Mike Cowlishaw’s FAQ on dec­i­mal arith­metic

If you’re in­ter­est­ed in dec­i­mal arith­metic in com­put­ers, you’ve got to check out Mike Cowlishaw’s FAQ on the sub­ject. There’s a ton of in­sight to be had there. If you like the kind of writ­ing that makes you feel smarter as you read it, this one is worth your time. For con­text: Cowlishaw is the ed­i­tor of the 2008 edi­tion of the IEEE 754 stan­dard, up­dat­ing the 1985 and 1987 stan­dards. The words thus car­ry a lot of au­thor­i­ty, and it would be quite un­wise to ig­nore Mike in these mat­ters. If you pre­fer sim­i­lar in­for­ma­tion in ar­ti­cle form, take a look at Mike’s Dec­i­mal Float­ing-Point: Al­go­rism for Com­put­ers. (Note the de­light­ful use of al­go­rism. Yes, it’s a word.) The FAQ fo­cused main­ly on float­ing-point dec­i­mal arith­metic, not ar­bi­trary-pre­ci­sion dec­i­mal arith­metic (which is what one might im­me­di­ate­ly think of when the one hears dec­i­mal arith­metic). Ar­bi­trary-pre­ci­sion dec­i­mal arith­metic is whole oth­er ball of wax. In that set­ting, we’re talk­ing about se­quences of dec­i­mal dig­its whose length can­not be spec­i­fied in ad­vance. Pro­pos­als such as dec­i­mal128 are about a fixed bit width—128 bits—which al­lows for a lot of pre­ci­sion, but not ar­bi­trary pre­ci­sion. One cru­cial in­sight I take away from Mike’s FAQ—a real mis­un­der­stand­ing on my part which is a bit em­bar­rass­ing to ad­mit—is that dec­i­mal128 is not just a 128-bit ver­sion of the same old bi­na­ry float­ing-point arith­metic we all know about (and might find bro­ken). It’s not as though adding more bits meets the de­mands of those who want high-pre­ci­sion arith­metic. No! Al­though dec­i­mal128 is a fixed-width en­cod­ing (128 bits), the un­der­ly­ing en­cod­ing is dec­i­mal, not bi­na­ry. That is, dec­i­mal128 isn’t just bi­na­ry float­ing-point with ex­tra juice. Just adding bits won’t un­break bust­ed float­ing-point arith­metic; some new ideas are need­ed. And dec­i­mal128 is a way for­ward. It is a new (well, rel­a­tive­ly new) for­mat that ad­dress­es all sorts of use cas­es that mo­ti­vate dec­i­mal arith­metic, in­clud­ing needs in busi­ness, fi­nance, ac­count­ing, and any­thing that uses hu­man dec­i­mal num­bers. What prob­a­bly led to my con­fu­sion is think­ing that the ad­jec­tive float­ing-point, re­gard­less of what it mod­i­fies, must be some kind of vari­a­tion of bi­na­ry float­ing-point arith­metic.