usr/local/bin/ld: failed to set dynamic section sizes: bad value
>
> my first reaction was to recommend using binutils instead of /usr/bin/ld
> which is LLD 16.0.6 or similar (since a while ago FreeBSD switched to that
> toolchain as part of the base system).
>
> My nightly tester has been using GNU ld since
>
>   # 2012-03-11  Configure using --with-as=$localbase/bin/as and
>   #             --with-ld=$localbase/bin/ld on *.freebsd.org.
>
> in the script invoked by cron, even before FreeBSD made that switch.
>
> Seeing /usr/local/bin/ld in the error message it appears you are doing
> that already, though?
>

Yes, I'm using GNU ld.

On IRC Andreas suggested that also using GNU objdump would solve my issue,
but I haven't tried yet.



> > Which suggests that -fPIE is missing from the default test flags.
> >
> > Have you seen this? Do you do something locally to work around it?
>
> All I have in terms of adjustments to the FreeBSD systems I build on via
> that script are the following
>
>   CONFIGUREFLAGS="--with-gmp=$LOCALBASE --with-as=$LOCALBASE/bin/as
> $CONFIGUREFLAGS"
>   LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$LOCALBASE/lib
>   PATH=$LOCALBASE/bin:$PATH
>
> where LOCALBASE looks like it would be /usr/local in your case. Nothing
> explicit around PIE there.
>
> Looking at the logs of a serialized build I triggered manually where I
> remove the --with-ld configure option, I see
>
>   checking linker PIE support with copy reloc... no
>   checking for -fno-PIE option... yes
>   checking for -no-pie option... yes
>
> and then build invocations like
>
>   c++ -std=c++11  -fno-PIE -c ...
>
> during all-stage1-target-libgcc which ultimately fails with - mystery?! -
>
>   ld: error: unable to find library -lc
>   collect2: error: ld returned 1 exit status
>   gmake[3]: *** [Makefile:1005: libgcc_s.so] Error 1
>
> (Disclaimer: all my tests on FreeBSD 13.2, not FreeBSD 14 as in your
> case.)
>
>
> Looking at the lang/gcc* ports that I maintained for two decades until
> Lorenzo (copied now) kindly took them over two years ago I see the
> following change among others for newer versions:
>
>   % git show b6a5871a0cf40
>   commit b6a5871a0cf40dfc194217704e2dc03e2e91fb62
>   Author: Lorenzo Salvadore <salvadore@FreeBSD.org>
>   Date:   Fri Feb 3 20:12:49 2023 +0100
>
>     lang/gcc10: Mark PIE_UNSAFE
>
>     Building the port with WITH_PIE fails if the BOOTSTRAP option is
>     enabled. Mark PIE_UNSAFE when this option is enabled until a better
>     solution is found.
>
>     PR:             268901
>
> where https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=268901 is the PR
> referenced.
>
> PIE_UNSAFE=yes in a port's Makefile (the equivalent to a spec file in RPM
> land) is used as follows in ports/Mk/Features/pie.mk:
>
>   .  if !defined(PIE_UNSAFE)
>   PIE_CFLAGS?=    -fPIE -fPIC
>   CFLAGS+=        ${PIE_CFLAGS}
>   CXXFLAGS+=      ${PIE_CFLAGS}
>   LDFLAGS+=       -pie
>   STATIC_PIE_ARGS+=       -static-pie
>   .  endif
>
>
> You are not using the Ports Infrastructure, I believe, so the above does
> not directly apply, may provide some additional background, though?
>
>
> Hope this helps - and please chime in Lorenzo and Andreas!
>
> Gerald
>