Potential things you could do

[wheremyfoodat]

Hello. I saw your project on Github and Discord and would like to provide some further feedback to improve the code

• Switching to a proper emitter library for one. You've been looking into dynasm, though I would like to recommend xbyak (C++-only and the one I use) and asmjit (C++ only as well, I don't like it but it exists). If you don't want to use C++ then that's fine too, dynasm is cool.

• Performance improvements

  • xor r64, r64 where r64 = a 64-bit register is an anti-pattern. 64-bit operations are inherently more bloated because to encode them, you need to insert a REX.W prefix. Thankfully, when writing to a 32-bit register, the top 32 bits of the whole 64-bit reg are zeroed out (The same is NOT true for 8 and 16-bit register wirtes). So xor eax, eax for example is a more efficient version of xor rax, rax, mov eax, 0 and so on (Though sometimes mov is preferred over xor when the preservation of flags is important).
  • If you would like to experiment with SIMD a bit, I would recommend converting CLS to use a memset written in SSE or AVX (when available), which you could use to write 16 or 32 bytes at a time, and maybe even partially unroll the clear loop for maximum efficiency. If you'd like an example, I have some in my MIPS->x64 recompiler for PCSX-Redux: https://github.com/grumpycoders/pcsx-redux/blob/1e18f27c13a19e248947afe549ac63f984df1093/src/core/DynaRec_x64/recompiler.cc#L254. If you want help learning about SIMD in general, I have various code snippets to help you, though most are decently complicated so it would be best to discuss over text. On a side note, one of the reasons I really recommend Xbyak is that it really helps with things like this, as it has tools to help me detect when eg AVX is available, without me having to write any code for it
  • https://github.com/thamugadi/jit-chip8/blob/30f53690e0cad1d9c85588b557f02905a2020407/src/jit.c#L97-L99 Due to how x86 immediates work, you can just do mov byte ptr [&context.V[ins.x]], ins.kk since kk is a compile time constant. Same for add and other similar instructions.
  • https://github.com/thamugadi/jit-chip8/blob/30f53690e0cad1d9c85588b557f02905a2020407/src/jit.c#L173C6-L173C11 This comment should probably say setc since you're setting vf to the carry out of the shift. The 2 opcodes are equivalent however
  • https://github.com/thamugadi/jit-chip8/blob/30f53690e0cad1d9c85588b557f02905a2020407/src/jit.c#L224 You really shouldn't be using a syscall for this as they are incredibly slow and non-portable. I can show you how to implement a very basic rand in assembly if you want, or you could use RDRAND for maximum randomness. I can also show you how to reimplement the original chip-8 rand algorithm.
  • For the BCD op: You can do it without divisions in a much faster way, which is what a compiler would do, but maybe it is best not to do this if you're aiming for easy to read and understand code.

  That's it for now. Have a nice day.

[wheremyfoodat]

You also don't need to preserve rax/rcx/rdx most of the time. The calling convention says they're volatile on both MS ABI (Windows) and Sys-V (Linux, Mac, BSD, and everything not Windows).

One of my introduction-to-x86 PDFs in the emudev devnull channel explains this stuff. Of course you should respect the ABI of whatever system you're running on, and the ABIs do have some stack-related requirements that may break if you remove the pushes n pops

[thamugadi]

Thank you very much for the very detailed analysis !! :heart_eyes: This will help me enormously to improve the code :smile: