The problem here is much more the unix wars and a lack of confidence in BSD under legal fire rather than a lack of ability. The principal concern of unix vendors of the early PC era was to maintain their market share in mini and mainframe product sectors rather than growth into the consumer market. This spurred a rewrite of BSD fragments tied to the legacy unix codebase which fully portablized C and the GCC downstream projects which ended up benefiting the weird hobby OS linux disproportionately, and had it not had to be written from scratch we may have ended up with a wonderful 286-BSD rather than a 486-BSD, which at the time was still not fully clean room foss and unburdened. This was a time when large customers of OS products were trying to squeeze all the performance juice out of the existing systems instead of looking at new paradigms. We have things like the full SAST and warn-free release of sunOS around this time, where Sun was focused on getting a rock stable platform to then optimize around rather than efforts to produce products for the emerging Micro market. We can see that the concept of a portable unix system and c library as early as Xenix on the Apple Lisa in 1984. That's only 3 short years after the IBM collaboration for PC-DOS, showing even a rookie uncoordinated and low technical skill team such as microsoft (Paraphrasing Dave Cutler, chief NT kernel lead - Zachary, G. Pascal (2014). Showstopper!: The Breakneck Race to Create Windows NT and the Next Generation at Microsoft. Open Road Media. ISBN 978-1-4804-9484-8).
Xenix was my introduction to UNIX, I wouldn't claim it would win any performance price, specially when considering graphics programming.
Also my first C book was "A book on C", which had a type in listing for RatC dialect, like many others in those early 1980's, which were nothing more than a plain macro assembler without opcodes, for all practical purposes.
Compiler optimizations in those 8 and 16 bit compilers were what someone nowadays would do in a introduction to compilers, as the bare minimum, like constant propagation and peephole optimizations.
