Wunk

[AVX512 doubles the amount of SIMD registers but requires instructions to emit in EVEX form](https://en.wikipedia.org/wiki/AVX-512#SIMD_modes). It might be worth detecting when AVX512{F,VL} is available and having an alternative `UseScratchXmmEVEX` function...

Actually a better idea with likely larger gains is maybe even just keeping the allocator as it is and using those new registers as extra spill-space.

Note that xbyak's detection of GFNI was improperly coupled with detection of AVX512 up until this PR https://github.com/herumi/xbyak/pull/154

`waitpkg` instructions have been added to xbyak as of https://github.com/herumi/xbyak/commit/898c354e67313b194efe3a66e0f502ed4dac35ed

I've made a draft of this, but I do not currently have the required hardware available to me at the moment to validate that the implementation works or that it...

Might not like the design, but maybe something could allow the backend to emit IR opcodes of its own so it can hint, to itself, to emit certain instructions. Something...

The `vcvt.f32.f16` issue seems to be an issue when processing float16 NaN values like `0xff88` turning into `0xfff10000` while unicorn is expecting `0x7fc00000`(default nan) despite the DN flag not being...

> Sometimes, when reading the RLE stream (step 4), I sometimes get a 'stack_buffer_overflow' asan crash, pointing at the saistream in libsai: Each RLE stream is prefixed with a `uint16_t...

I think the issue might be that at the time of writing, `Read` used to return the number of bytes read(`return Size`), while now(probably in my attempts at fixing that...

I've implemented some basic rearchitecting and added a new sample `Document.cpp` that dumps layer images, so now I have a proper testbed to get some much cleaner code in and...