Probably not secret anymore. CMS (the "Code Morphing Software" that implemented the x86 emulation) originally went straight to translation which was difficult to get correct and was expensive to do for code that might only be run once. Linus, when he joined said "That's stupid" and wrote an x86 interpreter which then acted as the first tier in the emulation. That let to a massive improvement in quality as more workloads could be tested and enabled an awesome creation by Jim Mattson (IIRC): self-cosimulation. CMS could be run in a mode where all translation were cross checked with the interpreter before the results were committed.
This was before my time and I'm sure he did much more. I only have first-hand knowledge of his work on TVM, the Transmeta x86 Virtualization which predated Intel (and AMD's) hardware support for x86 virtualization. Sadly it never productized. I suspect we couldn't find a way to monetize it.
At least the journal reports I read at that time implied that much. He was one of the technical leads on this as far as I recall. So he would have had to get a very good knowledge of the Transmeta CPU and of the x86 instruction set for that task. I think it shows here.
This was before my time and I'm sure he did much more. I only have first-hand knowledge of his work on TVM, the Transmeta x86 Virtualization which predated Intel (and AMD's) hardware support for x86 virtualization. Sadly it never productized. I suspect we couldn't find a way to monetize it.