Note that the benchmark code is inconsistent - nanb3 and nanb4 allocate 10x the number of objects of nanb and nanb2, so I was getting very surprising results until I realised that and fixed it.
On a 32-bit system, it seems tagged pointers have the advantage:
nanb 18.562s - NaN boxing
nanb2 17.484s - tagged pointer
nanb3 25.468s - tagged structure
nanb3 with pack(1) 26.000s
nanb4 20.093s - tagged union
nanb4 with pack(1) 19.109s
I also tried with pack(1) to see if cache effects would show up, and although it made the tagged structure slightly slower, indeed you can see that the tagged union became faster without alignment padding.
On a 32-bit system, it seems tagged pointers have the advantage:
I also tried with pack(1) to see if cache effects would show up, and although it made the tagged structure slightly slower, indeed you can see that the tagged union became faster without alignment padding.