> A subset of an ISA will be incompatible with the full ISA and therefore be a new ISA. No existing software will run on it. So this won't really help anyone.

This isn't an issue in any way. Vendors have been routinely taking out rarely used instructions from the hardware and simulating them in the software for decades as part of the ongoing ISA revision.

Unimplemented instruction opcodes cause a CPU trap to occur where the missing instruction (s) is then emulated in the kernel's emulation layer.

In fact, this is what was frequently done for «budget» 80[34]86 systems that lacked the FPU – it was emulated. It was slow as a dog but worked.

There are software compiled today without using MMX support. I was thinking the idea of something that is open or for licensing is an 86 ISA that is forward compatible. And for customers that requires strict backward compatibility they could still source it from AMD and Intel.

i.e Software compiled for 86 should work on x86. The value for backward compatibility is kept with both Intel and AMD. If the market wants something in between they now have an option.

I know this isn't a sexy idea because HN or most tech people like something shiny and new. But I have always like the idea of extracting value from the "old and tried" solutions.

Sadly over the past year, Spotify builds require AVX extensions. Had an issue updating my 2008 Dell semi-upgraded bench PC that has a Q9300 in it (no AVX on it)

But thankfully I could install an old bin and lock it out from updating.

Intel’s software development emulator might run the newest bin but variable how slow it might be.

In other circumstances, the AVX extensions aren’t required but the app is compiled to fail if they’re not required: https://www.reddit.com/r/pcgaming/comments/pix02j/hotfix_for...

Software or microcode emulation works pretty well.

So it would be faster and more efficient when sticking to the new subset and Nx slower then using the emulation path.

You could argue that microcode emulation is what they do now.

True, the only part that can not be solved by microcode emulation is faster decoding.

Most architectures other than x86 have fixed sized machine instructions now, making decoding fast and predictable.