You can't PGO that code, though. Also, I'm not sure it can generate architecture specific optimized (-march, -mtune) code as well.
Considering it just pulls and remixes blocks, this is a very tall order for today's LLMs, and maybe a couple of generations as well.
Calling GCC/clang at the back and emitting ASM is both cheating and has no advantages whatsoever as well.
That is fighting the last war on the early days of 5 GLs, and taken care by agentic runtimes.
This isn't futurology, besides a few projects that already came up on Phoronix and around here, with Assembly generation, there are enough SIGPLAN papers on ML driven compilers already.