[titzer]

She's exactly right. I'm really not sure how we came to accept the worse-is-better argument that high performance required exposing machine details into the source language, and letting compiler writers just "do whatever is fastest", regardless of safety. It gave rise to today's situation where none of application writers, library writers, kernel writers, or compiler writers are quite sure what will happen if a new optimization is introduced. Will it break programs? By how much? Well, let's roll the dice! This is a direct result of allowing--no, even expecting--programmers to peer behind the curtains to see machine details that they frankly, shouldn't. It's why porting programs from one machine to another is even a thing, as opposed to just selecting a different output from a compiler, or running a different VM.

Proper abstractions define exactly what happens in every situation, and design unobservable details out. It is the basis for optimizations that fundamentally cannot observed (except side-channels, like timing and other resource usage). Proper abstractions is why we have fast processors. The ISA is a hard boundary.