A bug was introduced in 3480230e: The divide-and-conquer method multiplies
with a power of the base, but this power is one too much if there is a
decimal point. This may happen because digits_to_I(...) is also called
from read_float(...). As a result, the number containd spurious zeros
(in the base used for reading it).
Thanks to Thomas Luthe <tluthe@physik.uni-bielefeld.de>.
The previous patch cb76b5eb used 4-byte word alignment.
Thanks to Peter Green and Hector Oron for pointing out that
8-byte alignment should be used (cf. Debian bug #757623).
Yunquiang Su <wzssyqa@gmail.com> proposed to use 64 bit [su]intD and
long [su]intC on mips64 and mips64el.
While at it, cleaned up the AArch64 definitions from cb76b5eb.
While building Fedora 21 for AArch64 we found out that your software is
not building for us. Attached is a fix which adds this architecture in
all required places.
Reading leading '0' characters in integers of base 2, 4, 8, 16, or 32
could result in a NUDS with leading zero digits. This is against the
rules. The result was a misbehaving cl_I down the road.
Thanks to Morgan Deters <mdeters@cs.nyu.edu> of the CVC4 team.
CLN fails to build on s390x and sparc64 as it detects them as 32-bit
architectures, due to __s390__ and __sparc__ matching respectively both
s390 and s390x, and both sparc and sparc64. The patch below fixes the
issue. (Reported as Debian bug#639494.)
Thanks to Aurelien Jarno <aurel32@debian.org>.
The routine scale_float(cl_LF, cl_I) in file cl_LF_scale_I.cc had not
been adapted to 64-bit exponents on systems where uintD is 32 bit.
Reported by Michael Miller <millermj@lemoyne.edu>.
This patch works around problems MSVC has with extern "C" declarations
inside namespace cln. As a result, it should work with MS 32-bit compiler
version 16.00.30319.01.
Thanks to Jan Rheinländer <jrheinlaender@gmx.de>.
When we extended immediate types on 64-bit platforms back for CLN
1.2.0, we broke the assumption that integers 2^53..2^55-1 are
bignums. This lead to segfaults when approximating a non-integer cl_RA
to double, as reported by Igor Khavkine <igor.kh@gmail.com>.
Recent GMP multiplication routines (mpn_mul) are somewhat faster than CLN's.
For huge operands, this is due to better tuned FFT. For medium-sized operands
it's due to more algorithms (Toom 3-way.) For small operands, I couldn't find
a difference. Let's just use mpn_mul if it's available.
For huge binsplit sums, this can half the runtimes, on amd64.
Automake adds all kinds of subdirectories to -I, the list of directories
to be searched for header files. However, CLN's header files are not
supposed to be included without the cln/ subdirectory, so adding
-Iinclude/cln can lead to conflicts with other header files. (The
problem at hand was a conflict between CLN's string.h and the C
library's string.h).
Use std::iostream functions instead of EOF and cl_EOF macros.
This enables us to get rid of the freadchar and funreadchar functions.
While at it, removed the extra signatures introduced for CLN-1.2.0
compatibilty. After all, the soname has been bumped.
Also updated largest known Mersenne prime in examples/perfnum.cc.
Richard Kreckel
Marcin Juszkiewicz
Daniel Schepler
Alexei Sheplyakov
Alexei Sheplyakov