cross-posted from: https://lemmy.ml/post/41409887

I've been building a bytecode VM in Rust and recently implemented NaN boxing for value representation. Sharing here for anyone interested.

I needed all VM values (booleans, integers, string pool indices, bytecode references) to fit in 64 bits (stack is Vec).

My implementation encodes 5 distinct types using a 3-bit tag and 32-bit payload, all within a single u64. It also has 15 unused bits, they may be used later for types expansion.

I'm using a 64-bit layout:

• Bits 63-51: Quiet NaN signature (0x7FFC...)
• Bits 50-18: 32-bit payload (integers, string pool indices, etc.)
• Bits 17-3: Unused/ (15 bits)
• Bits 2-0: 3-bit type tag

So it allows me to have 5 tagged types: TRUE_VAL, FALSE_VAL, STRING_VAL, CALLDATA_VAL, U32_VAL