Is there a way to make ml64 or some assembler accept ifort .asm files? ml64 rejects the code labels beginning with '.'.
I'm looking for a way to test the "ALIGN 16" code alignment directives which come and go with various ifort releases, particularly when AVX2 is set.
A related confusing issue comes up when setting AVX2 reduces the unrolling. Opt-report (and Advisor) don't report unrolling when it is not done consistent with Qunroll: or setting.
When I run VTune or Advisor on ifort generated .exe, sometimes where ifort decides to align, the alignments come out different by up to 3 bytes. This doesn't appear to hurt, even when the loop is placed 3 bytes ahead of an aligned address.
In a few cases where gfortran decides to align code, but ifort doesn't, there may be a 30% speedup with alignment on my Haswell. I'm guessing it makes Loop Stream Detection more effective.
