2007/4/3, David Kastrup :

New Cicero or what?

From the pdftex NEWS for 1.30:

- pdftex knows two new units: - new Didot (1nd=0.375mm) - new Cicero (1nc=12nd) But it's broken in pdftex as well. :-(

Who does things like that without actually checking that they work? It is possible to rework the arithmetic to get along with 65535, but it is certainly not a matter of merely changing the _documentation_ and not the code.

Me. I got a patch and didn't check it. :-( Best Martin

David Kastrup wrote:

But I consider it more likely that nc should be changed to 1370:107

And also, I applied this change to luatex.web (and fixed the documentation sting back). TeX's conversions are approximations anyway, I see little point in changing the conversion arithmetic. Taco

Taco Hoekwater writes:

David Kastrup wrote:

...

But I consider it more likely that nc should be changed to 1370:107

And also, I applied this change to luatex.web (and fixed the documentation sting back). TeX's conversions are approximations anyway, I see little point in changing the conversion arithmetic.

Well, but I would find a change rendering "1in"="72.27pt" attractive. I'll propose one shortly. Anyway, those fractions carry _way_ more precision than TeX's general arithmetic anyway, and TeX's truncation (which this was all about in the first place) introduces more bias than this particular fraction does. One could argue that 32163:2512 will be a nicer fraction since it overshoots, and TeX forgets to round the final result, so we err on different sides. But I'd rather have the closest ratio and fix TeX's, excuse me (hopefully Knuth didn't hear), LuaTeX's arithmetic on units by letting it round properly. -- David Kastrup

Taco Hoekwater writes:

David Kastrup wrote:

...

But I consider it more likely that nc should be changed to 1370:107

And also, I applied this change to luatex.web (and fixed the documentation sting back). TeX's conversions are approximations anyway, I see little point in changing the conversion arithmetic.

Unfortunately, Knuth is a practical joker. The old documentation string, while sounding reassuring, is actually also wrong. With the code he has written there, the requirement is that the _sum_ of num and denum does not exceed 32768. Which happens to be the case for his ratios, but not for _either_ of the new ones. Check the following: This is luaTeX, Version 3.141592-snapshot-2007040210 (Web2C 7.5.6) **\relax *\dimen0=9.9998nd *\showthe\dimen0

7.44438pt. <*> \showthe\dimen0

? *\dimen0=10nd *\showthe\dimen0

10.66977pt. <*> \showthe\dimen0

? * So I recommend that you change the ratio of nd either to 1370:1284 (which reinstates the 12nd=1nc relation), or better change it to the closest usable ratio which is 11183:10481. So it looks like neither the new cicero nor the new didot have a chance to be represented exactly. Too bad. I am still being sidetracked from what I really wanted to look at. Does anybody actually _use_ those new dimensions? -- David Kastrup

2007/4/3, Taco Hoekwater :

12pt!=12pc, so I suspect the same is true for nds and ncs (assuming anyone actually uses these units, which is doubtful at this stage).

The bug has been in the wild since 2005-08-01, so I agree: Nobody uses them. Not even the author of the patch. :-{ Best Martin

"Martin Schröder" writes:

2007/4/3, Taco Hoekwater :

...

12pt!=12pc, so I suspect the same is true for nds and ncs (assuming anyone actually uses these units, which is doubtful at this stage).

The bug has been in the wild since 2005-08-01, so I agree: Nobody uses them. Not even the author of the patch. :-{

Well, at least not with large numbers behind the decimal point. The nc is rather easy to break with those (0.6nc does the trick), but you have to cherry-pick values for breaking nd (9.99998nd is the smallest value of form x.99998nd that actually breaks). So it is possible that the author uses nd, and possibly multiples of 0.5nc without seeing the problem. -- David Kastrup