Changes commited to PARI GIT during the last two weeks

add a CHANGES entry for numbpart improvements

+ minor improvement to memory use

1	0	CHANGES
1	1	src/modules/part.c

rename L -> SelbergL and compute cosines via roots of unity with caching

- compute cos((6l+1)/(6q) * Pi) as real(r^(1+6l)), r = e(1/(12q))

- compute r^(1+6l) as r^(1+6lastl) * R^d, d = l - last, R = r^6

- compute the successive R^d via a naive addition chain for d, caching
  values. N.B. There are few values of d, with collisions, this improves
  further on caching the R^d by reusing all powers. One could implement
  other improvements, sliding 2^k window or Yao's multipowering,
  but d is "small" and almost all the time is spent computing e(1/(12q)).

N.B there's an optimal formula for L(n,q) in terms of the factorization of q.
Not worth the complication for now.

Before:
    numbpart(10^7); \\ time = 45 ms.
    numbpart(10^8); \\ time = 181 ms.
    numbpart(10^9); \\ time = 4,050 ms.
    numbpart(10^10);\\ time = 56,031 ms.
    numbpart(10^11);\\ time = 54min, 41,505 ms.

After:
    numbpart(10^7); \\ time = 43 ms.
    numbpart(10^8); \\ time = 147 ms.
    numbpart(10^9); \\ time = 1,793 ms.
    numbpart(10^10);\\ time = 17,593 ms.
    numbpart(10^11);\\ time = 8min, 12,924 ms.

44	15	src/modules/part.c

psi: remove unused argument

6	6	src/modules/part.c

2- abs(t_QUAD) would not return exact answer even if possible

For instance
  abs(3 + 4*I) -> 5
but
  abs(3 + 4*quadgen(-4)) -> 5.000000000000
Now the latter is 5 as well.

1	0	CHANGES
14	3	src/basemath/gen2.c
11	2	src/functions/transcendental/abs
5	0	src/test/32/quad
5	0	src/test/in/quad

cgetr(realprec()) -> rtor_lg

2	3	src/headers/pariinl.h

numbpart: reorganize computation + allow n < 2^54

- factor out sqrt(q) term in Psi and sqrt(q/3) in L: avoid square roots
  and replace full precision multiplication by simple multiplication by q

- avoid low precision exponentials: with Selberg formula the neglectible
  sum is actually 0

- precompute pi / 6*q in L

- simplify pinit

- rename estim -> log2pn and return a double

- bound 10^15 was somewhat arbitrary, increase to 2^54. Both would require
  hundreds of years to compute using current implementation.

Asymptotic complexity remains the same (and bad, more than quadratic in
output size instead of almost linear). But 25% improvement for small values,
with diminishing returns.

Before:
  numbpart(10^7); \\ time = 54 ms.
  numbpart(10^8); \\ time = 289 ms.
  numbpart(10^9); \\ time = 5,841 ms.
  numbpart(10^10);\\ time = 1min, 11, 975 ms.
  numbpart(10^11);\\ time = 1h, 9min, 14,433 ms

After:
  numbpart(10^7); \\ time = 45 ms.
  numbpart(10^8); \\ time = 181 ms.
  numbpart(10^9); \\ time = 4,050 ms.
  numbpart(10^10);\\ time = 56,031 ms.
  numbpart(10^11);\\ time = 54min, 41,505 ms.

6	2	src/functions/combinatorics/numbpart
60	71	src/modules/part.c
4	4	src/test/32/partition
3	1	src/test/in/partition

fix make test-kernel: link with dummy Z_ispow2

1	0	src/test/dummy.c

rdivsi, rdivis, rdivss: use cgetg(l, t_REAL)

9	6	src/kernel/none/level1.h

cgetc: improve + fix doc

2	2	doc/usersch5.tex
5	1	src/headers/pariinl.h

update trans bench after commit cb6dca75d7

incgam(4,1,6) incgamm uses 2 more words

1	1	src/test/in/trans

stack corruption in mplog1p due to lg / prec confusion

fixlg argument is a length, not a prec

1	1	src/basemath/trans1.c

cleanup exp1r_abs

17	16	src/basemath/trans1.c

rpowuu: cgetr -> cgetg(,t_REAL)

1	2	src/basemath/trans1.c

bessel: confusions between lg and prec

+ 1 => + EXTRAPREC64

5	4	src/basemath/trans3.c

final prec2nbits cleanups

2	3	doc/usersch5.tex
10	10	src/language/sumiter.c

further prec2nbits cleanups

9	9	src/basemath/rootpol.c

fix bestappr doc and comment in code

4	2	src/basemath/arith1.c
4	1	src/functions/number_theoretical/bestappr

rootpol.c: prec2nbits cleanups

Also fix incorrect paranoia setting in mygprec
  if (bit < 0) bit = 0
to
  if (bit < BITS_IN_LONG) bit = BITS_IN_LONG

26	30	src/basemath/rootpol.c

simple rewrite of prec2nbits uses (= identity now)

3	4	src/basemath/alglin2.c
12	14	src/basemath/bern.c
15	15	src/basemath/bibli2.c
21	23	src/basemath/elltrans.c
13	14	src/basemath/grossenchar.c
41	45	src/basemath/hypergeom.c
7	6	src/basemath/lambert.c
10	10	src/basemath/lerch.c
4	4	src/basemath/lll.c
20	21	src/basemath/mftrace.c
3	5	src/basemath/random.c
65	82	src/basemath/trans1.c
38	43	src/basemath/trans2.c
71	84	src/basemath/trans3.c
14	16	src/basemath/zetamult.c

remove uses of prec2nbits_mul

2	3	src/basemath/bern.c
3	4	src/basemath/bibli1.c
2	3	src/basemath/elliptic.c
3	3	src/basemath/elltrans.c
1	1	src/basemath/hypergeom.c
1	1	src/basemath/mftrace.c
1	1	src/basemath/trans1.c
3	3	src/basemath/trans2.c
11	11	src/basemath/trans3.c

divri: handle divisions by powers of 2 more efficiently

+ document how to input an exact floating point number (in binary).

8	5	doc/usersch2.tex
1	0	src/kernel/gmp/mp.c
1	0	src/kernel/none/mp.c

doc clarifications

9	4	doc/usersch2.tex

absrtostr_width_frac: confusion between lg / prec [similar to f3d63fab9]

4	4	src/language/es.c

deprecate real_1_bit: use real_1

5	0	CHANGES
1	0	COMPAT
2	5	doc/usersch5.tex
1	1	src/basemath/rootpol.c
2	2	src/basemath/trans1.c
1	1	src/basemath/trans2.c
1	1	src/basemath/trans3.c
0	1	src/headers/paridecl.h
1	0	src/headers/pariold.h
1	1	src/kernel/none/add.c
1	3	src/kernel/none/level1.h
8	8	src/language/sumiter.c

fix mpatanh: confusion EXTRAPREC64 -> BITS_IN_LONG + wrong exponent bound

2	2	src/basemath/trans2.c

2- [libpari] rename real_0_bit -> real_0_expo

The argument is not a bitprec, but an exponent. Inconsistent naming with
- real_1_bit (argument is actually a bitpec)
- real_0 (argument is actually a prec)

And inconsistent semantic with real_0(prec) = real_0_bit(- prec).

1	0	CHANGES
6	0	COMPAT
5	3	doc/usersch5.tex
2	2	src/basemath/bibli2.c
1	1	src/basemath/lfun.c
1	1	src/basemath/mftrace.c
2	2	src/basemath/random.c
9	9	src/basemath/rootpol.c
13	13	src/basemath/trans1.c
8	8	src/basemath/trans2.c
1	1	src/headers/paridecl.h
4	4	src/headers/pariinl.h
2	1	src/headers/pariold.h
2	2	src/kernel/gmp/mp.c
4	4	src/kernel/none/add.c
2	2	src/kernel/none/level1.h
2	2	src/kernel/none/mp.c
11	11	src/kernel/none/mp_indep.c
1	1	src/language/anal.c

absrtostr: confusion between lg / prec

lx = lg(x) was changed to lx = realprec(x) without updating occurences of
lx + 1 or lx++

10	10	src/language/es.c

fix prec / length confusions

12	10	doc/usersch4.tex
38	22	doc/usersch5.tex
11	10	src/functions/transcendental/HEADER

update 32-bit benches (getheap) after 859750f9a

EXTRAPRECWORD -> EXTRAPREC64 causes mplog2 to cache a value with one
more word of accuracy

1	1	src/test/in/analyz
1	1	src/test/in/elliptic
1	1	src/test/in/number
1	1	src/test/in/polyser
1	1	src/test/in/program
1	1	src/test/in/sumiter

2- [libpari] fractor_lg, itor_lg, rdivii_lg, rtor_lg, stor_lg, utor_lg

The experimental change of 'prec' parameter from a word length to a bit
precision has created 2 paradigms 'lg' vs. 'prec' and conversions
between each all over the code.

Both still coexist in ugly ways, with conversions that were introduced
to ensure backward compatibility through search and replace. This commit
introduces variants of conversion routines accepting 'lg' parameters in
case it's available.

Goal: avoid direct or indirect uses of consructions like nbits2lg(lg2prec(l)).
If we have an 'lg' l we might as well use it.

For instance

  r = cgetr(l, t_REAL);

instead of

  r = cgetr(lg2prec(l));

cgetr doesn't need its full 'prec' parameter and only needs an 'lg'
(it performs an nbits2lg on its argument)

Or tests comparing lg2prec(lg(x)) against some constant 'prec' where we
might as well compare lg(x) with some 'lg'.

Design choice: introducing a constant EXTRALG64 = EXTRAPREC64 / BITS_IN_LONG
(= 1 or 2) would simplify some code. Don't do this: the default is to
use 'prec', unless 'l' is cleaner and more efficient.

1	0	CHANGES
36	13	doc/usersch5.tex
1	1	src/basemath/F2x.c
4	4	src/basemath/trans1.c
18	17	src/basemath/trans2.c
6	0	src/headers/paridecl.h
12	10	src/kernel/none/add.c
27	10	src/kernel/none/level1.h
15	14	src/kernel/none/mp.c
4	4	src/kernel/none/mp_indep.c
6	5	src/language/intnum.c

typo

1	1	src/basemath/mellininv.c

add proto.c to src/desc/.gitignore

1	0	src/desc/.gitignore

Kderivsmallinit: simplify code

2	6	src/basemath/mellininv.c

fix conceptual difference 32/64-bit due to EXTRAPRECWORD/EXTRAPREC64

The logic of EXTRAPRECWORD is to add "a few" guard bits,actually BITS_IN_LONG,
to cater for terms bounded by log2(ulong).

The logic of incrprec(bit) or bit + EXTRAPREC64 is to do it in an
architecture-independent way (add 64 bits no matter what).

This commit replaces EXTRAPRECWORD by EXTRAPREC64. The odd speed gain on
32-bit architectures is not worth the lack of consistency and potential
for bizarre bugs (further differences in behaviour between 32/64 bits).

EXTRAPRECWORD wasn't actually used consistently or frequently (15 occurences)
and it wasn't documented.

CAVEAT: This is about the precreal ~ bitprec world. A parallel issue
in the 'lg' world is lgprec + 1 or lx++ instead of
  lgprec + EXTRAPREC64 / BITS_IN_LONG
This is much more frequent than EXTRAWORD and will be dealt with later.
For historical reasons a 'bitprec' is actually a "bit accuracy" whereas
a 'prec' is a multiple of BITS_IN_LONG (making the range of attainable
precs depend on the architecture. This is an issue: it should be a
multiple of 64 always.

N.B. It is worth keeping 2 notions of "bit precision" and "mantissa size
in words" even though the latter can be inferred from the former.

1	1	src/basemath/base1.c
3	3	src/basemath/dirichlet.c
2	2	src/basemath/ellanal.c
1	1	src/basemath/ellrank.c
1	1	src/basemath/hyperell.c
1	1	src/basemath/mellininv.c
3	3	src/basemath/trans1.c
3	3	src/basemath/trans2.c
0	1	src/headers/parigen.h

expQ: fix conceptual bug introduced 499574bea

lx++ for t_REAL is achitecture dependent, one should use
  lx += EXTRAPREC64 / BITS_IN_LONG
This occurs elsewhere in the code and will be fixed in other commits.

1	1	src/basemath/trans1.c

Replace gerepilecopy by gc_GEN

2	2	src/basemath/buch2.c
1	1	src/basemath/ellanal.c

cleanup zetamult_i

14	10	src/basemath/zetamult.c

cleanup expQ

9	9	src/basemath/trans1.c

update bench after bestappr rewrite (commit 96846c82)

3	3	src/test/32/gchar-lfun

inefficient use of cgetr when lg would be available

9	7	src/basemath/lll.c
4	4	src/kernel/none/level1.h
12	7	src/modules/stark.c

doc typos related to lg <-> prec conversions

10	10	doc/usersch5.tex

doc typos

4	5	doc/usersch5.tex

reorganize block of functions related to lg <-> accuracy conversions

14	10	src/headers/pariinl.h

cleanup logr_abs

14	15	src/basemath/trans1.c

RgM_hess: cleanup prec use

16	15	src/basemath/alglin2.c

replace nbits2ndec uses by prec2ndec

was always used in precreal context.

1	1	doc/usersch5.tex
1	1	src/basemath/ellanal.c
7	8	src/language/default.c
1	1	src/language/es.c

bestappr_frac: add diagnostic in GC at \gm2

2	1	src/basemath/arith1.c

fix bestappr(t_REAL or t_FRAC, B) issues [#2688]

- rewrite bestappr_real in terms of bestappr_frac: this allows the
  accuracy to drop down and is one order of magnitude faster. For instance
    localbitprec(80000); bestappr(Pi, 2^40000);
  goes down from 6,242 ms. to 122 ms.

- best approximation sometimes missed bestappr(-1/7, 6) -> 0 [instead of -1/6]

- bestappr(t_REAL) now choses B = 2^((realprec(x) + 1) \ 2) instead of
  2^realprec(x); with the latter value we might as well return the
  floating point representation of x

4	0	CHANGES
25	99	src/basemath/arith1.c
7	7	src/functions/number_theoretical/bestappr
2	0	src/test/32/bestappr
2	0	src/test/in/bestappr

Replace bit_prec uses by realprec.

Deprecate the bit_prec alias.

2	0	doc/usersch4.tex
1	1	doc/usersch5.tex
1	1	src/basemath/arith1.c
2	2	src/basemath/bit.c
1	1	src/basemath/buch2.c
1	1	src/basemath/elliptic.c
1	1	src/basemath/gen2.c
1	1	src/basemath/gen3.c
1	1	src/basemath/polarit2.c
1	1	src/basemath/trans1.c
3	3	src/basemath/trans2.c
2	2	src/kernel/none/mp_indep.c
1	1	src/modules/galois.c

bestappr cleanups

- factor out cf_updatepq in bestappr_frac and bestappr_real
- add GC in bestappr_frac
- fix GC in bestappr_real (x clobbered)

37	11	src/basemath/arith1.c

small improvement to bestappr_real

19	2	src/basemath/arith1.c

pariold.h: move #ifdef PARI_OLD_NAMES up the list

No longer define (by default) aliases for functions obsoleted more than
10 years ago.

3	2	src/headers/pariold.h

improve isint2n

3	2	src/kernel/none/halfgcd.c

bestappr_real: add GC + restrict variable scopes

12	6	src/basemath/arith1.c

update NEW for 2.19-stable

181	195	NEW

doc glitches

3	3	src/functions/number_fields/idealhnf

document QM_ishnf

2	0	CHANGES
3	0	doc/usersch5.tex

cleanup idealhnf use

- fix typo in vec_mulid
- clarify in idealhnf doc that idempotent usage (where input is already
  in HNF) is *not* deprecated. We deprecate other non-trivial "vector of
  generators" use in favor of idealfromgens
- uniformize "idempotent" usage in libpari: first idealtyp, then
  idealhnf if not already a t_MAT
- make sure idealhnf recognize matrix in HNF also over the rationals

18	0	src/basemath/ZV.c
16	11	src/basemath/base4.c
23	2	src/basemath/base5.c
14	5	src/functions/number_fields/idealhnf
1	0	src/headers/paridecl.h

idealhnf: clarify doc

17	19	src/functions/number_fields/idealhnf

typo in vec_mulid: the code was inteded to skip x[i] if D | x[i]

3	3	src/basemath/base4.c
8	1	src/test/32/ideal
1	0	src/test/in/ideal

hyperellchangecurve doc: switch confusing notations (x,y) <-> (X,Y)

4	4	src/functions/elliptic_curves/hyperellchangecurve

hyperell* docs: add a reference to hyperellchange curve

whenever a "change of coordinates" is mentioned and not explained

1	1	src/functions/elliptic_curves/hyperellauto
2	1	src/functions/elliptic_curves/hyperellchangecompose
3	2	src/functions/elliptic_curves/hyperellchangecurve
2	1	src/functions/elliptic_curves/hyperellchangeinvert
1	1	src/functions/elliptic_curves/hyperellchangepoint
1	1	src/functions/elliptic_curves/hyperellchangepointinv
1	1	src/functions/elliptic_curves/hyperellisisom

examples: openmp, thread: force compilation error if ENABLE_TLS is not set.

4	0	examples/openmp.c
4	0	examples/thread.c

ellmaninconstant: add reference to Cremona's explanation

https://johncremona.github.io/ecdata/manin.txt

1	0	src/functions/elliptic_curves/ellmaninconstant

ellmaninconstant: minor simplification

3	3	src/basemath/ellisog.c

kill underfull hboxes

8	6	doc/refcard-mf.tex

impove doc, kill overfull hbox

10	11	doc/usersch5.tex

doc improvement

2	2	src/functions/number_fields/HEADER

improve ellheegner doc

8	9	src/functions/elliptic_curves/ellheegner

doc:ellheegner: mention ellheegnertwist

10	1	src/functions/elliptic_curves/ellheegner

fix comment: reference stable version where function is obsoleted

1	1	src/headers/pariold.h

Bump date to 2026

2	2	README
1	1	doc/INSTALL.tex
1	1	doc/gp.1
1	1	doc/gphelp.1
1	1	src/gp/gp.c

Bump version to 2.19.0

3	0	CHANGES
2	2	config/version
1	1	doc/refmacro.tex
1	1	src/test/32/version

add parmatrix

6	6	doc/refcard.tex

expIPiR: don't call absrnz_equal2n on 0.0

15	10	src/basemath/elltrans.c

upowers: compute powers x^0 to x^n as documented

2	2	src/basemath/trans1.c

GCOVFLAGS: add -fprofile-update=atomic

1	1	config/get_cc

expIPiC: typo, expo(z) => expo(x)

1	1	src/basemath/elltrans.c