Assuming ATTRIBUTE_PURE means that the function must return,
a few more functions that should not be declared with ATTRIBUTE_PURE.
The GCC manual (and even the C23 standard, with [[reproducible]])
is not clear about this, and it’s better to be safe.
* src/bignum.h (mpz_get_d_rounded):
* src/lisp.h (bignum_to_double):
No longer pure, as it does not return if memory is exhausted.
* src/fns.c (Fproper_list_p): No longer pure, as it does not
return if the user quits.
* src/gnutls.c (Fgnutls_errorp): No longer pure, as it does not
return if it runs into an eassert failure in XSYMBOL_WITH_POS via EQ.
* src/lisp.h (bignum_to_intmax, bignum_to_uintmax, bignum_bufsize):
No longer pure, as it does not return if it runs into an
eassert failure in XBIGNUM via xbignum_val.

Problem for mpz_get_d_rounded reported by Helmut Eller in:
https://lists.gnu.org/r/emacs-devel/2025-11/msg00795.html
* lib-src/make-docfile.c (DEFUN_pure): New constant.
(write_globals, scan_c_stream): Support "attributes: pure".
* src/bignum.h (mpz_get_d_rounded):
* src/data.c (Fsymbolp, Fmodule_function_p, Fintegerp, Fnumberp):
* src/lisp.h (bignum_to_double, bignum_to_intmax)
(bignum_to_uintmax, bignum_bufsize):
Now pure, not const, since they depend on current state.
For example, Fsymbolp now inspects symbols_with_pos_enabled,
and the bignum functions inspect bignum contents in memory.
* src/data.c (Feq):
* src/xfaces.c (Fface_attribute_relative_p):
No longer const, since they might abort when debugging.
* src/pdumper.h (pdumper_object_p, pdumper_cold_object_p)
(pdumper_find_object_type, pdumper_object_p_precise):
These are not const functions.  But there is no need to declare
them to be pure, either, as they’re inline so the compiler can
figure it out.

Run "TZ=UTC0 admin/update-copyright".

* src/bignum.h (FASTER_BIGNUM): New macro.
(mpz_set_intmax, mpz_set_uintmax): Optimize only if FASTER_BIGNUM.
Also, use ckd_add to test for overflow instead of doing it by hand.

* src/bignum.c (get_random_limb, get_random_limb_lim)
(get_random_bignum): New functions, for more-efficient
generation of random bignums without using Frem etc.
* src/fns.c (get_random_fixnum): New function.
(Frandom): Use it, and get_random_bignum.
Be consistent about signalling nonpositive integer arguments;
since zero is invalid, Qnatnump is not quite right here.
* src/sysdep.c (get_random_ulong): New function.

Run "TZ=UTC0 admin/update-copyright".

Instead of doing GMP by hand, use the Gnulib libgmp module.
* .gitignore: Add lib/gmp.h.
* admin/merge-gnulib (GNULIB_MODULES): Add libgmp.
* configure.ac (GMP_LIB, GMP_OBJ): Remove.  Gnulib uses the name
LIB_GMP, so all uses changed.  All uses of GMP_OBJ removed.
(HAVE_GMP): Set this from Gnulib’s variables.
* lib/gnulib.mk.in, m4/gnulib-comp.m4: Regenerate.
* lib/mini-gmp-gnulib.c, lib/mini-gmp.c, lib/mini-gmp.h, m4/libgmp.m4:
New files, copied from Gnulib.
* src/bignum.h, test/data/emacs-module/mod-test.c:
Include gmp.h unconditionally.
* src/mini-gmp-emacs.c, src/mini-gmp.c, src/mini-gmp.h:
Remove.  This moves these files from src to lib, and
updates them to the current GMP version.
* test/Makefile.in (GMP_H): New macro.
($(test_module)): Use it to decide whether to compile
mini-gmp-gnulib.c too.