←back to thread

C++ proposal: There are exactly 8 bits in a byte

(www.open-std.org)
288 points Twirrim | 1 comments | 17 Oct 24 22:21 UTC | HN request time: 0.215s | source
Show context
favorited ◴[17 Oct 24 23:54 UTC] No.41875023[source]▶
Previously, in JF's "Can we acknowledge that every real computer works this way?" series: "Signed Integers are Two’s Complement" <https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p09...>
replies(1): >>41875200 #
jsheard ◴[18 Oct 24 00:25 UTC] No.41875200[source]▶
Maybe specifying that floats are always IEEE floats should be next? Though that would obsolete this Linux kernel classic so maybe not.

https://github.com/torvalds/linux/blob/master/include/math-e...

replies(9): >>41875213 #>>41875351 #>>41875749 #>>41875859 #>>41876173 #>>41876461 #>>41876831 #>>41877394 #>>41877730 #
FooBarBizBazz ◴[18 Oct 24 02:03 UTC] No.41875749[source]▶
Whether double floats can silently have 80 bit accumulators is a controversial thing. Numerical analysis people like it. Computer science types seem not to because it's unpredictable. I lean towards, "we should have it, but it should be explicit", but this is not the most considered opinion. I think there's a legitimate reason why Intel included it in x87, and why DSPs include it.
replies(2): >>41875950 #>>41876023 #
stephencanon ◴[18 Oct 24 03:04 UTC] No.41876023[source]▶
Numerical analysis people do not like it. Having _explicitly controlled_ wider accumulation available is great. Having compilers deciding to do it for you or not in unpredictable ways is anathema.
replies(2): >>41876108 #>>41879704 #
FooBarBizBazz ◴[18 Oct 24 14:16 UTC] No.41879704[source]▶
I get what you mean and agree, and have seen almost traumatized rants against ffast-math from the very same people.

After digging, I think this is the kind of thing I'm referring to:

https://people.eecs.berkeley.edu/~wkahan/JAVAhurt.pdf

https://news.ycombinator.com/item?id=37028310

I've seen other course notes, I think also from Kahan, extolling 80-bit hardware.

Personally I am starting to think that, if I'm really thinking about precision, I had maybe better just use fixed point, but this again is just a "lean" that could prove wrong over time. Somehow we use floats everywhere and it seems to work pretty well, almost unreasonably so.

replies(1): >>41884242 #
1. stephencanon ◴[18 Oct 24 22:58 UTC] No.41884242[source]▶
Yeah. Kahan was involved in the design of the 8087, so he’s always wanted to _have_ extended precision available. What he (and I, and most other numerical analysts) are opposed to is the fact that (a) language bindings historically had no mechanism to force rounding to float/double when necessary, and (b) compilers commonly spilled x87 intermediate results to the stack as doubles, leading to intermediate rounding that was extremely sensitive to optimization and subroutine calls, making debugging numerical issues harder than it should be.

Modern floating-point is much more reproducible than fixed-point, FWIW, since it has an actual standard that’s widely adopted, and these excess-precision issues do not apply to SSE or ARM FPUs.