hi @klausler ran into a regression in our hpc2021 testing, bisects to this PR.
if you have access to the suite, should be reproducible with any of the MPI, OMP, or TGT builds
Failure seen in 519, 528.,535
if you need a smaller reproducer , i will try and produce one

Thread 1 "weather_base.cl" received signal SIGSEGV, Segmentation fault.
0x00007ffff74a547e in __GI___libc_free (mem=0x1) at ./malloc/malloc.c:3368
warning: 3368 ./malloc/malloc.c: No such file or directory
(gdb) bt
#0 0x00007ffff74a547e in __GI___libc_free (mem=0x1) at ./malloc/malloc.c:3368
#1 0x00007ffff7e406b0 in Fortran::runtime::io::UnitMap::CloseAll(Fortran::runtime::io::IoErrorHandler&) ()
from /opt/openmpi-5.0.6-flang-new/lib/libmpi_usempif08.so.40
#2 0x00007ffff7e3a0db in Fortran::runtime::io::ExternalFileUnit::CloseAll(Fortran::runtime::io::IoErrorHandler&) ()
from /opt/openmpi-5.0.6-flang-new/lib/libmpi_usempif08.so.40
#3 0x0000000000258c96 in CloseAllExternalUnits(char const*) ()
#4 0x00000000002590fd in _FortranAProgramEndStatement ()
#5 0x0000000000228012 in main () at miniWeather.fppized.f90:894
#6 0x00007ffff7429d90 in __libc_start_call_main (main=main@entry=0x228000

small reproducer, requires an openmpi compiler invocation, so some kind of interaction
$ cat goo.f90

                dt =1.0
                write(*,*) 'dt: ',dt
               end

$ OMPI_FC=/COD/2025-07-26/trunk/bin/flang /opt/openmpi-5.0.6-flang-new/bin/mpif90 -O2 goo.f90 -g
$ ./a.out
dt: 1.
Segmentation fault (core dumped)

Works fine for me here. Could you try it with valgrind in your environment?

(This patch affects the input side of user-defined derived type I/O, so perhaps it exposes something in your setup, but I don't see anything in it that would be at risk of smashing memory in the runtime support library's open unit map, and your stack traceback looks like that's what led to the crash.)

==2250846== Conditional jump or move depends on uninitialised value(s)
==2250846== at 0x208772: Fortran::runtime::io::IoStatementState& Fortran::runtime::io::ExternalFileUnit::BeginIoStatement<Fortran::runtime::io::ExternalListIoStatementState<(Fortran::runtime::io::Direction)0>, Fortran::runtime::io::ExternalFileUnit&, char const*&, int&>(Fortran::runtime::Terminator const&, Fortran::runtime::io::ExternalFileUnit&, char const*&, int&) (in ./a.out)

Thanks for that. I'm not seeing it here with valgrind myself, using a flang-rt built with gcc 9.3.0 for x86-64. What are you building with? (It might be a red herring.)

i think its a mismatch of using an openmpi build that is slightly out of date with respect to latest flang sources. Just now I rebuilt openmpi using todays local build of llvm/main . test passes and valgrind errors no longer present.

i think we can close this out as stale build.

I'm happy if you're happy. Thanks for the help.