Technical specs for Rosetta 2?

[Kainrzog]

Could anyone point me to some more detailed technical specs (meaning: not marketing) on how Rosetta 2 actually works?

Does such a document even exist?

I'm wondering, because some pseudo-technical docs that I've read state that Rosetta 2 is based on ahead of time compilation, so in theory macOS should take the amd64 binary, recompile it to aarch64, and run the compiled binary instead of the original one.

That's not what I've experienced though. My experiments indicate something else: when I compile on an amd64 macOS a trivial application that allocates a small memory buffer, and I start writing amd64 instructions directly into the buffer, then I set executable rights on the allocated memory region and I run the buffer, it runs without problems on an ARM macOS. That means it can't be an ahead of time compilation, it needs to be some kind of amd64 instruction emulation. Or something else that works similarly.

 

[bobmans]

Sadly I can't find any technical details other than this but from what I've read it tries translating everything it can ahead of time and translates what can't be done ahead of time on the fly.

 

[Kainrzog]

Thanks. Yeah, I've seen this document. I just had hopes there will be more info about it, I'm really interested in how R2 is implemented.

 

[leman]

It supports both AOT and JIT compilation. I don't think that Apple has disclosed any details on how it actually works. I suppose they mark executable pages with the architecture tag and when code in a x86-64 page is executed for the first time, Rosetta compiler kicks in. For applications on disk, they can cache the transpiration results or even do it during installation.

You could make some fun experiments using your simple code generator, like measuring the execution start latency and checking how modifying the code in the buffer affects it.