Potential issue with Where statements in debug mode

[JosephMouallem]

Hi all,

I am encountering issues with where statements when running SHiELD/MOM6 in debug mode.
For example, the model performs a division by zero and crashes in xgrid.F90 when computing the inverse area, where grid%area>0 so not sure how the division by zero is happening in this context:

FMS/exchange/xgrid.F90

Lines 1973 to 1976 in 23df6d7

	if(grid%on_this_pe) then
	grid%area_inv = 0.0_r8_kind;
	where (grid%area>0.0_r8_kind) grid%area_inv = 1.0_r8_kind/grid%area
	endif

A similar issue arises in the monin obukhov routine:

FMS/monin_obukhov/include/monin_obukhov_inter.inc

Lines 248 to 253 in 23df6d7

	where (mask_1)
	rzeta = 1.0_lkind/zeta
	zeta_0 = zeta/z_z0
	zeta_t = zeta/z_zt
	zeta_q = zeta/z_zq
	end where

I am using the standard SHiELD template and compile flags: https://github.com/NOAA-GFDL/SHiELD_build/blob/main/site/intel.mk.

These issues are specific to DEBUG mode; everything works fine when compiling in REPRO mode.

[bensonr]

There are some nuances with respect to where statements and the order things are evaluated leading to FPEs when you would expect to be protected against them. Whether you hit them or not is completely up to how the compiler is implemented.

[J-Lentz]

Closing because this isn't an issue with FMS per se.

Floating point exceptions are enabled by the main program due to -fpe0 and -ftrapuv in the debug-mode mkmf flags, while optimization of FMS code causes speculative execution of divide-by-zero cases in the where statement without regard for the mask condition. Possible solutions are to add -fp-speculation=safe to the FMS build flags or to build FMS in debug mode when linking it with a debug-mode main program.