> On Oct 23, 2017, at 1:47 AM, Brent Royal-Gordon via swift-dev > <swift-dev@swift.org> wrote: > >> On Oct 21, 2017, at 6:27 PM, Xiaodi Wu via swift-dev <swift-dev@swift.org> >> wrote: >> >> Steve can describe the exact number of additional machine instructions on >> each architecture, but from my cursory reading the performance cost is not >> good and there's little cleverness to be had. > > I'm not sure why you think there's no cleverness to be had. For instance, > suppose we loosen our requirement a bit: We require the `==` operator to be > reflexive, but we allow `BinaryFloatingPoint.==` to treat NaNs with different > payloads as different values. That very nearly allows us to use integer > equality for floating-point comparison; we only need a +0/-0 special case:
This way leads to madness: a set S that contains a NaN, but S.contains(T.nan) is false. You’re going to run into worse problems when you get to Decimal; is 1e2 == 100e0? They have totally different encodings. If we’re going to define another notion of floating-point equality in order to satisfy the usual mathematical axioms, it should treat all NaNs as equal, and all encodings of any non-zero value as equal. Whether it treats +0 and -0 as the same or different is the only really question from my perspective (I come down *very* weakly on the side of “they’re different”). – Steve _______________________________________________ swift-dev mailing list swift-dev@swift.org https://lists.swift.org/mailman/listinfo/swift-dev
