I looked at the patch
<http://launchpadlibrarian.net/20858726/sort_human_v3.diff> and have a
few comments:
* There's no documentation, either in the .texi file or in the --help
string. That's often the hardest part to get right.
* The patch slows down "sort -n" a bit, and (worse) complicates the code
for "sort -n". There are a lot of gotos. Better, I think, would be
to (1) search for a trailing suffix character, and compare based on
that, and then (2) fall back on numeric comparison only if the
trailing suffix characters match. This will certainly make plain
"sort -n" faster and will result in a less-intrusive change to the
existing code; arguably it will make the new code faster in many
cases, too.
* Please don't take the address of an often-used local variable; this
hurts optimization, since it often inhibits optimizing that variable
into a register.
* I suggest ignoring any suffix after the 'k' or 'M' or 'G' or whatever;
it's simpler.
* I also suggest allowing only the upper-case version (e.g., 'P', not
'p'), as this will allow later support for ISO prefixes like 'p' (for
"pico") better. (It's OK to allow both 'k' and 'K' to mean kilo,
though.)
* exponent_order should be a simple indexed lookup like "table[a]";
there shouldn't be a need for a loop checking for "inverse_table[i] ==
a".
Hope this helps.