Developer of a research CFD code. Teaching Fortran programming for students of meteorology.
From my fans:
Many comments have been made that are misleading, manipulative, and in some cases outright falsehoods by one of the commenter's: Vladimir. My primary interest here is to shed light on those egregious items that may cause the casual reader to accept his comments, and end up in a bad way. It is a lesser but legitimate interest to set the record straight in relation to the abuses and falsehoods lavished upon me by Vladimir, and which also speak to allowing others to be misled by his suspicious items.
Dear Vladimir: It seems a waste of time to converse with you, so this and the comment below will be my last to you. CLEARLY, my results above DO include the "Do" version also, which is identically copied from the original posting. THAT is why there are two sets of data points in each chart (a "Do" and an SAS), and in the discussion. So, no idea what you are taking about. Also, as explained repeatedly, using just TWO compiler settings is sufficient to PROVE my point, no need for a tome ... so, no idea what you are up to, mate. You were wrong, be a man about it.
As for your "answer" below: A lot of "hand waving", but no actual results. Why not just report your timing results instead of all the "loop unrolling distraction"? IT is a certainty, as proven above, the "Do" executes slower, and very much slower for low compiler opt. You may "believe" anything you like, but science is about data and the "what it is", not the "what you wish it would be". I also checked some of your other bits, and there is pattern to you "imagined" facts ... hey, long live narcissism, what a shame for SO.