I mean, it’s all a matter of tradeoffs. As I see it, there are really three key questions that’d need to be considered:
- Is it breaking? It changes the interval from closed-open to open-open, and it’s a change in the documented behaviors. We have said, though, that changing RNG results is ok. It’s worth reviewing the time we looked into unbiasing
rand()by making just the change from half-open to open: make rand unbiased for floats #33222. My read there is that maybe but the weighting of tradeoffs in that change (which is different than this) weren’t worth it. - Complexity and correctness. It’s quite clear what the distribution of
(0:2^53-1) ./ 2^53looks like, flaws and all — and there’s additionally very well-established literature and norms in other languages. Apparently zig didn’t quite get its newfangled implementation right: Oh, cute. It looks like they ever-so-slightly overweight the probability of valu... | Hacker News. - Performance.