I patched OpenMPI 1.4.2 to fix some problems/warnings when using the
PGI compilers.
• openmpi-1.4.2/configure has the libtool patch to correctly parse PGI
compiler versions 1 through 5, but the libtool.m4 files do not. I
don't know if they are actually used -- I looked for this error before
I tried compiling. The fixes to the libtool.m4 files are:
# Fixes to correctly identify PGI compiler versions 1 through 5
mv openmpi-1.4.2/config/libtool.m4{,.original}
sed -e '5899s/\[\[1-5\]\]\*/\[\[1-5\]\].\*/g' \
openmpi-1.4.2/config/libtool.m4.original \
>openmpi-1.4.2/config/libtool.m4
mv openmpi-1.4.2/opal/libltdl/m4/libtool.m4{,.original}
sed -e '5899s/\[\[1-5\]\]\*/\[\[1-5\]\].\*/g' \
openmpi-1.4.2/opal/libltdl/m4/libtool.m4.original \
>openmpi-1.4.2/opal/libltdl/m4/libtool.m4
• Inline assembly of the atomic operators is explicitly disabled for
PGI C in openmpi-1.4.2/configure, but not for PGI C++. I applied the
same fix in openmpi-1.4.2/configure for PGI C++.
• openmpi-1.4.2/configure tests whether the C and C++ compilers accept
#pragma ident. However, PGI C does not perform macro substitution in
#pragma ident, which OpenMPI relies on. The fix is to use a more
complete test of the #pragma ident feature in openmpi-1.4.2/
configure. The fixes to openmpi-1.4.2/configure are:
# Disable inline assembly for PGI C++, as is done for PGI C (34579), and
# Fix PGI C compiler warning (10393, 26106): Pragma ignored – string
# expected after #pragma ident
mv openmpi-1.4.2/configure{,.original}
sed -e '34579{x;s/^.*$/if test "$ompi_cv_cxx_compiler_vendor" =
"portland group" ; then/;p;
s/^.*$/# PGI seems to have some issues with
our inline assembly./;p;
s/^.*$/# Disable for now./;p;
s/^.*$/asm_result="no (Portland Group)"/;p;
s/^.*$/else/;G;}' \
-e '34702{x;s/^.*$/fi/;G;}' \
-e '10393{s/#pragma ident/#define IDENT/;p;
s/^.*$/#pragma ident \$IDENT/;}' \
-e '26106{s/#pragma ident/#define IDENT/;p;
s/^.*$/#pragma ident \$IDENT/;}' \
openmpi-1.4.2/configure.original \
>openmpi-1.4.2/configure
chmod +x openmpi-1.4.2/configure
• PGI C issues spurious warnings when an array name is used as the
first operand in a conditional expression. This occurs in the
definition of the LT_STRLEN() macro in openmpi-1.4.2/opal/libltdl/
ltdl.h: #define LT_STRLEN(s) (((s) && (s)[0]) ? strlen (s) : 0). Even
though I think the PGI compiler is complaining too much, I also find
C's use of 0 to mean zero, false, and nil poor programming practice.
I changed the guard expression in the definition of LT_STRLEN to
simply (s!=NULL), which also eliminated the short-circuit test (s)
[0]. The extra test is not required for strlen() to succeed, and only
makes sense if zero-length strings are a common occurrence worth hand-
optimizing. The fix to openmpi-1.4.2/opal/libltdl/ltdl.h is:
# Fix PGI compiler warning: Array name used in logical expression
mv openmpi-1.4.2/opal/libltdl/ltdl.h{,.original}
sed -e '44s/((s) && (s)\[0\])/(s!=NULL)/' \
openmpi-1.4.2/opal/libltdl/ltdl.h.original \
>openmpi-1.4.2/opal/libltdl/ltdl.h
• PGI C issues spurious warnings when 0 is one of the operands in a
conditional expression of pointer type. Even though I think the PGI
compiler is again complaining too much, I replaced such occurrences of
0 with NULL, which is my preferred programming practice. The fixes to
the hooks.c and malloc.c files in openmpi-1.4.2/opal/mca/memory/
ptmalloc2 are:
# Fix PGI compiler warning: Pointer value created from a nonlong
integral type
mv openmpi-1.4.2/opal/mca/memory/ptmalloc2/hooks.c{,.original}
sed -e '434s/: 0;/: NULL;/' \
-e '449s/: 0;/: NULL;/' \
openmpi-1.4.2/opal/mca/memory/ptmalloc2/hooks.c.original \
>openmpi-1.4.2/opal/mca/memory/ptmalloc2/hooks.c
mv openmpi-1.4.2/opal/mca/memory/ptmalloc2/malloc.c{,.original}
sed -e '3446s/: 0,/: NULL,/' \
-e '3664s/: 0,/: NULL,/' \
-e '3789s/: 0,/: NULL,/' \
openmpi-1.4.2/opal/mca/memory/ptmalloc2/malloc.c.original \
>openmpi-1.4.2/opal/mca/memory/ptmalloc2/malloc.c
• The only remaining complaint I get now from the PGI C compiler is
for orte_base_log() in openmpi-1.4.2/orte/mca/notifier/base/
notifier_base_open.c:
static void orte_base_log(int priority, const char *msg, ...)
{
/* just do nothing - it is here just so
* someone calling it won't segv
*/
}
PGI C issues the following warning:
PGC-W-0155-No va_start() seen (base/notifier_base_open.c: 85)
This is completely legal C, as far as I can tell from my copy of the
1990 ISO C standard. I didn't bother to fix this; I sent a bug report
to PGI.
Larry Baker
US Geological Survey
650-329-5608
ba...@usgs.gov
