- Testing requirements after changes:
  * Test all functions return either native or bigints.  Functions that
    return raw MPU::GMP results will return strings, which isn't right.
  * Valgrind, coverage
  * use:  -O2 -g -Wall -Wextra -Wdeclaration-after-statement -fsigned-char
  * Test on 32-bit Perl.  Test on Win32.


- Move .c / .h files into separate directory.
  version does it in a painful way.  Something simpler to be had?

- finish test suite for bignum.  Work on making it faster.

- An assembler version of mulmod for i386.

- More efficient Mertens.  The current version has poor growth.
  For 32-bit, consider doing XS followed by sum for remainder.

- It may be possible to have a more efficient ranged totient.  We're using
  the sieve up to n/2, which is better than most people seem to use, but I'm
  not completely convinced we can't do better.  The method at:
  http://codegolf.stackexchange.com/a/26747/30069 ends up very similar.  For
  the monolithic results the main bottleneck seems to be the array return.

- Big features:
   - QS factoring

- Figure out a way to make the internal FOR_EACH_PRIME macros use a segmented
  sieve.

- Rewrite 23-primality-proofs.t for new format (keep some of the old tests?).

- Factoring in PP code is really wasteful -- we're calling _isprime7 before
  we've done enough trial division, and later we're calling it on known
  composites.  Note how the XS code splits the factor code into the public
  API (small factors, isprime, then call main code) and main code (just the
  algorithm).  The PP code isn't doing that, which means we're doing lots of
  extra primality checks, which aren't cheap in PP.

- Consider using Test::Number::Delta for many tests

- More tweaking of LMO prime count.
    - OpenMP.  The step 7 inner loop is available.
    - Convert to 32-bit+GMP to support large inputs, add to MPU::GMP.
    - Try __int128.
    - Variable sieve size
    - look at sieve.c style prime walking
    - Fenwick trees for prefix sums

- Iterators speedup:
  1) config option for sieved next_prime.  Very general, would make
     next_prime run fast when called many times sequentially.  Nasty
     mixing with threads.
  2) iterator, PrimeIterator, or PrimeArray in XS using segment sieve.

- Perhaps have main segment know the filled in range.  That would allow
  a sieved next_prime, and might speed up some counts and the like.

- Benchmark simple SoEs, SoA.  Include Sisyphus SoE hidden in Math::GMPz.

- Try using malloc/free for win32 cache memory.  #define NO_XSLOCKS

- Investigate optree constant folding in PP compilation for performance.
  Use B::Deparse to check.

- Ensure a fast path for Math::GMP from MPU -> MPU:GMP -> GMP, and back.

- We don't use legendre_phi for other functions any more, but it'd be nice
  to speed it up using some ideas from the Ohana 2011 SAGE branch.  For example
  (10**13,10**5) takes 2.5x longer, albeit with 6x less memory.

- More Pari:  parforprime

- znlog:
    = GMP BSGS for znlog.
    = Clean up znlog (PH, BSGS, Rho).
    = Experiment with Wang/Zhang 2012 Rho cycle finding

- consider using Ramanujan Li for PP code.

- xt/pari-compare:  add chinese, factorial, vecmin, vecmax,
                        bernfrac, bernreal, LambertW.

- Proth test using LLR.  Change mersenne test file to test both.
  Note: what does this mean?  Both LLR and Proth are in GMP now.

- harmreal and harmfrac for general $k

- For PP, do something like the fibprime examples.  At load time, look for
  the best library (GMPz, GMP, Pari, BigInt) and set $BICLASS.  Then we
  should use that class for everything.  Go ahead and return that type.
  Make a config variable to allow get/set.

- Support FH for print_primes.  PerlIO_write is giving me fits.

- Test for print_primes.  Not as easy with filenos.

- We should have a way to exit a forprimes loop.  E.g. add a lastfor() call.
  Nested calls are doable, but threading is a bear.  Perhaps do a first
  version that is non-threadsafe.

- sum primes better than current method.  Especially using less memory.

- bernreal and bernfrac using MPFR

- divsum and divsummult as block functions.
  The latter does sum = vecprod(1 + f(p_i) + f(p_i^2) + ... f(p_i^e) for all p.

- surround_primes

- More Montgomery: znlog, catalan

- verify chinese to GMP is ok

- numtoperm and permtonum.
  Perhaps Bonet: https://github.com/soniakeys/perm/blob/master/lex.go
  Also consider randperm(n)

- polymul, polyadd, polydiv, polyneg, polyeval, polyorder, polygcd, polylcm, polyroots, ...
  A lot of our ops do these mod n, we could make ..mod versions of each.

- Berkane/Dusart 2016, Axler 2016, Axler 2017 could tune prime count bounds.

---- \/ ------  Next release

- publish info about Perl chacha20 (ianix, blog?).
- Remove primeinc config option.
- put entropy_bytes call in XS.
- have the RNG return a context and store it in our package context space.
  Should help threaded performance by removing lock.
  Careful to not leak state.
- Make a feature table for rand modules.  bits-per-double, secure, speed,
  seeding, interval, etc.  Lots of modules only fill doubles with 32 bits.
- get info on different drands.  TestU01, 45-bit look, cycles, x-y plots, etc.
- further processing on Ramanujan prime limits up to 4e14