[flang] performance issue with 108.leslie3d benchmark

[shivaramaarao]

llvm-flang is taking 50% more time compared to gfortran/classic flang for 108.leslie3d in mpi2007 suite in x86. Most of the time spent seem to be in communication as per report. Anyone looked into this benchmark and root caused what might be the issue in code generation?.

On AMD Genoa platform with 192 ranks

classic flang - 80 seconds
llvm flang - 120 seconds

if we use jemalloc, the runtime reduces to 100 seconds. So, wondering if these are due to any memory allocation differences compared to other compilers.