To do (or consider)
===================

Deccoeff type
-------------

* Add check that MAX_DIGITS/LIMB_DIGITS is small enough;  twice that
  value should fit into a Py_ssize_t.

* relax MAX_DIGITS on 64-bit platforms?

* make it possible to choose LIMB_DIGITS at configure time?  if so,
  add LIMB_DIGITS = 8 option, for purposes of experimentation.

* change default choice of LIMB_DIGITS: it's probably not wise to
  automatically use LIMB_DIGITS = 18 when the 128-bit unsigned type is
  available, since LIMB_DIGITS = 9 can be faster, even on a 64-bit
  machine.

* expand Deccoeff-specific tests

* improve and correct documentation; remove outdated deccoeff.txt; ReST!

* fast recursive algorithms for division, base conversion

* (fast recursive) square root

* implement div_nearest and sqrt_nearest: they're bottlenecks

* make scale_Py_ssize_t safer: use deccoeff instead of Py_ssize_t for
  high-precision case

* minor optimization opportunities:

  - in Karatsuba, accumulate all additions for middle term without
    propagating carry

  - make limb_error checks only for debug build

  - limbs_lshift and limbs_rshift could be faster when the shift count
    is a multiple of LIMB_DIGITS.

  - when LIMB_DIGITS == 9, base conversion could be a factor of 2 faster
    (asymptotically) by operating on two PyLong_Digits at a time instead of
    one.

_Decimal type
-------------

* recode the various Decimal operations in C.  str and cmp are
  particularly high priority.  C version of __str__ should write
  unicode directly, rather than writing bytes and then converting from
  bytes to unicode.

* fill in missing docstrings

* make exponent a Deccoeff (first requires making Deccoeff signed).
  this should offer improvements in parsing and printing of Decimal.
  OR:

* make exponent a C Py_ssize_t (or long? or int64_t?), for speed.  Would
  require some thinking about exponent and coefficient length
  constraints, with reference to the IBM specification.  Unless
  there's a huge speed advantage from this, I'm more inclined to leave
  exponents being arbitrary precision integers.

_Context type
-------------

To do in its entirety.