Nah, if you actually read into what's available there, it's clear that the compilers have never implemented features to make this broadly usable. You only get runtime instruction selection if you've manually tagged each individual function that uses SIMD to be compiled with function multi-versioning, so that's only really useful for known hot spots that are intended to use autovectorization. If you just want to enable the latest SIMD across the whole program, GCC and clang can't automatically generate fallback versions of every function they end up deciding could use AVX or whatever.
The alternative is to make big changes to your build system and packaging to compile N different versions of the executable/library. There's no easy way to just add a compiler flag that means "use AXV512 and generate SSE2 fallbacks where necessary".
The people that want to keep running new third-party binaries on 12+ year old CPUs might want to work with the compiler teams to make it feasible for those third parties to automatically generate the necessary fallback code paths. Otherwise, there will just be more and more instances of companies like Google deciding to start using the hardware features they've been deploying for 15+ years.
But you already know all that, since we discussed it four months ago. So why are you pretending like what you're asking for is easy when you know the tools that exist today aren't up to the task?
reply