MPIR doesn't compile with GCC-4.7.0 on ia64

[jdemeyer]

On Skynet iras (SUSE ES10 ia64):

Host system:
Linux iras 2.6.16.46-0.12-default #1 SMP Thu May 17 14:00:09 UTC 2007 ia64 ia64 ia64 GNU/Linux
****************************************************
C compiler: gcc
C compiler version:
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/local/gcc-4.7.0/ia64-Linux-suse/libexec/gcc/ia64-unknown-linux-gnu/4.7.0/lto-wrapper
Target: ia64-unknown-linux-gnu
Configured with: /usr/local/gcc-4.7.0/src/gcc-4.7.0/configure --enable-languages=c,c++,fortran --with-gnu-as --with-as=/usr/local/binutils
-2.22/ia64-Linux-suse-gcc-4.6.2/bin/as --with-gnu-ld --with-ld=/usr/local/binutils-2.22/ia64-Linux-suse-gcc-4.6.2/bin/ld --with-gmp=/usr/l
ocal/mpir-2.5.1/ia64-Linux-suse-gcc-4.6.2 --with-mpfr=/usr/local/mpfr-3.1.0/ia64-Linux-suse-mpir-2.5.1-gcc-4.6.2 --with-mpc=/usr/local/mpc
-0.9/ia64-Linux-suse-mpir-2.5.1-mpfr-3.1.0-gcc-4.6.2 --prefix=/usr/local/gcc-4.7.0/ia64-Linux-suse
Thread model: posix
gcc version 4.7.0 (GCC)
****************************************************
Your compiler does not support '-m32' nor '-m64'.  Leaving ABI unset.
Checking what CFLAGS MPIR would use if they were empty...
Settings chosen by MPIR when configuring with CFLAGS unset:
  CC:      gcc
  CFLAGS:  -O2
Settings required to properly build MPIR, taking into account SAGE_DEBUG etc.:
  CFLAGS:
  LDFLAGS:  -Wl,-z,noexecstack
  ABI:
Settings from the "global" environment:
  CC:      gcc
  CFLAGS:
  (CPPFLAGS, CXX and CXXFLAGS are listed below; these don't get modified.)
Using MPIR's settings (plus mandatory ones).
Finally using the following settings:
  CC=gcc
  CFLAGS=-O2
  CPPFLAGS=
  CXX=g++
  CXXFLAGS=
  LDFLAGS= -Wl,-z,noexecstack
  ABI=
(These settings may still get overridden by 'configure' or Makefiles.)
Configuring MPIR with the following options:
    --prefix="/home/jdemeyer/iras/sage-5.0.beta8-gcc/local" --enable-shared --enable-static --enable-gmpcompat --enable-cxx=yes --libdir=/
home/jdemeyer/iras/sage-5.0.beta8-gcc/local/lib
[...]
/bin/sh ../libtool --tag=CC   --mode=compile gcc -std=gnu99 -DHAVE_CONFIG_H -I. -I.. -D__GMP_WITHIN_GMP -I.. -DOPERATION_`echo divrem_eucl
idean_r_1 | sed 's/_$//'`    -O2   -c -o divrem_euclidean_r_1.lo divrem_euclidean_r_1.c
 gcc -std=gnu99 -DHAVE_CONFIG_H -I. -I.. -D__GMP_WITHIN_GMP -I.. -DOPERATION_divrem_euclidean_r_1 -O2 -c divrem_euclidean_r_1.c  -fPIC -DP
IC -o .libs/divrem_euclidean_r_1.o
divrem_euclidean_r_1.c: In function '__gmpn_mod_1_k':
divrem_euclidean_r_1.c:157:5: error: 'asm' operand requires impossible reload
make[2]: *** [divrem_euclidean_r_1.lo] Error 1
make[2]: Leaving directory `/home/jdemeyer/iras/sage-5.0.beta8-gcc/spkg/build/mpir-2.1.3.p9/src/mpn'
make[1]: *** [all-recursive] Error 1
make[1]: Leaving directory `/home/jdemeyer/iras/sage-5.0.beta8-gcc/spkg/build/mpir-2.1.3.p9/src'
make: *** [all] Error 2
Error building MPIR.

This is likely a GCC bug, see ​http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48496

There is a work-around (by essentially disabling all optimization) spkg at #11616.

See #12751 for the GCC-4.7.0 metaticket.

[leif]

Doesn't look like they've already fixed this upstream, they only work around the problem in libffi.

Or did you test some svn version?

[jdemeyer]

Replying to leif:

Doesn't look like they've already fixed this upstream, they only work around the problem in libffi.

Or did you test some svn version?

I haven't tested anything, I just believed the GCC Bugzilla.

[leif]

The MPIR 2.4.0.p2 spkg at #11616 now includes a work-around for this bug (almost completely disabling optimization on Linux ia64 with GCC 4.7.x).

[leif]

Looks like we'd have the same problem with NTL:

...
g++ -I../include -I.  -O2 -g -O3 -g -DHONORS_CXXFLAGS  -fPIC -c  FFT.c
In file included from ../include/NTL/ZZ.h:1804:0,
                 from ../include/NTL/FFT.h:7,
                 from FFT.c:3:
../include/NTL/SPMM_ASM.h: In function ‘void NTL::FFT(long int*, const long int*, long int, long int, const long int*)’:
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
make[3]: *** [FFT.o] Error 1
make[3]: Leaving directory `/home/leif/Sage/release/build/iras/sage-5.0.beta12-gcc-4.7.0/spkg/build/ntl-5.5.2/src/src'
make[2]: *** [libntl.so] Error 2
make[2]: Leaving directory `/home/leif/Sage/release/build/iras/sage-5.0.beta12-gcc-4.7.0/spkg/build/ntl-5.5.2/src/src'
Error creating ntl shared library.
...

[leif]

... also zn_poly.

[leif]

Replying to leif:

... also zn_poly.

And overrides -O0 ... in CFLAGS with -O3... >8(

[leif]

Replying to leif:

Looks like we'd have the same problem with NTL:

...
g++ -I../include -I.  -O2 -g -O3 -g -DHONORS_CXXFLAGS  -fPIC -c  FFT.c
In file included from ../include/NTL/ZZ.h:1804:0,
                 from ../include/NTL/FFT.h:7,
                 from FFT.c:3:
../include/NTL/SPMM_ASM.h: In function ‘void NTL::FFT(long int*, const long int*, long int, long int, const long int*)’:
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
../include/NTL/SPMM_ASM.h:103:69: error: ‘asm’ operand requires impossible reload
make[3]: *** [FFT.o] Error 1
make[3]: Leaving directory `/home/leif/Sage/release/build/iras/sage-5.0.beta12-gcc-4.7.0/spkg/build/ntl-5.5.2/src/src'
make[2]: *** [libntl.so] Error 2
make[2]: Leaving directory `/home/leif/Sage/release/build/iras/sage-5.0.beta12-gcc-4.7.0/spkg/build/ntl-5.5.2/src/src'
Error creating ntl shared library.
...

Despite that, after a second try with -fno-strict-aliasing -O3 (and again at least parts of these errors), the spkg is reported to have successfully been installed... ?

[leif]

Replying to leif:

Replying to leif:

... also zn_poly.

And overrides -O0 ... in CFLAGS with -O3... >8(

And that's not upstream. From spkg-install (which is a mess):

...
CFLAGS="$CFLAGS -g -fPIC -O3 -L."; export CFLAGS
LDFLAGS="$LDFLAGS"; export LDFLAGS
...

OMG.

[leif]

Furthermore: GMP-ECM, FLINT, MPFR and PARI.

[leif]

Jeroen, should I open separate (GCC 4.7 on ia64) tickets for these, or perhaps just one meta-ticket?

FLINT and zn_poly are currently "under construction" anyway; Andrew Ohana slightly fixed these just to respect CC and CXX for the clang port, so we could also do (little) changes on those open tickets.

(For FLINT there's also the ARM port ticket, which upgrades to 1.5.2 and thereby makes a couple of changes unnecessary, or at least different, with less effort. I also have an old, broadly cleaned-up FLINT 1.5.2 spkg, which I believe would need some work, at least some rebasing.)

[GMP-ECM, MPFR and PARI seem to build with -O0 [-finline-functions -fschedule-insns] in C/CPP/CXXFLAGS, without further changes to their spkg-install or Makefiles etc.; haven't yet tried with SAGE_CHECK=yes though.]

[jdemeyer]

Replying to leif:

Jeroen, should I open separate (GCC 4.7 on ia64) tickets for these, or perhaps just one meta-ticket?

To be honest, I don't care too much about these issues... It makes more sense to open a ticket to blacklist GCC-4.7.0 on ia64, i.e. build Sage's GCC-4.6.3 in this case.

I'm currently building an svn version of GCC-4.7.x to see whether the problem is fixed.

[jdemeyer]

MPFR though is a dependency for GCC, so that's more important.

[leif]

Replying to jdemeyer:

MPFR though is a dependency for GCC, so that's more important.

Yep.

[leif]

This one is more funny:

...
g++: error: unrecognized command line option ‘-m64’
make[2]: *** [gfan] Error 1
make[2]: Leaving directory `/home/leif/Sage/release/build/iras/sage-5.0.beta12-gcc-4.7.0/spkg/build/gfan-0.4plus.p2/src'
Error building gfan.
...

[jdemeyer]

That's weird because GCC on ia64 never supported -m64 AFAIK. So it should fail with any GCC version.

[leif]

Replying to leif:

This one is more funny:

...
g++: error: unrecognized command line option ‘-m64’
make[2]: *** [gfan] Error 1
make[2]: Leaving directory `/home/leif/Sage/release/build/iras/sage-5.0.beta12-gcc-4.7.0/spkg/build/gfan-0.4plus.p2/src'
Error building gfan.
...

The funniest thing being that that is my fault... B)

(since I made the p2, apparently without looking at the patched Makefile, and without testing on non-x86_64 systems)

[jdemeyer]

Replying to leif:

Doesn't look like they've already fixed this upstream, they only work around the problem in libffi.

You are right, it is not fixed upsteam (I tried the latest svn version of the 4.7 branch on cleo).

[leif]

Replying to leif:

Replying to leif:

Replying to leif:

... also zn_poly.

And overrides -O0 ... in CFLAGS with -O3... >8(

And that's not upstream. From spkg-install (which is a mess):

...
CFLAGS="$CFLAGS -g -fPIC -O3 -L."; export CFLAGS
LDFLAGS="$LDFLAGS"; export LDFLAGS
...

OMG.

Forgot: A fixed zn_poly spkg (not yet taking care of GCC 4.7.0 on ia64 IIRC) is available at #12433, still needing review. (The main purpose of that ticket was to make zn_poly respect CC etc.; it now also does a major clean-up.)