D Drmmr, i did run optimized code. however, there could be problems w/ smarty modes of compilers because algo has used rXX, xmmXX & mmX registers as global vars. nevertheless nothing prevents us to run tests in separate files (one for fsort(no-if) & second for others).

That's right and that's why I'm not agreeing with you. You're claiming that application programmers should start applying hardware specific micro-optimizations throughout and that's more like a suicide pill.

Of course if you're a systems programmer you should apply every trick in the book to make it run as fast as possible on the target system/hardware. I don't disagree with that. I'm not even saying that programmers should never micro-optimize. Certainly they should occasionally but not routinely.

And of course there's room for improvements on systems software and tools like the article you referrenced suggests.
Razzle, i have no suggested to micro-code throughout whole thing. each algo has critical parts which are suitable to gain the significant speed-up. For instance, if block in algo consumes only 0.1% of overall time -- the're zeroed reasons to spend any efforts in the rest of cases.

Furthermore, the article specifically points out the challenges posed by the multiple cores of todays CPUs and GPUs but nowhere does it say that "modern software engineering practices" should be abandoned like you suggest.
Author just has expressed to us his wishes/dreams/hopes/etc-etc. In fact, we've no had even something a little resembling towards those magic tools.