Ok, I feel you know what this is actually about, but here spelled out:

Do any of the optimizations you mentioned affect the underlying behavior of the run code? Semantics...

If not, it doesn't fit to LLM use, where this is very much the case. We can talk about bugs in compilers and undefined behaviours and so on and so forth all the livelong day, but you know as well as I that this is not of the same scope of impact.

I could and in the past would have written a long diatribe about how pure compilation to assembly is deterministic, thus making this not really suitable and that while yes, there are things like JVM that aren't 100%, we have contracts, docs, specs and, because of them, an expectation of consistency that LLMs do not deliver. I did not because I thought that was not necessary, but see my error now, that is my mistake.

There is the observable behaviour, and the machine code instructions that actually get generated in every single case for the same kind of text input.

Which for example, unless you can exactly replicate the world every single time a PGO data gathering session is done, there are no guarantees that the compiler will generate exactly the same machine code pattern, when given various PGO files generated at different occurrences.