Mark mpif.h as deprecated

[jeffhammond]

I think it would be neat if MPICH also warned users that mpif.h is deprecated when the Fortran compiler supports it.

open-mpi/ompi#13123

[hzhou]

I don't think new code include mpif.h. It is likely legacy code a graduate student inherits from his professor that passed down from past students who no longer work in the area. Issuing such warnings are just nagging for no real benefit.

A better idea is to make newer MPI features not available via mpif.h. That plus if the forum can add real compelling new features that applications have real motivations to adopt, then eventually we will see fewer and fewer code using mpif.h.

[jeffhammond]

Some of the top HPC apps in the world under active development use mpif.h because they see no reason to change. They are unaware that it is deprecated and may be deleted some day.

We want to warn existing users of mpif.h to stop using it and switch at least to use MPI.

Many HPC codes are still using only MPI-1 features because that's all the math requires. I don't see "make the standard more interesting" as a useful solution to mpif.h usage.

[hzhou]

If mpif.h works for them, why do we want to nag them and threaten them? mpif.h is simpler than modules and if they understand and are comfortable with the C/Fortran linking works.

[hzhou]

[hzhou]

To clarify, I am not against it. I am just not motivated to it. And I think with OpenMPI nagging about it, it is already a good enough dissuasion.

[jeffhammond]

If mpif.h works for them, why do we want to nag them and threaten them? mpif.h is simpler than modules and if they understand and are comfortable with the C/Fortran linking works.

If K&R C or C89 work for people, why do GCC and Clang nag and threaten them?

To clarify, I am not against it. I am just not motivated to it. And I think with OpenMPI nagging about it, it is already a good enough dissuasion.

I agree. I just thought the implementation of this is trivial and having both major MPIs do it will accelerate progress in MPI Fortran usage.

[hzhou]

If mpif.h works for them, why do we want to nag them and threaten them? mpif.h is simpler than modules and if they understand and are comfortable with the C/Fortran linking works.

If K&R C or C89 work for people, why do GCC and Clang nag and threaten them?

Really? When did they started the nag? But people hated C89 in the light of C99 since 1999 to the point that I don't think many even noticed the nagging, for example, me. Which brings back to my point, if the new features are desirable enough, old code will slowly dissipate, until a point, a final kick of nagging will nail it.

Modern Fortran is not like C99. It fixes some sore spots of F77, but it also directly leaps to the future where not many are ready for. It's like, try converting people from C89 to C++

[hzhou]

I guess the conclusion is we will watch Open MPI's venture and kick the can down until next time this issue is raised again.

[jeffhammond]

If mpif.h works for them, why do we want to nag them and threaten them? mpif.h is simpler than modules and if they understand and are comfortable with the C/Fortran linking works.

If K&R C or C89 work for people, why do GCC and Clang nag and threaten them?

Really? When did they started the nag? But people hated C89 in the light of C99 since 1999 to the point that I don't think many even noticed the nagging, for example, me. Which brings back to my point, if the new features are desirable enough, old code will slowly dissipate, until a point, a final kick of nagging will nail it.

nwchemgit/nwchem#1082, among other examples. Enforcement of modern C standards broke builds all over the place, and people have been forced to fix them.

Modern Fortran is not like C99. It fixes some sore spots of F77, but it also directly leaps to the future where not many are ready for. It's like, try converting people from C89 to C++

mpif.h is not compliant with any version of the Fortran standard, which is but one of the reasons we deprecated it.

use MPI isn't either, but at least it supports partial type-safety.

[hzhou]

nwchemgit/nwchem#1082, among other examples. Enforcement of modern C standards broke builds all over the place, and people have been forced to fix them.

Your example is not good. gcc-14/15 are very recent -- but I did notice they become much more aggressive in warnings and errors recently. For 2, -std= is asking for pedantic mode.

Also quoting K&R C and C89 together is devious. I haven't encountered people using K&R C since I learned programming. C89 on the hand, is still the norm when I joined MPICH, and not many folks are annoyed by C89. I was the one could not tolerate C89 and advocated to move to C99.

[jeffhammond]

nwchemgit/nwchem#1082, among other examples. Enforcement of modern C standards broke builds all over the place, and people have been forced to fix them.

Your example is not good. gcc-14/15 are very recent -- but I did notice they become much more aggressive in warnings and errors recently. For 2, -std= is asking for pedantic mode.

We have gone through the exact same process over and over again for the past 15 years, as GCC and other compilers started enforcing C99 requirements for having function signatures declared. I merely forwarded you the latest example that hit my inbox, because I do not think it is necessary to have to document everything I know about legacy C code and compiler warnings, because you can easily verify this yourself by taking code from the K&R C book and compiling it, or by doing the same with C89 code.

Also quoting K&R C and C89 together is devious. I haven't encountered people using K&R C since I learned programming. C89 on the hand, is still the norm when I joined MPICH, and not many folks are annoyed by C89. I was the one could not tolerate C89 and advocated to move to C99.

K&R C is literally used right now in NWChem: https://github.com/GlobalArrays/ga/blob/develop/ma/ma.c#L588. I don't see why you accuse me of being devious just because you are unaware of things that are happening in other parts of the HPC world.

[hzhou]

K&R C is literally used right now in NWChem: https://github.com/GlobalArrays/ga/blob/develop/ma/ma.c#L588. I don't see why you accuse me of being devious just because you are unaware of things that are happening in other parts of the HPC world.

No no, I am not pointing fingers at all but try to hear the story. What is the NWChem story that K&R C code still lives today despite repeated nagging each year? 😃

PS: I apologize for the choice of word, meant jokingly.