Lots of low-level optimizations also made their way into compilers that sometimes do a better job than a human. There's not much to be gained by writing everything in assembly which means it doesn't interest people as much as in the 90s.
There is a lot to be gained from writing your code in a way that makes use of SIMD instructions. Also, a lot of things that you can write in assembler code is insanely hard to express in higher-level languages, so of course
- the compiler may implement some specific low-level optimization
- but the compiler can (in general) not easily change the programming language so that low-level programming tricks can (without "contortions") be formulated in the programming language.
I agree that if the compiler was "allowed" to change the (higher-level) programming language, too, by quite a lot, so that quite a lot of low-level programming tricks can be formulated or much better formulated in it, then I would somewhat agree that the advantages of low-level coding have become smaller.
But this is currently clearly not the case.
The problem was in my opinion not that the constraints were gone, but the fact that the PC did not provide a very stable platform anymore on which you could do some crazy low-level optimizations.