GNU v12 Debug runs are crashing

[mathomp4]

After the last tranche of V12 updates, my GNU Debug nightly builds fails at run time with a floating overflow in the Moist process library (via evap3 it seems):

1 0x00000000033eb0e8 __geosmoist_process_library_MOD_ldradius4()  /discover/nobackup/mathomp4/SystemTests/builds/AGCM_V12GNU/CURRENT/GEOSgcm/src/Components/@GEOSgcm_GridComp/GEOSagcm_GridComp/GEOSphysics_GridComp/GEOSmoist_GridComp/Process_Library.F90:758
2 0x00000000033eba59 __geosmoist_process_library_MOD_evap3()  /discover/nobackup/mathomp4/SystemTests/builds/AGCM_V12GNU/CURRENT/GEOSgcm/src/Components/@GEOSgcm_GridComp/GEOSagcm_GridComp/GEOSphysics_GridComp/GEOSmoist_GridComp/Process_Library.F90:643
3 0x000000000381731a __geos_gfdl_1m_interfacemod_MOD_gfdl_1m_run()  /discover/nobackup/mathomp4/SystemTests/builds/AGCM_V12GNU/CURRENT/GEOSgcm/src/Components/@GEOSgcm_GridComp/GEOSagcm_GridComp/GEOSphysics_GridComp/GEOSmoist_GridComp/GEOS_GFDL_1M_InterfaceMod.F90:747

I did some prints and at least one point had:

LDRADIUS4 inputs: RHO=   1.20311737      QC=   1.06696154E+36

which was then used in:

          !- liquid water content
          WC = 1.e3*RHO*QC ! air density [g/m3] * liquid cloud mixing ratio [kg/kg]

and I think it then overflowed a 32-bit real.

Now the latest changes in v12 GEOSgcm_GridComp as a whole are tiny:

https://github.com/GEOS-ESM/GEOSgcm_GridComp/compare/GCMv12-rc22...GCMv12-rc23?w=1

and only the gfdl_mp one seems like it could do anything here.

I tried adding (back) qa (i, k) = qaz (k) in the if (.not. do_qa) then part of the code in gfdl_mp.F90. But it is also non-zero-diff if I do that.

Before your fix, qa was always set to qaz so that seems to make GNU happy. But well, since we are in a .not. do_qa block, we aren't probably meant to do qa!

cc: @sdrabenh @wmputman

[mathomp4]

Apparently fixed now. Closing.