Use DWARF unwinding with Clang to avoid overproduction of linking warnings with clang + gfortran

[GiudGiud]

refs idaholab/moose#29009

[jwpeterson]

I don't have any issues with this, but I also don't use clang on Darwin so I don't have an easy way of testing it. @roystgnr should we just merge this and propagate the changes to libmesh to make it easier for @GiudGiud to actually test it in MOOSE?

[roystgnr]

@GiudGiud reports verifying that this fixed his warnings for clang on Darwin. I'm just worried that it might cause problems on other configurations. We're passing linker options here based on autodetection of compiler type, right? But I believe it's common for a system to combine LLVM compilers (which would cause us to detect clang) with non-LLVM linkers (which might scream and die if they don't recognize -keep_dwarf_unwind -no_compact_unwind).

And ... I was going to test my theory manually, but it looks like that's what already happened here: https://civet.inl.gov/job/2659310/

The configure script misdiagnoses a failure to link to libhdf5 as "HDF5 not found", but if you look at the immediate failure then it's clearer:

/usr/bin/ld: unrecognized option '-keep_dwarf_unwind'
/usr/bin/ld: use the --help option for usage information
clang-15: error: linker command failed with exit code 1 (use -v to see invocation)

So we can't safely set these flags unless we verify them first at configure time, e.g. by temporarily setting them in LDFLAGS and using AC_TRY_LINK on a "Hello world" program to make sure they're usable.

[jwpeterson]

Yes, definitely sounds like an AC_LINK_IFELSE (I think that's the preferred one) type test is called for here.

[GiudGiud]

something like that?

[GiudGiud]

Should I add a message in the "yes" case too?

[roystgnr]

I'd think we only want a message in the yes case, personally. If we don't add any flags, maybe there's a problem, but "your environment has problems unless we carefully find the right flags to overcome them" sounds like a "your environment" problem we don't need to bug the user about rather than a "we" problem we do. But if we do add flags, even with the best of intentions, ... it's maybe not morally required to talk about what we're doing, but it's at least polite. Something like "retaining dwarf unwinding for gfortran compatibility" that would fit on one line, leaving the more detailed explanations for the .m4 comment, maybe?

[jwpeterson]

I think the way AC_LINK_IFELSE works is that you first set the LIBS variable to the flags you want to try, then you call AC_LINK_IFELSE and if it passes you enable the flags in the libmesh LDFLAGS, otherwise you don't. There's an example in xdr.m4:

AC_DEFUN([TEST_XDR],
[
  old_CPPFLAGS="$CPPFLAGS"
  old_LIBS="$LIBS"
  CPPFLAGS="$CPPFLAGS $XDRINCLUDES"
  LIBS="$LIBS $XDRLINKLIBS"

  AC_LINK_IFELSE([AC_LANG_PROGRAM([@%:@include <stdio.h>
                                   @%:@include <rpc/rpc.h>
                                   @%:@include <rpc/xdr.h>],
                                  [
                                    XDR * xdr;
                                    FILE * fp;
                                    xdrstdio_create(xdr, fp, XDR_ENCODE);
                                  ])],
                 [
                   AC_MSG_RESULT(yes)
                   AC_DEFINE(HAVE_XDR, 1, [Flag indicating headers and libraries for XDR IO are available])
                   enablexdr=yes
                 ],
                 [
                   AC_MSG_RESULT(no)
                   enablexdr=no
                 ])

  dnl Reset flags after testing
  CPPFLAGS="$old_CPPFLAGS"
  LIBS="$old_LIBS"
])

[GiudGiud]

I think the way AC_LINK_IFELSE works is that you first set the LIBS variable to the flags you want to try, then you call AC_LINK_IFELSE and if it passes you enable the flags in the libmesh LDFLAGS, otherwise you don't. There's an example in xdr.m4:

I think I have something similar to the example now

[jwpeterson]

As I understand it, AC_LINK_IFELSE specifically uses the variables LIBS and LDFLAGS to test linking:

https://www.gnu.org/software/autoconf/manual/autoconf-2.69/html_node/Running-the-Linker.html

So I think you need to do like in the xdr example, and set the LIBS variable prior to calling AC_LINK_IFELSE. Also, it seems to me that there is an issue with your test program, it will need to #include <iostream> in order to be well-formed.

PR not working yet