ts-native VS dynarmic

I have gone through 3 iterations and the third was faster than Lua and came close (enough for me) to native C performance with the benchmarks I tried. I'm now on the fourth and hopefully last iteration of this masochistic quest and I'm tired. I want to do this right. I've even made it through much of the dragon book. Apart from the naive solutions I end up wanting to replace in each previous iteration, the problem I keep coming up against is that I can't find many resources that describe interfaces between the host (C++) and the compiled code. I come up with my own ideas for how things should work and look, and then once I finish I can't shake the feeling that there's a better/cleaner/faster way. I start over and come up with better solutions only to feel the same eventually. I want to write solid code that I can feel confident about so I can put this all behind me.

Here's an example: test.gjs

They're probably referring to dynarmic: https://github.com/merryhime/dynarmic

As an HLE, touchHLE is radically different from a low-level emulator (LLE) like QEMU. The only code the emulated CPU executes is the app binary and a handful of libraries; touchHLE takes the place of iPhone OS and provides its own implementations of the system frameworks (Foundation, UIKit, OpenGL ES, OpenAL, etc).

https://github.com/merryhime/dynarmic/pull/697 - Dynarmic implementing Arm backend support. Dynarmic is a dynamic recompiler for ARM.

I did some work on Dynarmic, so there's my bias I guess. But I def believe Dynarmic would emit better context-aware assembly than a CIL->Native jitteror the JVM would emit. Dynarmic takes advantage of instruction sets like BMI2 and AVX{2,512} and uses a pretty darn concise intermediate representation to the original ARM assembly to emit efficient x86 while a C# JIT probably still only emits baseline x86-64 and struggles with efficient vectorization of things like Arm's NEON instructions. Can't even imagine a C# or Java VM trying to automatically emit optimal assembly for an FMADD including handling NaN propagation or detecting that FMINNM can be very quickly emulated with a single vrangep{s,d} x86 instruction.

For emulators, it's currently used in dynarmic(Yuzu's ARM-to-x86 recompiler backend) for the fast emulation of ARM's NEON instruction set.