- Testing requirements after changes:
  * Test all functions return either native or bigints.
  * Valgrind, coverage
  * use:  -O2 -g -Wall -Wextra -Werror=vla -Wdeclaration-after-statement -fsigned-char
  * Test on 32-bit Perl, Cygwin, Win32.
  * Test on gcc70 (NetBSD), gcc119 (AIX/Power8), gcc22 (MIPS64), gcc115 (aarch)
  * prove -b -I../Math-Prime-Util-GMP/blib/lib -I../Math-Prime-Util-GMP/blib/arch

- For new functions:
  XS,.h,.c, PP, PPFE, export,t 02+92, lib/ntheory.pm,Changes,TL, doc, test(2)


- 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.

- 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?).

- Research: create random proven prime method using ECPP blocks.

- Use BLS5 in PrimalityProving.pm for random_maurer_prime

- 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.

- More Pari:  parforprime

- znlog rho: Wang/Zhang 2012, Teske/r-adding, or negation-map rho

- consider using Ramanujan Li for PP code.

- xt/pari-compare:  add chinese2, 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

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

- Test for print_primes.  Not as easy with filenos.

- 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.

- Consider Lim-Lee random prime generation, optionally with proof.
  https://pdfs.semanticscholar.org/fd1d/864a95d7231eaf133b00a1757ee5d0bf0e07.pdf
  libgcrypt/cipher/primegen.c

- More formal random prime generation for pedantic FIPS etc. users, with
  guarantee of specific algorithm.

- surround_primes

- More Montgomery:
   - catalan
   - test znlog code

- Generic 64-bit montgomery

- 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.

- poly_is_reducible

- use word-based for-sieve for non-segment.

- remove start/end partial word tests from inner loop in for-sieve.

- sieve.h and util.h should get along better.

- compare wheel_t with primes separated and possibly cached.

- Destroy csprng context on thread destruction.

- submit bug report for Perl error in 30b8ab3

- localized a/b in vecreduce, see:
  https://metacpan.org/diff/file?target=REHSACK/List-MoreUtils-XS-0.428/&source=HERMES%2FList-MoreUtils-XS-0.427_001#XS.xs
  perl #92264 (sort in 5.27.7)

- consider #define PERL_REENTRANT

- add back formultiperm optimization if we can get around lastfor issue.

- make a uint128_t version of montmath.  Needs to handle 64-bit.
  also see:
  https://github.com/int128-libraries/curtint128/blob/master/include/cuda_uint128.h

- srand with no args should be calling GMP's srand with the selected seed
  value.  This is all a hacky artifact of having the two codebases.

- Look at using Ramanujan series for PP Li.

- There may be more we can get from Dusart's 2016 paper:
  https://piyanit.nl/wp-content/uploads/2020/10/art_10.1007_s11139-016-9839-4.pdf
- clean up the prime count bounds for small inputs.
- Dusart 2018 also:  ~/Downloads/2018_Estimatesofthekthprime_10.1007_s11139-017-9984-4.pdf

- circular primes ... just use repdigits after 1M?  https://oeis.org/A068652

- perhaps square-free flag for factor for early stop.  Use in moebius etc.

- make a NVCONST, define log, sqrt, etc. for quadmath vs. long double

- move most of our long double routines to NVCONST (see above).

- Change from Kalai to Bach's algorithm for random factored integers
  https://maths-people.anu.edu.au/~brent/pd/multiplication-HK.pdf

- Adjust crossover in random_factored_integer PP code for Kalai vs. naive

- Things from Pari/GP 2.12 beta:
   - rewritten (much faster) Bernoulli.
   - factorial
   - divisors?
   - DLP/PH

- semiprime_count PP just walk if small range.

- add b125527.txt to oeis 125527.  (semiprime counts 2^n)

- improve 64-bit limit estimates for twin prime and ramanujan prime.

- add to A033843 (twin prime count < 2^n).  Oliviera e Silva has good data.

- multicall in forpart/forcomp.

- check memory use for non-multicall.  We need enter/leave which were removed.

- testing:  lehman_factor, print_primes, aks

- factor, factor_exp should accept negative inputs

- NEGMAXINT testing in PP.

- NEGMAXINT input in XS.

- In theory IFLAG_ABS in _validate_and_set should allow up to neg(~0).
  Looks a bit messy but possible.

- Make prime_omega, prime_bigomega, and liouville take a range like moebius.

- think about making an iterator for range omega/bigomega.  We can precalc
  the primes and offsets, which should enable fast sieving of small windows.

- dickman_rho, debruijn_psi
  See: https://arxiv.org/pdf/1301.5293.pdf
       https://www.ams.org/journals/mcom/1969-23-106/S0025-5718-1969-0247789-3/S0025-5718-1969-0247789-3.pdf
       Hunter/Sorenson(1997)
       https://cosec.bit.uni-bonn.de/fileadmin/user_upload/teaching/08us/08us-cryptabit/AnatInt_Crypto-sld.pdf
       https://arxiv.org/pdf/2604.22058

- consider random_smooth_integer, random_rough_integer.  nbit or range or mod?
  See: https://cr.yp.to/papers/epsi.pdf
       https://arxiv.org/pdf/2006.07445.pdf

- ipowsafe could have the limits using hard-coded sqrt and cbrt to avoid div.

- almost primes check and enter new for
   3 http://oeis.org/A109251
   4 http://oeis.org/A114106
   5 http://oeis.org/A114453
   6 http://oeis.org/A120047
   7 http://oeis.org/A120048
   8 http://oeis.org/A120049
   9 http://oeis.org/A120050
  10 http://oeis.org/A120051
  11 http://oeis.org/A120052
  12 http://oeis.org/A120053
  4^n https://oeis.org/A116426
      mpu 'for (0..60) { say "$_ ",almost_prime_count($_,powint(4,$_)); }'
  6^n https://oeis.org/A116427
  8^n https://oeis.org/A116428
  9^n https://oeis.org/A116429
  nth https://oeis.org/A101695
  Try sequence A052130.
    a(n) = {my m = ceil(n*log(3)/log(1.5)); return apc(1<<m,m-n); }
    Or: apc(m=floor(n*(1+1/sqrt(2))); return apc(2^(n+m+2), m+2)
    mpu '$n=19; $m=int($n*log(3)/log(1.5)+1); say almost_prime_count($m-$n,1<<$m);'
    mpu '$n=19; $m=int($n*(1+1/sqrt(2))); say almost_prime_count($m+2,1<<($n+$m+2));'
    It looks like the Mathematica table has a misplaced paren.
  Sequence A078843:
    mpu 'use POSIX; $|=1; sub app3i { my($k,$n)=@_; almost_prime_count($k,$n) - (($k < 1) ? 0 : almost_prime_count($k-1, divint($n,3))); } my $a=1; for my $n (1..60) { $k=POSIX::ceil($n*log(5/3)/log(5/2)); $a +=app3i($n-$k,divint(powint(3,$n),powint(2,$k))); print "$a, "; }'


- Almost primes things:
    almost_prime_count(n,~0): n=17   23m   n=16   41m   n=15  71m   n=14  127m
                              n=13  225m   n=12  397m   n=11 699m   n=10 1217m
                              n=9    42hr  n=8    73hr  n=7  123hr  n=6  185hr
                              n=5   256hr  n=4   288hr  n=3  209hr  n=2   10hr

  - optimize PP is_almost_prime
  - better PP almost_prime_count bounds for k > 3
  - optimize PP nth_almost_prime and bounds (this is super slow)
  - Consider getting rid of the unused construction code.
  - almost_prime_count(k,beg,end)
  - almost_prime_count_approx revisit weighting.

- almost_prime_count lower/upper
  mmpu '$n=subint(powint(2,64),1); ($k,$c)=(26,3323560145881); for (0..100000000) { $c-- if is_almost_prime($k,$n); $n--; die "$n $c" if almost_prime_count_lower($k,$n) > $c; }'

- optimize nth_powerful, both in C and PP.

- Add OEIS sequence, a(n) = the k-th k-powerful number
  mpu 'for my $k (1..40) { say "$k ",nth_powerful($k,$k); }'

- For almost primes in PP, maybe use Lagrange estimator to start, given
  that we have no bounds.

- omega primes, work on figuring out formula for omega_prime_count
  better construction and counting
  OEIS sequence for omega k=3 counts for 10^i
  look into the code for omega prime count k=2, especially overflow pr[]

- There might be a better method for _sqrtmod_prime_power.

- extend qnr with optional root argument, cubic non-residue, etc.

- practical numbers: make OEIS sequence with count <= 10^n.
  mpu '$s=1; for my $e (1..9) { $s += is_practical($_) for 10**($e-1)+1..10**$e; say "10^$e  $s"; }'
  mpu 'for my $e (1..9) { $s=0; $s += is_practical($_) for 1..10**$e; say "10^$e  $s"; }'
  5, 30, 198, 1456, 11751, 97385, 829157, 7266286, 64782731, 582798892, 5283879886

- consider forprimepowers { ... } beg,end

- sum_prime_powers(start,end)  or  prime_power_sum(start, end)
  Look at sum_primes and do similar 128-bit
  See Sidef for implementation

- test coverage for PP.

- Revisit AKS in PP.  Essentially all the time is spent in two lines of
  poly_mod_mul.  (2) try using bigint multiply (Kronecker) like we do in GMP.

- any way to make non-GMP random_strong_prime faster

- Speed up XS RiemannR with quadmath.

- binomialmod:
   - PP implementation for primes, squarefree, and general composites.
   - better factorialmod_without_p.  E.g.
       binomialmod(powint(10,18),5*powint(10,17),powint(2,60))'
       binomialmod(1234567890123456789,123456789012345678,powint(999979,3)

- Possible new:
  - checksums?  Rather than print | md5sum.  adler32, sha1/2/3, blake2 (b2sum)

- inverse sigma.  Better, determine how to generalize this somehow.

- consider negative k in sigma:
  https://github.com/trizen/sidef/commit/00f083f4645b810292f9fbf4020e762279fdb4d4

- stronger BPSW test:  https://arxiv.org/abs/2006.14425

- faster prime gaps:  https://arxiv.org/abs/2012.03771
                      https://github.com/sethtroisi/prime-gap

- Pari 2.13 has faster exp, Catalan, log2, gamma, factorial, lngamma.
  Completely new Bernoulli and bernvec.
  eulerreal, eulervec, ramanujantau
  MPQS
  Take a look.

- Pari 2.13 added an optional third argument to sqrtint, just like rootint.
  Not high priority since we can just call rootint with k=2.

- Faster Mertens. Helfgott/Thompson 2020.  https://arxiv.org/pdf/2101.08773.pdf
  Also:  https://arxiv.org/pdf/2607.07566

- gcdext, see Sorenson, Jabelean, and kernel/none/gcdll.c.

- OEIS A186995
  mmpu 'for $b (2..29) { forprimes { if (is_delicate_prime($_,$b)) { say "$b $_"; lastfor(); } } 1e13; }'

- refactor lucas code.
  split out all the different codes, benchmark them all.

- Complete reviewing docs for positive vs. non-negative, and comparing
  to XS.xs.  Everything after modular functions was not reviewed.

- PP needs a revamp of the bigint->int downgrade.  Using the babs(BMAX) is
  wrong.  We can use max/min.  Better come up with something more consistent.
  Possibly XS, or try "$x=$n->numify; return ($x eq $n) ? $x : $n;"

- GMP
  - hi  partitionsq
  - hi  fubini
  - hi  vecprefixsum
  - med smooth_count
  - med rough_count
  - low qnr
  - low almost primes
  - low omega primes

- an overload option or module, to call our *int functions.
  have: add, sub, mul, pow, div, rem, neg, ++, --, <<, >>, <=>
  need: not, bnot, and, or, xor, gt, lt, geq, leq, eq, neq,
        as_bool, as_string, as_num, clone

- 32-bit testing

- Lots of older PP code, especially factoring, is built on a bigint vs PP
  idea.  It might be useful to write normally using Mpowmod etc. and
  benchmark various inputs.

- incremental factoring.  Maybe a stateful iterator?

- consider forfactors { } n.  Iterate through factors using iterator.
  Allows convenient early exit.

- new semiprime approximations should be used in PP.

- semiprime bounds (for count or nth, inverse for the other)

- twin prime bounds (for count or nth, inverse for the other)

- omega prime bounds (for count or nth, inverse for the other)

- omega prime approx (for count or nth, inverse for the other)

- faster squarefree count in PP.

- faster squarefree sieving.   https://arxiv.org/pdf/2411.01964

- PP lucky number functions should be improved.

- 64-bit max lucky index should be improved if possible

OEIS lucky numbers:
  - add program
  - add link to file

- OEIS https://oeis.org/A161170
  We can use approx and limits to create a small window containing the
  crossover.  Then use forfactored to find the exact point, or binary search.
  mmpu '$n = 1e10; while (1) { $c=prime_count_approx($n); $C=almost_prime_count_approx(8,$n); if ($c < $C) { die "$n $c $C\n"; } $n = int($n * 1.001); }'

- make GMPFE.pm.  No validation, just call GMP and objectify.
  Use _gmpcall consistently to prevent auto-vivify
  XS should call it if possible.  Maybe gmpobj?
  PPFE should direct to it after validation.
  PP.pm should remove the first-thing GMP checks
  included by PPFE and in the XS call.  Not by default.
  See about moving more GMP related things to that file.
  If/when the GMP backend finally does objects, remove it all.

- faster next_prime_power / prev_prime_power?
- PP next_prime_power, prev_prime_power: skip evens.

- prime sum bounds and approx
  https://arxiv.org/pdf/2309.16007.pdf

- prime_count_cache.c:
  - direct cache should not skip 0,1.  Removes a test
  - try cache with wheel30 instead of odds.  240/128 fewer entries.

- Better inverse interpolation in PP.

- test Brent and/or Chandrupatla with inverse interp
  https://www.embeddedrelated.com/showarticle/855.php
  also see May 2025 email about this

- xt validation test for monotonic approx/bounds.

- is_congruent_number

  A006991
  tmmpu 'forsquarefreeint { say if ($_%8) < 4 && is_congruent_number($_) } 1e7;' >/tmp/prim-cn.txt

  A274264:  tmmpu 'forsquarefreeint { $n++ if ($_%8)>4; } 1e7; say $n;'

  A274403:
  3039648+213318
  or:  mpu 'forsquarefreeint { $n++ if is_congruent_number($_) } 1e7; say $n;'

  A290449:
  mpu 'sub a { vecfirst { is_congruent_number($_*$_[0]) } (1,2,3,5); } say "$_ ",a($_) for 1..100000'

- R\'edei symbol, Borromean primes
  https://arxiv.org/pdf/2403.17957.pdf
  http://www.rnta.eu/4MSRNTA/PROC4MSRNTA.pdf
  https://arxiv.org/pdf/1806.06250.pdf
  https://www.researchgate.net/publication/367529395_Primes_knots_and_periodic_orbits
  (p,q,r), each 1 mod 4, (p|q)=(p|r)=(q|r)=1, [p,q,r]=-1
  See misc/borromean-primes.pl

- cornacchia negative d?

- nonpowerfree:    is_npf, mpf_count, npf_sum, nth_npf
- pf_numbers(lo,hi,k), npf_numbers(lo,hi,k)

- consider moving leaf PP functions to new file, so PP loads faster

- find a proper BigInt workaroud for binomial with bigint n.  bnok
  powersum uses binomial
  stirling uses binomial
  falling_factorial and rising_factorial uses binomial for m >= 250
  subfactorial uses binomial
  can use falling_factorial and factorial

- factor_range.  Get rid of N[i].  Always complete sieve.
  Get the last factor when we return the value in next (n-vecprod(factors).

- rewrite parts of RandomPrimes using our math
- remove Math::BigInt assumptions
- rewrite _random_prime using our math functions

- is there any desire for perfect power count on negative ranges?

- Look at things that loop over floor(n/i).
  https://math.stackexchange.com/a/1740370/117584

== almost_prime_count bounds verify
   DONE 2^20 - 2^32 complete
   DONE 2^32 - 2^37 complete 2.5+5+11+23+46 hours
        2^37 - 2^38 k=2-63 (est 4 days)
        2^38 - 2^39 k=2-63 (est 8 days)
        2^39 - 2^40 k=2-63 (est 16 days)

== almost_prime_count_approx revisit weighting (e.g. 6,1<<56)
   Perhaps use one or two linear interpolations for weight.
   Different behavior for high k

- omega primes, work on figuring out formula for omega_prime_count
= omega primes construction, single call routine and construct primes
  tmmpu 'omega_primes(8,powint(10,14), powint(10,14) + 1e6)'

- A253595, should clean this code:
  mmpu '@cyc = grep { is_cyclic($_) } 0..10000; for $i (3..100) { $n = $cyc[$i-1]; for $t (1..1e10) { next unless is_carmichael($n*$t); say "$i $n ",$n*$t; last; } }'

- we should have an extra debug mode for asserts.

- C99 designated initializers

- consider inversions, see 22 Apr 2025 email

- we could split reading arrays into signed and unsigned parts, possibly
  putting them into the same array front/back.

- trizen sigma_sum, pillai_sum, etc.

- more efficient kth_farey in PP

= OEIS sequence for omega k=3 counts for 10^i

= taupower
= rtau

= chacha update to rfc8439
- more tests
- compare with https://github.com/DavyLandman/portable8439
               https://github.com/smuellerDD/chacha20_drng

- chacha
  - init_context
  - neon+sse
  - refill return have
  - refill not zero key from buf

- congruent_numbers
  - permuted factors

= lobby email
  - almost_prime_count ranged

- totient sum, look at paper from Sukhoy and Stoytchev (s41598-021-99545-w.pdf)
  Better:  https://arxiv.org/pdf/2506.07386
  The lucasaugustus method is slower than ours until 10^16 where the memory
  difference heavily favors the new method.
  See https://github.com/lucasaugustus/oeis/A064018.py

- moebius_iterator
  generate a batch, dole it out, make the next, etc.
  primes to sqrt(n)
  Also test it using a sliding window so we generate monolithic
  see how that performs vs. range_moebius or windowed range_moebius

- consider my_svuv redefined, like in validate_int.  Call sv_2uv if needed.
  See https://github.com/Perl/perl5/issues/22653
  We avoid this most of the time already.

- harmfrac for negative n
  https://mathworld.wolfram.com/HarmonicNumber.html

- default bigint to better package
- revisit PP factorial

- t/81-bignum should try different bigint classes

- waring_goldbach(n,k,t)  {p1}^k + {p2}^k + ... + {pt}^k

- add pillai composite sequence to OEIS
  mpu 'forcomposites { say ++$n," $_"  if is_pillai($_) } 100000;'

- rosettacode Calin-Wilf needs this line:
     splice @f, 0, 1, 1, $f[0]-1 if $#f & 1;

- nth_ramanujan_lower, should be redone like upper.

- ramanujan_prime_count calls ramanujan_primes, and
  ramanujan_primes generates all primes up to high

- primality
  https://arxiv.org/pdf/2411.01638  (Pell Cubic)
  https://eprint.iacr.org/2025/2083.pdf (BFW extension to BPSW, SuperBFW)
  https://arxiv.org/pdf/1908.06964 (Phatak, similar to Khashin)
  1. Different QNR selection (faster?)
  2. Adds Euler test
  3. Khashin:  (1+q)^n = 1-q  or (2+q)^n = 2-q
     Phatak:   (1+q)^n - 1 - q^n = 0  =>  (1+q)^n = 1 + q^n
  https://arxiv.org/pdf/2505.02167
- sqrtmod_prime
  https://eprint.iacr.org/2025/2083.pdf

- move to 5.10.1 minimum

- second parameter for Real functions, and/or default precision.
  Get rid of looking at input accuracy.

- put all real funcs in separate Real.pm file.
  ei       ExponentialIntegral
  li       LogarithmicIntegral
  lambertw LambertW
  riemannr RiemannR
  zeta     RiemannZeta
  plus Pi(n) and Schoenfeld distance

- what should irand64() do on a 32-bit Perl?

- I don't think 53-real.t will pass on NV = float machine

- keyval.h sizes should be size_t, not long/UV.

- Some more functions to add to the GMP backend:
  fubini, partitionsq, vecprefixsum

- XS.xs call/return helpers need work to handle stack correctly.  The caller
  shouldn't have to worry about how many input args there were.  There's also
  cruft from the system evolving over years.
  = RETURN_BIGINT_STR(str, len) — collapses stack, calls _to_bigint, returns.
    Replaces the RETURN_128 bigint branch and CALLROOTSUB_ONE_SCALAR pattern
  = OBJECTIFY_RESULT(input_sv) — revamped objectify_result.
    Normalizes PL_stack_sp before calling out to Perl

- PP znlog needs pollard rho
  tmmpu 'prime_set_config(bigint=>Math::GMPz,verbose=>2); say znlog(13, 17, "10000000000000000000000000117")'

- In PP, we register callbacks to free cached data.  We need something like
  that for the C code.

- nth_lucky bounds:  https://arxiv.org/pdf/2604.07142

- Better psi(x) and Riemann(psi(x))
  https://github.com/kimwalisch/primecount/pull/112/changes#top
  https://doi.org/10.48550/arXiv.1410.1083

- toint for bigints does a redundant second parse.
  split out into something like sv=xs_to_canonical_bigint(sv)

- Check PP validation on 5.6.2.

- xops in priority:
  urandomm urandomr
  is_square
  is_congruent
  kronecker
  addmod submod mulmod
  is_prime
  sqrtint
  absint
  negint

- srand should perhaps do better input parsing.

- put the secure_bzero somewhere to export.  We use it in 3 files.

- Destructors for XS iterators, to cleanup before croak.

- permtonum better algorithm (maybe for GMP).
- numtoperm performance.  Maybe split native/all-bigint.

- setbinop: Consider pushing any SV that doesn't go into the iset into a
  separate list.  If that list is non-empty at the end, push the iset plus
  the list into a call to PP toset.  Or better setinsert(iset,list).

- make a hash.h to hold our UV hash that is used in multiple places.
  E.g. ds_iset.c.

- @v = intrange($a, $b)
  use in allrootmod, some other places

- factor_light(n) [rethink the name], where it's like factor but some
  values might be composite.  The idea is to spend a very little time
  to find small factors.  Maybe we want a difficulty setting but I wouldn't
  go too far there.  Perhaps just "very light", "light", "mild" or even just
  "very light" and "mild" as the choices.  I don't think we can safely do it
  in this code, but we could imagine a "time limit" argument.
- Also consider the step method that GMP's ECPP uses.

- GMP using ChaCha20.  See misc/gmp-using-chacha-plan.txt.

- dispatch_external R_FLOAT support:  add R_FLOAT type, dispatch_external
  calls xs_to_float to do _update_to_float on result.
  = Pi, bernreal, harmreal, li, ei, riemannr, lambertw
  RiemannZeta cannot be directly done
  Naming conversions will be an issue

- inverse_sigma0 in PP should have a generator method. e.g. (5040,1<<40)

- tstbit(n,k) (bitget, bit_test).
- scan1(n,start) / scan0(n,start)

- forpart new option order => 'revlex' to match newer Pari
- partitions_list function to return all partitions

- sidef inverse_count is very nifty.
  https://github.com/trizen/sidef/commit/cb4094f8

- is_power deprecate k=undef, deprecate k=0.

- factor128: Huchalla positive-inverse REDC

- trialf128.  Maybe live in factor128.

- string trial factor, should bump 'limit' to 5 (see docs, always find 2/3/5).

- PP sum_primes $high <= ~0  =>  Mmulint(2,Msqrtint($high)) < SINTMAX

- we ought to have a factor64 with full support given HAVE_UINT64
  this really only benefits 32-bit.  holf32 and tinyecm64 support.
  otherwise it's nice code rearrangement, but also adding primality code.

- send mail about Sort::DJB (av_fetch/av_store overhead)
  we're only testing vs the generic version, not the SIMD one.
