Perl 5.24+ Configure loses hints script changes

[p5pRT]

Migrated from rt.perl.org#131415 (status was 'open')

Searchable as RT131415$

[p5pRT]

From al.dunsmuir@sympatico.ca

This is a bug report for perl from al.dunsmuir@​sympatico.ca,
generated with the help of perlbug 1.40 running under perl 5.24.1.

---

My Perl 5 port for AIX 5.3 (built May 2016) is 5.22.1, for both 32
and 64bit. I'm using XL C (xlc_r).

I started the port of 5.26.0, and the first perl file build with
miniperl (which happens to be 64-bit) failed with​:

A check of the Configure output for 5.26.0 showed
Any additional cc flags? [-q64 -brtl -bdynamic -DNEED_PTHREAD_INIT]
I attempted the same build with 5.24.1 and it had the same output.

The correct Configure output for 5.22.1 showed
Any additional cc flags? [-q64 -brtl -bdynamic -D_ALL_SOURCE
  -D_ANSI_C_SOURCE -D_POSIX_SOURCE -qmaxmem=-1 -qnoansialias
  -qlanglvl=extc99 -DUSE_NATIVE_DLOPEN -DNEED_PTHREAD_INIT]

The ccflags value passed to Configure in all cases was
'-q64 -brtl -bdynamic' for the 64 bit build.

There have been some large changes to Configure since 5.22.1, so
my first pass at diagnosing the cause was to add "set -x" to the
start of the 5.22 and 5.26 Configure scripts.

In 5.22, the options are set by calling hints/aix.sh as expected.

In 5.26, the options are initially set by calling a copy of the
hints/aix.sh (as UU/config.sh) at line 3818. Immediately after
at line 3826, the updated options are reverted to the original
command line options by calling optfile.sh (which appears to have
been added in 5.24 for resetting options after some new test
compiles).

Removing that optfile.sh leaves the options set, but really is not
sufficient to generate a correct build.
1) The hints file needs to be processed before the optfile.sh
  file is generated, so that any new tests are done using the
  recommended options for the platform/compiler.
2) Care needs to be taken for any inline tests to ensure that
  they are bypassed, or fail correctly when the build compiler
  is not gcc.
3) There seems to be overlap between the new long float inline
  tests in Configure and those in hints/aix.sh - I suspect that
  it would be better to take these into account in Configure
  and minimize any new redundant tests.

I'm going to continue with my 5.26 port effort, starting by
removing the call at line 3826.

Any port relying on options from a hints script will likely fail.
A workaround would be to specify any flags/options that should have
been added by the hints script to the Configure invocation.

A Google search showed what appear to be related error report​:.
- http​://www.nntp.perl.org/group/perl.mvs/2017/04/msg2458.html
  shows an identical error for the os390 port, using the similar
  IBM z/OS XL C compiler).
 
  In my day job, I'm a z/OS C programmer, and was part of IBM's C/C++
  compiler team for a number of years. I suspect -qnoansialias
  missing is the main problem for both platforms.

- I saw another another 5.24 miniperl error reported for an old Sun
  OS system that sounds suspeciously related.

Al Dunsmuir

Reported from my Fedora F25 system. This is not relevant to the
bug, so I removed all the following info except the flags.

---

---

Flags​:
  category=core
  severity=medium

---

  Message-Id​: <5.24.1_6508_1496260882@​wallace.alba.lan>
Subject​: Perl 5.24+ Configure loses hints script changes
Reply-To​: al.dunsmuir@​sympatico.ca
From​: al.dunsmuir@​sympatico.ca
To​: perlbug@​perl.org

This is a bug report for perl from al.dunsmuir@​sympatico.ca,
generated with the help of perlbug 1.40 running under perl 5.24.1.

---

My Perl 5 port for AIX 5.3 (built May 2016) is 5.22.1, for both 32
and 64bit. I'm using XL C (xlc_r).

I started the port of 5.26.0, and the first perl file build with
miniperl (which happens to be 64-bit) failed with​:

A check of the Configure output for 5.26.0 showed
Any additional cc flags? [-q64 -brtl -bdynamic -DNEED_PTHREAD_INIT]
I attempted the same build with 5.24.1 and it had the same output.

The correct Configure output for 5.22.1 showed
Any additional cc flags? [-q64 -brtl -bdynamic -D_ALL_SOURCE
  -D_ANSI_C_SOURCE -D_POSIX_SOURCE -qmaxmem=-1 -qnoansialias
  -qlanglvl=extc99 -DUSE_NATIVE_DLOPEN -DNEED_PTHREAD_INIT]

The ccflags value passed to Configure in all cases was
'-q64 -brtl -bdynamic' for the 64 bit build.

There have been some large changes to Configure since 5.22.1, so
my first pass at diagnosing the cause was to add "set -x" to the
start of the 5.22 and 5.26 Configure scripts.

In 5.22, the options are set by calling hints/aix.sh as expected.

In 5.26, the options are initially set by calling a copy of the
hints/aix.sh (as UU/config.sh) at line 3818. Immediately after
at line 3826, the updated options are reverted to the original
command line options by calling optfile.sh (which appears to have
been added in 5.24 for resetting options after some new test
compiles).

Removing that optfile.sh leaves the options set, but really is not
sufficient to generate a correct build.
1) The hints file needs to be processed before the optfile.sh
  file is generated, so that any new tests are done using the
  recommended options for the platform/compiler.
2) Care needs to be taken for any inline tests to ensure that
  they are bypassed, or fail correctly when the build compiler
  is not gcc.
3) There seems to be overlap between the new long float inline
  tests in Configure and those in hints/aix.sh - I suspect that
  it would be better to take these into account in Configure
  and minimize any new redundant tests.

I'm going to continue with my 5.26 port effort, starting by
removing the call at line 3826.

Any port relying on options from a hints script will likely fail.
A workaround would be to specify any flags/options that should have
been added by the hints script to the Configure invocation.

A Google search showed what appear to be related error report​:.
- http​://www.nntp.perl.org/group/perl.mvs/2017/04/msg2458.html
  shows an identical error for the os390 port, using the similar
  IBM z/OS XL C compiler).
 
  In my day job, I'm a z/OS C programmer, and was part of IBM's C/C++
  compiler team for a number of years. I suspect -qnoansialias
  missing is the main problem for both platforms.

- I saw another another 5.24 miniperl error reported for an old Sun
  OS system that sounds suspeciously related.

Al Dunsmuir

Reported from my Fedora F25 system. This is not relevant to the
bug, so I removed all the following info except the flags.

---

---

Flags​:
  category=core
  severity=medium

---

[p5pRT]

From al.dunsmuir@sympatico.ca

Actual miniperl failure on AIX 5.3 follows. Same error as OS390 port.
--
Al Dunsmuir

xlc_r -q64 -c -DPERL_CORE -q64 -brtl -bdynamic -DNEED_PTHREAD_INIT -q64 -DUSE_64_BIT_ALL -q64 -O miniperlmain.c
xlc_r -q64 -o miniperl -q64 -s -Wl,-bmaxdata​:0x80000000 -L/opt/freeware/lib64 -L/usr/lib64 -L/opt/freeware/lib -L/usr/lib -b64 opmini.o perlmini.o gv.o toke.o perly.o pad.o regcomp.o dump.o util.o mg.o reentr.o mro_core.o keywords.o hv.o av.o run.o pp_hot.o sv.o pp.o scope.o pp_ctl.o pp_sys.o doop.o doio.o regexec.o utf8.o taint.o deb.o universal.o globals.o perlio.o perlapi.o numeric.o mathoms.o locale.o pp_pack.o pp_sort.o caretx.o dquote.o time64.o miniperlmain.o -lpthread -lbind -lnsl -ldl -lld -lm -lcrypt -lpthreads -lc
LIBPATH=/apps/rpmbuild/BUILD/perl-5.26.0-RC2/64bit ./miniperl -w -Ilib -Idist/Exporter/lib -MExporter -e '<?>' || sh -c 'echo >&2 Failed to build miniperl. Please run make minitest; exit 1'
Invalid version object at dist/Exporter/lib/Exporter.pm line 3.
Compilation failed in require.
BEGIN failed--compilation aborted.
Failed to build miniperl. Please run make minitest
make​: *** [makefile​:385​: lib/buildcustomize.pl] Error 1

[p5pRT]

From @jkeenan

On Wed, 31 May 2017 21​:37​:46 GMT, al.dunsmuir@​sympatico.ca wrote​:

This is a bug report for perl from al.dunsmuir@​sympatico.ca,
generated with the help of perlbug 1.40 running under perl 5.24.1.

-----------------------------------------------------------------
My Perl 5 port for AIX 5.3 (built May 2016) is 5.22.1, for both 32
and 64bit. I'm using XL C (xlc_r).

I started the port of 5.26.0, and the first perl file build with
miniperl (which happens to be 64-bit) failed with​:

A check of the Configure output for 5.26.0 showed
Any additional cc flags? [-q64 -brtl -bdynamic -DNEED_PTHREAD_INIT]
I attempted the same build with 5.24.1 and it had the same output.

The correct Configure output for 5.22.1 showed
Any additional cc flags? [-q64 -brtl -bdynamic -D_ALL_SOURCE
-D_ANSI_C_SOURCE -D_POSIX_SOURCE -qmaxmem=-1 -qnoansialias
-qlanglvl=extc99 -DUSE_NATIVE_DLOPEN -DNEED_PTHREAD_INIT]

The ccflags value passed to Configure in all cases was
'-q64 -brtl -bdynamic' for the 64 bit build.

Would you be able to attach 'config.sh' for each of 5.22.1 and 5.26.0?

Our Configure experts will probably get some insight from them.

Thank you very much.

--
James E Keenan (jkeenan@​cpan.org)

[p5pRT]

The RT System itself - Status changed from 'new' to 'open'

[p5pRT]

From al.dunsmuir@sympatico.ca

On Wed, 31 May 2017 19​:18​:38 -0700, jkeenan wrote​:

Would you be able to attach 'config.sh' for each of 5.22.1 and 5.26.0?

Our Configure experts will probably get some insight from them.

Thank you very much.

File 1 is perl-5.22.1 and file 2 is perl-5.26.0
--
Al Dunsmuir

[p5pRT]

From al.dunsmuir@sympatico.ca

config.sh

[p5pRT]

From al.dunsmuir@sympatico.ca

Looks like this form gets confused when you attach 2 files with the same basename.
The last file (5.26.0 version) ended up being uploaded.

Now attaching the 5.22.1 version.

--
Al Dunsmuir

[p5pRT]

From al.dunsmuir@sympatico.ca

config.sh

[p5pRT]

From [Unknown Contact. See original ticket]

Looks like this form gets confused when you attach 2 files with the same basename.
The last file (5.26.0 version) ended up being uploaded.

Now attaching the 5.22.1 version.

--
Al Dunsmuir

[p5pRT]

From @Tux

On Thu, 01 Jun 2017 08​:26​:05 -0700, "Al Dunsmuir via RT"
<perlbug-followup@​perl.org> wrote​:

On Wed, 31 May 2017 19​:18​:38 -0700, jkeenan wrote​:

Would you be able to attach 'config.sh' for each of 5.22.1 and 5.26.0?

Our Configure experts will probably get some insight from them.

Thank you very much.

File 1 is perl-5.22.1 and file 2 is perl-5.26.0

I see just one file (without version in the name)

--
H.Merijn Brand http​://tux.nl Perl Monger http​://amsterdam.pm.org/
using perl5.00307 .. 5.25 porting perl5 on HP-UX, AIX, and openSUSE
http​://mirrors.develooper.com/hpux/ http​://www.test-smoke.org/
http​://qa.perl.org http​://www.goldmark.org/jeff/stupid-disclaimers/

[p5pRT]

From @jkeenan

On Thu, 01 Jun 2017 15​:32​:12 GMT, al.dunsmuir@​sympatico.ca wrote​:

Looks like this form gets confused when you attach 2 files with the
same basename.
The last file (5.26.0 version) ended up being uploaded.

Now attaching the 5.22.1 version.

--
James E Keenan (jkeenan@​cpan.org)

[p5pRT]

From @jkeenan

131415-5.22.1-config.sh

[p5pRT]

From @jkeenan

131415-5.26.0-config.sh

[p5pRT]

From @jkeenan

On Fri, 02 Jun 2017 10​:20​:55 GMT, jkeenan wrote​:

On Thu, 01 Jun 2017 15​:32​:12 GMT, al.dunsmuir@​sympatico.ca wrote​:

Looks like this form gets confused when you attach 2 files with the
same basename.
The last file (5.26.0 version) ended up being uploaded.

Now attaching the 5.22.1 version.

I have uploaded the two config.sh files, only now with names including their Perl version numbers.

I notice one peculiarity in 131415-5.26.0-config.sh. In the config_args you have​:

#####
-Dinc_version_list=5.26.1
#####

This must be a typo, as version 5.26.1 does not yet exist. Shouldn't it be​:

#####
-Dinc_version_list=5.26.0
#####

--
James E Keenan (jkeenan@​cpan.org)

[p5pRT]

From @doughera88

On Wed, May 31, 2017 at 02​:37​:47PM -0700, Al Dunsmuir wrote​:

# New Ticket Created by Al Dunsmuir
# Please include the string​: [perl #131415]
# in the subject line of all future correspondence about this issue.
# <URL​: https://rt-archive.perl.org/perl5/Ticket/Display.html?id=131415 >

-----------------------------------------------------------------
My Perl 5 port for AIX 5.3 (built May 2016) is 5.22.1, for both 32
and 64bit. I'm using XL C (xlc_r).

I started the port of 5.26.0, and the first perl file build with
miniperl (which happens to be 64-bit) failed with​:

A check of the Configure output for 5.26.0 showed
Any additional cc flags? [-q64 -brtl -bdynamic -DNEED_PTHREAD_INIT]
I attempted the same build with 5.24.1 and it had the same output.

The correct Configure output for 5.22.1 showed
Any additional cc flags? [-q64 -brtl -bdynamic -D_ALL_SOURCE
-D_ANSI_C_SOURCE -D_POSIX_SOURCE -qmaxmem=-1 -qnoansialias
-qlanglvl=extc99 -DUSE_NATIVE_DLOPEN -DNEED_PTHREAD_INIT]

The ccflags value passed to Configure in all cases was
'-q64 -brtl -bdynamic' for the 64 bit build.

There have been some large changes to Configure since 5.22.1, so
my first pass at diagnosing the cause was to add "set -x" to the
start of the 5.22 and 5.26 Configure scripts.

In 5.22, the options are set by calling hints/aix.sh as expected.

In 5.26, the options are initially set by calling a copy of the
hints/aix.sh (as UU/config.sh) at line 3818. Immediately after
at line 3826, the updated options are reverted to the original
command line options by calling optfile.sh (which appears to have
been added in 5.24 for resetting options after some new test
compiles).

I suspect this is another unintended consequence of

  commit 41d7307
  Author​: Ivan Pozdeev <vano@​mail.mipt.ru>
  Date​: Wed Aug 12 20​:33​:12 2015 +0300

  Make -O behaviour the default

which, unforutnately, did not include an option to skip the -O behaviour.
If it had, using that option would have reverted to the old behavior,
and would probably fix at least this initial problem.

One solution might be to adjust the way you call Configure to use
-Accflags instead of -Dccflags. (Similarly, you probably should change
-Dldflags to -Aldflags.) I'd try that first.

If that fails, your workaround of deleting line 3826 in Configure
is a sensible path forward.

Once that is in hand, it can make sense to look at specific failing
Configure or hints checks. I haven't looked at any of the long float
stuff or at AIX in ages, so I'm unlikely to be of much help there.

--
  Andy Dougherty doughera@​lafayette.edu