NetBSD/external/lgpl3/mpfr/dist/BUGS
mrg 299c6f0c6b import mpfr 4.0.1. main changes since 3.1.5 are:
Changes from version 4.0.0 to version 4.0.1:
- Bug fixes (see ChangeLog file), in particular in mpfr_div_ui, which
  could yield an incorrectly rounded result to nearest when using
  different precisions; this bug had been present since the introduction
  of mpfr_div_ui, and in MPFR 4.0.0, it was affecting mpfr_div too.

Changes from versions 3.1.* to version 4.0.0:
- Partial support of MPFR_RNDF (faithful rounding).
- New functions: mpfr_fpif_export and mpfr_fpif_import to export and import
  numbers in a floating-point interchange format, independent both on the
  number of bits per word and on the endianness.
- New function mpfr_fmodquo to return the low bits of the quotient
  corresponding to mpfr_fmod.
- New functions mpfr_flags_clear, mpfr_flags_set, mpfr_flags_test,
  mpfr_flags_save and mpfr_flags_restore to operate on groups of flags.
- New functions mpfr_set_float128 and mpfr_get_float128 to convert from/to
  the __float128 type (requires --enable-float128 and compiler support).
- New functions mpfr_buildopt_float128_p and mpfr_buildopt_sharedcache_p.
- New functions mpfr_rint_roundeven and mpfr_roundeven, completing the
  other similar round-to-integer functions for rounding to nearest with
  the even-rounding rule.
- New macro mpfr_round_nearest_away to add partial emulation of the
  rounding to nearest-away (as defined in IEEE 754-2008).
- New functions mpfr_nrandom and mpfr_erandom to generate random numbers
  following normal and exponential distributions respectively.
- New functions mpfr_fmma and mpfr_fmms to compute a*b+c*d and a*b-c*d.
- New function mpfr_rootn_ui, similar to mpfr_root, but agreeing with the
  rootn function of the IEEE 754-2008 standard.
- New functions mpfr_log_ui to compute the logarithm of an integer,
  mpfr_gamma_inc for the incomplete Gamma function.
- New function mpfr_beta for the Beta function (incomplete, experimental).
- New function mpfr_get_q to convert a floating-point number into rational.
- Dropped K&R C compatibility.
- Major speedup in mpfr_add, mpfr_sub, mpfr_mul, mpfr_div and mpfr_sqrt when
  all operands have the same precision and this precision is less than twice
  the number of bits per word, e.g., less than 128 on a 64-bit computer.
- Speedup by a factor of almost 2 in the double <--> mpfr conversions
  (mpfr_set_d and mpfr_get_d).
- Speedup in mpfr_log1p and mpfr_atanh for small arguments.
- Speedup in the mpfr_const_euler function (contributed by Fredrik Johansson),
  in the computation of Bernoulli numbers (used in mpfr_gamma, mpfr_li2,
  mpfr_digamma, mpfr_lngamma and mpfr_lgamma), in mpfr_div, in mpfr_fma
  and mpfr_fms.
2018-09-04 05:02:00 +00:00

73 lines
3.3 KiB
Plaintext

Copyright 1999, 2001-2018 Free Software Foundation, Inc.
Contributed by the AriC and Caramba projects, INRIA.
This file is part of the GNU MPFR Library.
The GNU MPFR Library is free software; you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation; either version 3 of the License, or (at your
option) any later version.
The GNU MPFR Library is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
License for more details.
You should have received a copy of the GNU Lesser General Public License
along with the GNU MPFR Library; see the file COPYING.LESSER. If not, see
http://www.gnu.org/licenses/ or write to the Free Software Foundation, Inc.,
51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
##############################################################################
Known bugs:
* The overflow/underflow exceptions may be badly handled in some functions;
specially when the intermediary internal results have exponent which
exceeds the hardware limit (2^30 for a 32 bits CPU, and 2^62 for a 64 bits
CPU) or the exact result is close to an overflow/underflow threshold.
* Under Linux/x86 with the traditional FPU, some functions do not work
if the FPU rounding precision has been changed to single (this is a
bad practice and should be useless, but one never knows what other
software will do).
* Some functions do not use MPFR_SAVE_EXPO_* macros, thus do not behave
correctly in a reduced exponent range.
* Function hypot gives incorrect result when on the one hand the difference
between parameters' exponents is near 2*MPFR_EMAX_MAX and on the other hand
the output precision or the precision of the parameter with greatest
absolute value is greater than 2*MPFR_EMAX_MAX-4.
Potential bugs:
* Possible incorrect results due to internal underflow, which can lead to
a huge loss of accuracy while the error analysis doesn't take that into
account. If the underflow occurs at the last function call (just before
the MPFR_CAN_ROUND), the result should be correct (or MPFR gets into an
infinite loop). TODO: check the code and the error analysis.
* Possible bugs with huge precisions (> 2^30) and a 32-bit ABI, in particular
undetected integer overflows. TODO: use the MPFR_ADD_PREC macro.
* Possible bugs if the chosen exponent range does not allow to represent
the range [1/16, 16].
* Possible infinite loop in some functions for particular cases: when
the exact result is an exactly representable number or the middle of
consecutive two such numbers. However for non-algebraic functions, it is
believed that no such case exists, except the well-known cases like cos(0)=1,
exp(0)=1, and so on, and the x^y function when y is an integer or y=1/2^k.
* The mpfr_set_ld function may be quite slow if the long double type has an
exponent of more than 15 bits.
* mpfr_set_d may give wrong results on some non-IEEE architectures.
* Error analysis for some functions may be incorrect (out-of-date due
to modifications in the code?).
* Possible use of non-portable feature (pre-C99) of the integer division
with negative result.