[mvapich-discuss] Fwd: Can't build MVAPICH2 w/ GCC 10.1 and PGI 20.4

[Korzennik, Sylvain]

GCC 10.1.0 config problem can be mitigated with
setenv  FCFLAGS "-std=legacy"
setenv  FFLAGS  "-std=legacy"

Got this answer to my GCC's 10.1.0 bug report:

> This is the intended behavior. Compile the faulty code with -std=legacy

As for PGI 20.4, no idea yet how to solve it.

    Stay sane, safe and healthy, 6+ft away! Cheers,
      Sylvain
--



---------- Forwarded message ---------
From: Korzennik, Sylvain <skorzennik at cfa.harvard.edu>
Date: Thu, Jul 2, 2020 at 9:40 AM
Subject: Can't build MVAPICH2 w/ GCC 10.1 and PGI 20.4
To: <mvapich-discuss at cse.ohio-state.edu>


I installed the following most recent compilers: GCC 9.3.0, GCC 10.1.0,
Intel 2020.1 and PGI 20.4 on our HPC cluster

I am able to build mvapich2-2.3.4 w/ GCC 9.3.0 and Intel 2020.1, but not
w/ GCC 10.1 and PGI 20.4

I use the same steps I used to build  mvapich2-2.3.1 that I
successfully build w/ older versions of GCC, Intel and PGI (close to 20
versions in total ;-).

GCC10.1.0 fails to configure on
checking whether gfortran allows mismatched arguments... no
configure: error: The Fortran compiler gfortran will not compile files that
call
the same routine with arguments of different types.

PGI 20.4 fails upon building (make) on
  CXXLD    lib/libmpicxx.la
pgc++-Error-Unknown switch: -pthread

I doubled checked and GCC pre-10.x can compile
conftest.f
::::::::::::::
      program main
      integer a
      real b
      character c
      call foo1(a)
      call foo1(b)
      call foo1(c)
      end
GCC 10.1 fails. I suspect that this is a GCC problem (IM(H)O their gfortran
is a catastrophe) and will report to GCC.

I looked into the PGI problem (reported to PGI as well), that step runs
/bin/sh ./libtool --silent --tag=CXX   --mode=link
/share/apps/pgi/linux86-64-llvm/20.4/bin/pgc++  -O4 -tp px  -DNDEBUG
-DNVALGRIND -version-info 13:1:1  -o lib/libmpicxx.la -rpath
/share/apps/pgi/20.4/mvapich2-2.3.4/lib src/binding/cxx/initcxx.lo lib/
libmpi.la -libmad -lrdmacm -libumad -libverbs -lrt -lpthread

and something is not set right since a -pthread flag is passed to pgc++,
that does not support it.

I was able to build mvapich2-2.3.1 w/ PGI 19.9, and trying the same thing
(building 2.3.1) w/ PGI 20.4 fails the same way.
Not clear if this is a PGI config problem or a libtool problem (to properly
handle PGI 20.4), PGI 19.9 pgc++ does not support -ptread, btw, so it looks
to me that it's the way libtool works w/ PGI 20.4

    Stay sane, safe and healthy, 6+ft away! Cheers,
      Sylvain
--
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20200702/daec6ef4/attachment.html>