One of the things I find interesting about SIMD is that a lot of behavior that is “undefined” for scalar types in C-derived languages is explicitly fully defined when you use SIMD intrinsics with the same integral types. UB exists to hide the fact that major CPU architectures give different results for basic ALU operations in some cases. SIMD makes no such pretense of abstraction. If I am using AVX-512 I explicitly get the full Intel architecture experience, the implementation details are not hidden behind UB. Same with ARM, etc.

For example, shift overflows are masked on x86, zero-filled on ARM, and undefined in C/C++. In SIMD-land, none of this is hidden and so you design your code to leverage the reality that those instructions behave differently, whereas in C/C++ only the behavior they have in common is “defined”.

The vector ISAs are sufficiently different from each other (and normal CPUs) that it is like trying to build a compiler that can automagically produce optimized code for both CPUs and GPUs from the same source tree. I am not optimistic that this will happen anytime soon. AVX-512 essentially started life as a GPU ISA, which probably explains the interesting fact that a modern x86 CPU core has more AVX-512 registers than x86 registers.