The key feature seems to be the dynamic nature while still being fast. Sure, they could also build it as a compiler that does all mentioned in the article and then dump optimized Go code. Maybe even use the Go PGO instead of their own. But this is another approach, what I mean is caching of the JIT generated code to avoid doing expensive part again while still being dynamic and adapt to incoming messages.

No, I think they want Profile-Guided Optimization. I think the C# AoT mode uses the results of a JIT first run.

My experience is that the practical performance achievable with Go is higher because the C++ lifetime issues are too difficult to reason about and therefore the developer is forced to copy for safety. In Go you can fairly easily alias everything from the physical buffer into your parsed object. In the official C++ library, protobuf refuses to acknowledge even the possibility of aliasing. Even if you say that your string types are "view" there is an owned buffer inside the generated class into which your data is copied. This is exasperating because inside Google they have several different ways to not copy a string into a protobuf, and they're all patched out of the open source edition, and you can read them and cry about it by looking at their git logs for "internal change" commits with baffling only-whitespaces changes that are symptomatic of where they are patching out the good stuff.