From: David Wells <wellsd2@rpi.edu>
Date: Sat, 2 Jul 2016 01:24:22 +0000 (-0400)
Subject: Add missing MPI function instantiations.
X-Git-Tag: v8.5.0-rc1~930^2
X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=6e68bf8dd51d5e9f310c056b681b44129cd642b7;p=dealii.git

Add missing MPI function instantiations.

The fixed-length array (i.e., things declared like T(&values)[N])
versions of MPI::sum, MPI::max, etc, all live in the header file mpi.h
since the length (N) is a compile-time constant. Those functions all
call Utilities::MPI::internal::all_reduce, which is declared in
mpi.templates.h, so that function (all_reduce) must be present in the
compiled library.

Note for the curious: it turns out that most compilers and most compiler
settings will not inline this function, meaning that it consistently
shows up in the shared object files without this explicit instantiation
in debug mode and usually in release mode. In particular, GCC 6 does not
inline it (so this instantiation is not necessary) but clang 3.8 does (so
this instantiation is necessary). Some versions of GCC (5.3 but not 4.8)
seem to inline it with the '-march=native' flag.
---

diff --git a/source/base/mpi.inst.in b/source/base/mpi.inst.in
index 19a06a0328..760a28fdac 100644
--- a/source/base/mpi.inst.in
+++ b/source/base/mpi.inst.in
@@ -36,6 +36,27 @@ for (S : MPI_SCALARS)
 
   template
   void min<S> (const std::vector<S> &, const MPI_Comm &, std::vector<S> &);
+
+  // The fixed-length array (i.e., things declared like T(&values)[N])
+  // versions of the functions above live in the header file mpi.h since the
+  // length (N) is a compile-time constant. Those functions all call
+  // Utilities::MPI::internal::all_reduce, which is declared in
+  // mpi.templates.h, so that function (all_reduce) must be present in the
+  // compiled library.
+  //
+  // Note for the curious: it turns out that most compilers and most compiler
+  // settings will not inline this function, meaning that it consistently
+  // shows up in the shared object files without this explicit instantiation
+  // in debug mode and usually in release mode. In particular, GCC 6 does not
+  // inline it (so this instantiation is not necessary) but clang 3.8 does (so
+  // this instantiation is necessary). Some versions of GCC (5.3 but not 4.8)
+  // seem to inline it with the '-march=native' flag.
+  template
+  void Utilities::MPI::internal::all_reduce<S> (const MPI_Op    &,
+                                                const S *const,
+                                                const MPI_Comm  &,
+                                                S               *,
+                                                const std::size_t);
 }