Several of the FCVS tests in the llvm-test-suite are expecting REALs to be printed with exactly 8 or 16 decimals, including trailing zeroes. For example, this WRITE is expected to produce this result. I think this is based on what gfortran produces. However, flang is currently skipping the trailing zeroes (so in that example it prints only
2.5 on my AArch64 Ubuntu machine).
IIUC the width and precision for the list-directed IO is processor dependent, so we can expect some reasonable variation here (the llvm-test-suite is designed to run with any compiler, not just flang). My proposal would be to change all the tests to use a fixed width format (i.e. replace
* with some
Fw.d or another, depending on the precision of the value being printed). Does this sound like the right thing to do?
Thanks in advance!