1) I know of no thread mentioning evals that has not got Richard on his soapbox.
Richard: for serious eval you 'borrow' a full version from the toolmaker the 'eval' version is for those that are "marginal customers" and what I see, today, mainly used for inclusion in chip demo packages.

As far as I know, there are no 8051s that use caches.
incorrect, the SILabs '51s do

All the same, most efficiency is gained by improvements in the algorithm rather than the low level ASM code.
absolutrly

Of course, you will get the best performance from a perfectly designed ASM program. Very few applications are that time critical. If your clients have few boards, your development time is very significant.
the issue is that you MAY need to "get the best performance from a perfectly designed ASM" from some little loop, never from the whole shebang. "Crafting" e.g. a user interface (timing = human reaction) is ridiculous.

I am sure that you can knock out perfect ASM faster and more reliably than others can use a HLL. This is due to your experience. At your age, I expect that PDIP-40 8051 chips will see you through. A young student would better spend her time learning how to use MCUs and tools that have a realistic future.
I feel for the "young student" in my time the chips have gotten smaller synchrounous with my eyes getting weaker and the "young students" starting point does not bode well for his abilities to scope when he get older.

Incidentally, when you meet your bus or jealous husband, will your clients really be able to maintain your ASM code? Or even find another expert?
this is the REAL issue, one client of mine chose to use a more expensive chip with the remark "this little bit of ASM is a great ides, but I need to be sure someone can maintain the thing in the future, so use C only"

Erik