I have some codes that I wrote 20 years ago or more, and performance wise for the most part I would rewrite them the same way as of today.
The remaining tiny part is because I am (hopefully) more skilled as a developer (although it’s not my primary job) than I was 20 years ago, and maybe there are some parts that I would write differently. But it has nothing to do with language features or whatever, a more skilled developer at that time would have written a code that would be still a very good one today.
Computers are much more powerful than they used to be, but the principles did not change that much, and performance-wise a good code from 40 years ago is generally still a good code today: compilers are here to exploit hardware improvements (such as vector intructions, etc…)
There are some exceptions to that: codes written for vector machines of the 80-90 era may not perform well on scalar machines (and vice versa). It was possible to write codes in a way they could perform well of both architectures, but sometimes it was better to have two versions. And of course, GPUs often require rewritting codes from scratch (but in this case even maintenance is pointless).
Netlib’s LAPACK is a so-called “reference implementation”. As such, it does not pretend (and has never pretended) to give the highest possible performances on every possible machine. MKL has some critical parts written in assembly, and nothing can beat that.