I thought FAT binaries don’t work like that - they included multiple instruction sets with a header pointing to the sections (68k, PPC, and x86)
Rosetta to the best of my understanding did something similar - but relied on some custom microcode support that isn’t rooted in ARM instructions. Do you have a link that explains a bit more in depth on how they did that?
Fat binaries contain both ARM and x86 code, but I was referring to Rosetta, which is used for x86-only binaries.
Rosetta does translation of x86 to ARM, both AOT and JIT. It does translate to normal ARM code, the only dependency on a Apple-specific custom ARM extension is that the M-series processors have a special mode that implements x86-like strong memory ordering. This means Rosetta does not have to figure out where to place memory barriers, this allows for much better performance.
So when running translated code Apple Silicon is basically an ARM CPU with an x86 memory model.
Apple’s been doing it for years. They try to do ahead of time transpiling wherever they can but they also do it on-the-fly for things like JITed code.
I thought FAT binaries don’t work like that - they included multiple instruction sets with a header pointing to the sections (68k, PPC, and x86)
Rosetta to the best of my understanding did something similar - but relied on some custom microcode support that isn’t rooted in ARM instructions. Do you have a link that explains a bit more in depth on how they did that?
Fat binaries contain both ARM and x86 code, but I was referring to Rosetta, which is used for x86-only binaries.
Rosetta does translation of x86 to ARM, both AOT and JIT. It does translate to normal ARM code, the only dependency on a Apple-specific custom ARM extension is that the M-series processors have a special mode that implements x86-like strong memory ordering. This means Rosetta does not have to figure out where to place memory barriers, this allows for much better performance.
So when running translated code Apple Silicon is basically an ARM CPU with an x86 memory model.
That makes a lot of sense - I wonder if they also do the SIGSEGV trick like HotSpot to know when they need to JIT the next chunk of instructions