This presentation will compare the performance of SPEC CPU® 2017 Fortran benchmarks, compiled using Flang versus GFortran, on AArch64. Then, for a benchmark exhibiting one of the largest slowdowns when compiled with Flang, it will be shown how it was profiled and how a particularly slow function was identified. Next, it will be presented how part of it was optimized, with a relatively simple MLIR transformation, by flattening multidimensional arrays, in the HLFIR dialect.