Fix Fortran preprocessor issue with CPPFLAGS #9899

jjhursey · 2022-01-20T16:57:07Z

Some C and Fortran compilers use different preprocessors. If one preprocessor
accepts -iquote and the other does not then a compiler error will occur
when Open MPI tries to use it.
- Nvidia/PGI v22.1-0 is one such. The C compiler supports -iquote
  while the Fortran compiler does not.
Similar to PR Fortran fixes #7265 we need to clear the CPPFLAGS and AM_CPPFLAGS

Signed-off-by: Joshua Hursey [email protected]

jjhursey · 2022-01-20T18:42:27Z

bot:aws:retest

jjhursey · 2022-01-20T18:48:48Z

Example output with Nvidia/PGI compiler v22.1-0:

checking for pgcc option to add a directory only to the search path for the quote form of include... "-iquote"
checking for pgfortran option to see if it also supports -iquote form of include... No. "-I" will be used instead

Example with gcc10:

checking for gcc-10 option to add a directory only to the search path for the quote form of include... "-iquote"
checking for gfortran-10 option to see if it also supports -iquote form of include... Yes. "-iquote" accepted

This will need to be cherry-picked to the v5.0.x branch.
The v4.0.x and v4.1.x branches do not have this logic. Additionally, PRTE and OpenPMIx do not need this fix even though they have the -iquote logic they do contain any Fortran code so are not impacted.

hjelmn · 2022-01-20T21:30:22Z

Does this skip the check for the --disable-mpi-fortran case? Sometimes I have a fortran compiler on the system but will never use it.

jjhursey · 2022-01-20T22:20:24Z

It looks like if you pass --disable-mpi-fortran it still runs the check with the fortran compiler. I'll work on adding an exception.

jjhursey · 2022-01-21T15:23:35Z

I just pushed a fix for the --disable-mpi-fortran case where it will skip the check against the Fortran preprocessor.

 ./configure --prefix=/workspace/exports/ompi CC=pgcc CXX=pgc++ FC=pgfortran --disable-cuda --disable-nvml --with-cuda=no --disable-mpi-fortran
...
checking for pgcc option to add a directory only to the search path for the quote form of include... "-iquote"
checking for the C compiler vendor... portland group
...
Open MPI configuration:
-----------------------
Version: 5.1.0a1
Build MPI C bindings: yes
Build MPI Fortran bindings: no
Build MPI Java bindings (experimental): no
Build Open SHMEM support: false (no spml)
Debug build: no
Platform file: (none)
...

jjhursey · 2022-01-21T15:24:30Z

Note that once this fix gets into master and v5 I can update the IBM CI build boxes to the latest Nvidia/PGI compilers. Without this fix they fail to build per PR #9716

jsquyres · 2022-01-23T16:28:45Z

This looks tangentially related to #9890 in that it's an issue about the C and Fortran compiler flags.

jsquyres · 2022-01-23T16:39:40Z

@jjhursey This feels very much like #7253 (issue) and #7265 (PR to fix it). E.g., in the commit message for ab398f4, I stated:

Automake's Fortran compilation rules inexplicably use CPPFLAGS and AM_CPPFLAGS.

Did we miss adding something like https://github.com/open-mpi/ompi/blob/master/ompi/mpi/fortran/use-mpi-f08/Makefile.am#L29-L34 somewhere? Can you cite the specific problem that you're solving -- e.g., are you seeing CPPFLAGS being used somewhere in the code base to compile Fortran code?

jjhursey · 2022-01-24T15:40:38Z

The error is:

make[2]: Entering directory '/opt/mpi_src/ompi/mpiext/ftmpi/use-mpi-f08'
  PPFC     comm_revoke_f08.lo
pgfortran-Error-Unknown switch: -iquote../../../..
make[2]: *** [Makefile:1464: comm_revoke_f08.lo] Error 1
make[2]: Leaving directory '/opt/mpi_src/ompi/mpiext/ftmpi/use-mpi-f08'
make[1]: *** [Makefile:2711: all-recursive] Error 1
make[1]: Leaving directory '/opt/mpi_src/ompi'
make: *** [Makefile:1477: all-recursive] Error 1

A bit more detail:

shell$ make V=1
/bin/sh ../../../../libtool  --tag=FC   --mode=compile pgfortran -DHAVE_CONFIG_H -I. -I../../../../opal/include -I../../../../ompi/include -I../../../../oshmem/include -I../../../../ompi/mpiext/cuda/c   -iquote../../../..  -I/opt/mpi_src/3rd-party/libevent-2.1.12-stable -I/opt/mpi_src/3rd-party/libevent-2.1.12-stable/include -I/opt/mpi_src/3rd-party/hwloc-2.7.0/include -I/opt/mpi_src/3rd-party/hwloc-2.7.0/include -I/opt/mpi_src/3rd-party/openpmix/include -I/opt/mpi_src/3rd-party/openpmix/include -I../../../../ompi/mpi/fortran/use-mpi -I../../../../ompi/mpi/fortran/use-mpi-f08/mod -I../../../.. -I../../../..   -c -o comm_revoke_f08.lo comm_revoke_f08.F90
libtool: compile:  pgfortran -DHAVE_CONFIG_H -I. -I../../../../opal/include -I../../../../ompi/include -I../../../../oshmem/include -I../../../../ompi/mpiext/cuda/c -iquote../../../.. -I/opt/mpi_src/3rd-party/libevent-2.1.12-stable -I/opt/mpi_src/3rd-party/libevent-2.1.12-stable/include -I/opt/mpi_src/3rd-party/hwloc-2.7.0/include -I/opt/mpi_src/3rd-party/hwloc-2.7.0/include -I/opt/mpi_src/3rd-party/openpmix/include -I/opt/mpi_src/3rd-party/openpmix/include -I../../../../ompi/mpi/fortran/use-mpi -I../../../../ompi/mpi/fortran/use-mpi-f08/mod -I../../../.. -I../../../.. -c comm_revoke_f08.F90  -fpic -o .libs/comm_revoke_f08.o
pgfortran-Error-Unknown switch: -iquote../../../..
make: *** [Makefile:1464: comm_revoke_f08.lo] Error 1

Is the solution roughly this to clear the CPPFLAGS and AM_CPPFLAGS flags before building this section of code? Is that preferred over the configure check in this PR?

jjhursey · 2022-01-24T16:20:25Z

I tried the alternative approach and it worked fine.

diff --git a/ompi/mpiext/ftmpi/use-mpi-f08/Makefile.am b/ompi/mpiext/ftmpi/use-mpi-f08/Makefile.am
index d06c0aa6..85d1b778 100644
--- a/ompi/mpiext/ftmpi/use-mpi-f08/Makefile.am
+++ b/ompi/mpiext/ftmpi/use-mpi-f08/Makefile.am
@@ -13,6 +13,14 @@
 # $HEADER$
 #
 
+# Note that Automake's Fortran-buidling rules uses CPPFLAGS and
+# AM_CPPFLAGS.  This can cause weirdness (e.g.,
+# https://github.com/open-mpi/ompi/issues/7253 and
+# https://github.com/open-mpi/ompi/issues/9716).  Let's just zero
+# those out and rely on AM_FCFLAGS.
+CPPFLAGS =
+AM_CPPFLAGS =
+
 # This file builds the use_mpi_f08-based bindings for MPI extensions.  It
 # is optional in MPI extensions.

I went ahead and added the change above to this PR as a second commit. That'll make it easier to pick the one the community wants.

I'm happy with either or both. I'm just looking for some guidance from the community on their preference. Let me know and I'll update the PR.

bwbarrett · 2022-01-24T16:37:10Z

I'd prefer the second approach (mucking with CPPFLAGS) for two reasons: 1) we already do it elsewhere and 2) I'm worried about the corner cases that led to -iquote in the first place biting us in the butt again.

jsquyres · 2022-01-24T17:25:56Z

@jjhursey While you have the hood up, could you also add the same fix to ompi/mpiext/example/use-mpi-f08/Makefile.am? I'm guessing that we ended up in this situation because the good ftmpi folks copied from the example, and we didn't fix the example back in #7265.

jjhursey · 2022-01-24T17:28:54Z

Sure. I'll make that adjustment and update this PR in a bit.

* Some C and Fortran compilers use different preprocessors. If one preprocessor accepts `-iquote` and the other does not then a compiler error will occur when Open MPI tries to use it. - Nvidia/PGI v22.1-0 is one such. The C compiler supports `-iquote` while the Fortran compiler does not. * Similar to PR open-mpi#7265 we need to clear the `CPPFLAGS` and `AM_CPPFLAGS` Signed-off-by: Joshua Hursey <[email protected]>

jjhursey · 2022-01-24T17:33:22Z

@jsquyres @bwbarrett I made those changes. This is ready for re-review. Thanks!

jjhursey requested review from jsquyres, hjelmn and bwbarrett January 20, 2022 16:57

jjhursey linked an issue Jan 20, 2022 that may be closed by this pull request

nvfortran-Error-Unknown switch: -iquote../../../../../.. #9716

Closed

jjhursey added the ⚠️ WIP-DNM! label Jan 20, 2022

jjhursey force-pushed the fix-iquote-check branch from 0a3ea64 to 4f3729c Compare January 21, 2022 15:22

jjhursey removed the ⚠️ WIP-DNM! label Jan 21, 2022

jjhursey force-pushed the fix-iquote-check branch from 0a6b56a to bbe5788 Compare January 24, 2022 17:30

jjhursey changed the title ~~Fix iquote check for different C and Fotran preprocessors~~ Fix Fortran preprocessor issue with CPPFLAGS Jan 24, 2022

jsquyres approved these changes Jan 24, 2022

View reviewed changes

bwbarrett approved these changes Jan 24, 2022

View reviewed changes

jjhursey merged commit fcc41d7 into open-mpi:master Jan 24, 2022

jjhursey deleted the fix-iquote-check branch January 24, 2022 20:36

jjhursey mentioned this pull request Jan 24, 2022

nvfortran-Error-Unknown switch: -iquote../../../../../.. #9716

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Fix Fortran preprocessor issue with CPPFLAGS #9899

Fix Fortran preprocessor issue with CPPFLAGS #9899

Uh oh!

jjhursey commented Jan 20, 2022 •

edited

Loading

Uh oh!

jjhursey commented Jan 20, 2022

Uh oh!

jjhursey commented Jan 20, 2022 •

edited

Loading

Uh oh!

hjelmn commented Jan 20, 2022

Uh oh!

jjhursey commented Jan 20, 2022

Uh oh!

jjhursey commented Jan 21, 2022

Uh oh!

jjhursey commented Jan 21, 2022

Uh oh!

jsquyres commented Jan 23, 2022

Uh oh!

jsquyres commented Jan 23, 2022 •

edited

Loading

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

bwbarrett commented Jan 24, 2022

Uh oh!

jsquyres commented Jan 24, 2022

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

Uh oh!

Fix Fortran preprocessor issue with CPPFLAGS #9899

Fix Fortran preprocessor issue with CPPFLAGS #9899

Uh oh!

Conversation

jjhursey commented Jan 20, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jjhursey commented Jan 20, 2022

Uh oh!

jjhursey commented Jan 20, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hjelmn commented Jan 20, 2022

Uh oh!

jjhursey commented Jan 20, 2022

Uh oh!

jjhursey commented Jan 21, 2022

Uh oh!

jjhursey commented Jan 21, 2022

Uh oh!

jsquyres commented Jan 23, 2022

Uh oh!

jsquyres commented Jan 23, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

bwbarrett commented Jan 24, 2022

Uh oh!

jsquyres commented Jan 24, 2022

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

jjhursey commented Jan 24, 2022

Uh oh!

Uh oh!

jjhursey commented Jan 20, 2022 •

edited

Loading

jjhursey commented Jan 20, 2022 •

edited

Loading

jsquyres commented Jan 23, 2022 •

edited

Loading