* libguile/numbers.c (log_of_shifted_double, scm_log10): Avoid complex
extension when the argument is a real nan.
* test-suite/tests/numbers.test: Tests for nans of either sign.
This avoids gmp aborting e.g. with (ash 1 (expt 2 37)). The new limit is
such that (ash 1 (expt 30)) is accepted but (ash 1 (expt 31)) throws.
Fixes https://bugs.gnu.org/48150
* libguile/numbers.c (ash, round-ash): As stated.
* test-suite/tests/numbers.test: Test a case known to make gmp abort before.
Reported by Marius Bakke <marius@gnu.org>
at <https://issues.guix.gnu.org/50696>.
Previously, the baseline compiler would incorrectly emit a right shift
when for, say, (ash x 2), and a left shift for (ash x -2).
* module/language/tree-il/compile-bytecode.scm (canonicalize): When Y is
negative, emit 'rsh', not 'lsh'.
* test-suite/tests/numbers.test ("ash at -O1"): New test.
Fixes <https://bugs.gnu.org/21901>.
Reported by Zefram <zefram@fysh.org>.
* libguile/numbers.c: Add another top-level 'verify' to ensure that
LONG_MIN is not a fixnum.
(scm_ash, scm_round_ash): Ensure that when the shift count is LONG_MIN,
it is not handled via the normal code path, to avoid signed overflow
when the shift count is negated.
* test-suite/tests/numbers.test: Add tests.
This is a followup to commit 011aec7e24.
When rounding, right shifting a negative integer by a huge shift count
results in 0, not -1.
* libguile/numbers.c: Add top-level 'verify' to ensure that the
assumptions in 'scm_ash' and 'scm_round_ash' are valid.
(scm_round_ash): In the case that handles huge right shifts, require
that the shift count _exceeds_ the integer length, and return 0 instead
of -1.
* test-suite/tests/numbers.test: Adjust tests accordingly.
Fixes <https://bugs.gnu.org/32644>.
Reported by Stefan Israelsson Tampe <stefan.itampe@gmail.com>.
The need for this arose because the type inferrer for 'ursh' sometimes
passes (- 1 (expt 2 64)) as the second argument to 'ash'.
* libguile/numbers.c (scm_ash, scm_round_ash): Gracefully handle several
cases where the shift count does not fit in a C 'long'.
* test-suite/tests/numbers.test: Add tests.
This bug was introduced by 35a9059250.
* module/language/cps/specialize-numbers.scm (specialize-operations):
Check that both operands are real as a condition for
specialize-f64-comparison.
* test-suite/tests/numbers.test: Add test.
* libguile/numbers.c (left_shift_exact_integer): Fix edge case where
N is -1 and count is SCM_I_FIXNUM_BIT-1 to return the most negative
fixnum. Previously this result was returned as a bignum.
* test-suite/tests/numbers.test (ash): Add tests.
* libguile/numbers.c (scm_ash): Fix (ash -1 SCM_I_FIXNUM_BIT-1) to
return a fixnum instead of a bignum.
* test-suite/tests/numbers.test (ash): Add tests.
* libguile/numbers.c (scm_numerator, scm_denominator): Handle signed
zeroes and infinities in accordance with the corresponding R6RS flonum
procedures.
* module/rnrs/arithmetic/flonums.scm (flnumerator, fldenominator):
Remove special handling of infinities.
* test-suite/tests/numbers.test (numerator, denominator): Add tests.
Convert existing tests to use 'pass-if-equal'.
* test-suite/tests/r6rs-arithmetic-flonums.test (flnumerator): Fix
broken test of (flnumerator -0.0).
Fixes <http://bugs.gnu.org/14905>.
Reported by Göran Weinholt <goran@weinholt.se>.
* libguile/numbers.c (scm_rationalize): Rewrite. Previously an
incorrect algorithm was used which failed in many cases.
* test-suite/tests/numbers.test (rationalize): Add tests.
* libguile/numbers.c (scm_num_eq_p): Fix bug comparing fractions to
infinities (reported by Göran Weinholt <goran@weinholt.se>). Fix
erroneous comment describing the logic behind inum/flonum comparison.
Use similar logic for inum/complex comparison to avoid rounding
errors. Make minor indentation fixes and simplifications.
* test-suite/tests/numbers.test (=): Add tests.
* libguile/numbers.c (INUM_LOSSLESSLY_CONVERTIBLE_TO_DOUBLE):
New macro.
(scm_i_divide2double): Use INUM_LOSSLESSLY_CONVERTIBLE_TO_DOUBLE to
determine if our fast path is safe. Previously, negative arguments
were not checked properly.
* test-suite/tests/numbers.test (exact->inexact): Add tests.
* libguile/numbers.c (exact_integer_is_perfect_square,
exact_integer_floor_square_root): New static functions.
(scm_sqrt): Use SCM_LIKELY. Add 'scm_t_inum' variable in inum case to
reduce the number of uses of SCM_I_INUM. Rename 'mpz_t' variable.
Remove unneeded sign check. Handle bignums too large to fit in a
double. Handle fractions too large or too small to fit in a
normalized double.
* test-suite/tests/numbers.test ("sqrt"): Add tests.
* libguile/numbers.c (scm_sqrt): Handle exact integers and rationals in
such a way that exact results are returned whenever possible.
* test-suite/tests/numbers.test ("sqrt"): Add tests.
* libguile/numbers.c (idbl2str): Print large numbers in scientific
notation only if the exponent is >= 7 and the least significant
non-zero digit has value >= radix^4.
* test-suite/tests/numbers.test ("number->string"): Add tests.
Fixes <http://bugs.gnu.org/13757>.
* libguile/numbers.c (idbl2str): Reimplement.
(mem2decimal_from_point): Accept negative exponents larger than
SCM_MAXEXP that produce subnormals.
(SCM_MAX_DBL_PREC): Removed preprocessor macro.
(scm_dblprec, fx_per_radix): Removed static variables.
(init_dblprec, init_fx_radix): Removed static functions.
(scm_init_numbers): Remove initialization code for 'scm_dblprec'
and 'fx_per_radix'.
* test-suite/tests/numbers.test ("number->string"): Restore tests that
previously failed. Remove comments about problems in the number
printer that are now fixed.
* libguile/numbers.c (scm_i_divide2double): New function.
(scm_i_divide2double_lo2b): New variable.
(scm_i_fraction2double, log_of_fraction): Use 'scm_i_divide2double'.
(do_divide): Removed. Its code is now in 'scm_divide'.
(scm_divide2real): Removed. Superceded by 'scm_i_divide2double'.
(scm_divide): Inherit code from 'do_divide', but without support for
forcing a 'double' result (that functionality is now implemented by
'scm_i_divide2double'). Add FIXME comments in cases where divisions
might not be as precise as they should be.
(scm_init_numbers): Initialize 'scm_i_divide2double_lo2b'.
* test-suite/tests/numbers.test (dbl-epsilon-exact, dbl-max-exp): New
variables.
("exact->inexact"): Add tests.
("inexact->exact"): Add test for largest finite inexact.
* libguile/numbers.c (scm_inexact_to_exact): Implement conversion of a
double to an exact rational without using the mpq functions.
* test-suite/tests/numbers.test (dbl-mant-dig): Simplify initializer.
(dbl-epsilon, dbl-min-exp): New variables.
("inexact->exact"): Add tests. Fix broken "2.0**i to exact and back"
test, and change it to "2.0**i to exact", to avoid use of
'exact->inexact'.
* libguile/numbers.c (scm_i_big2dbl_2exp): New static function.
(scm_i_big2dbl): Reimplement in terms of 'scm_i_big2dbl_2exp',
with proper rounding.
* test-suite/tests/numbers.test ("exact->inexact"): Add tests.
* libguile/numbers.c (left_shift_exact_integer,
floor_right_shift_exact_integer, round_right_shift_exact_integer): New
static functions.
(scm_round_ash): New procedure.
(scm_ash): Reimplement in terms of 'left_shift_exact_integer' and
'floor_right_shift_exact_integer'.
* libguile/numbers.h: Add prototype for scm_round_ash. Rename the
second argument of 'scm_ash' from 'cnt' to 'count'.
* test-suite/tests/numbers.test (round-ash, ash): Add new unified
testing framework for 'ash' and 'round-ash'. Previously, the tests
for 'ash' were not very comprehensive; for example, they did not
include a single test where the number to be shifted was a bignum.
* doc/ref/api-data.texi (Bitwise Operations): Add documentation for
'round-ash'. Improve documentation for `ash'.
Fixes <http://bugs.gnu.org/11887>.
* libguile/numbers.c (mem2ureal): New argument 'allow_inf_or_nan'.
Accept infinities and NaNs only if 'allow_inf_or_nan' is true and "#e"
is not present. Check for "inf.0" or "nan." case-insensitively. Do
not accept rationals with zero divisors.
(mem2complex): Pass new 'allow_inf_or_nan' argument to 'mem2ureal',
which is set if and only if a explicit sign was present.
* test-suite/tests/numbers.test ("string->number"): Add tests.
* libguile/numbers.c (scm_product): Avoid signed integer overflow, which
modern C compilers are allowed to assume will never happen, thus
allowing them to optimize out our overflow checks.
* test-suite/tests/numbers.test (*): Add tests.
* libguile/numbers.c (scm_angle): Check the sign of an inexact real
zero, to ensure that (angle -0.0) => pi and (angle 0.0) => 0.0.
* test-suite/tests/numbers.test (angle): Add tests, and increase
precision of tests where the angle should be pi.
* libguile/numbers.c (scm_exact_integer_sqrt): New C procedure to
compute exact integer square root and remainder.
(scm_i_exact_integer_sqrt): New Scheme procedure `exact-integer-sqrt'
from the R6RS, imported into core guile.
* libguile/numbers.h: Add prototypes.
* module/rnrs/base.scm: Remove broken stub implementation, which would
fail badly when applied to large integers.
* doc/ref/api-data.texi: Add documentation.
* doc/ref/r6rs.texi: Change documentation for `exact-integer-sqrt' to a
stub that xrefs the core docs, as is done for other operations
available in core.
* test-suite/tests/numbers.test: Add tests.
* NEWS: Add news entries.
* libguile/numbers.c (mem2decimal_from_point): Use scm_divide instead of
scm_divide2real when applying a negative exponent, to preserve
exactness in case the "#e" forced exactness specifier is present.
This fixes a bug where numeric literals such as "#e1e-5" yielded
incorrect fractions.
* libguile/numbers.c (scm_quotient, scm_remainder, scm_modulo): Accept
inexact integers as well as exact ones, as required by the R5RS.
* test-suite/tests/numbers.test (quotient, remainder, modulo): Add tests.
* libguile/numbers.c (log_of_shifted_double, log_of_exact_integer,
log_of_exact_integer_with_size, log_of_fraction): New internal static
functions used by scm_log and scm_log10.
(scm_log, scm_log10): Robustly handle large integers, large and small
fractions, and fractions close to 1. Previously, computing logarithms
of fractions close to 1 yielded grossly inaccurate results, and the
other cases yielded infinities even though the answer could easily fit
in a double. (log -0.0) now returns -inf.0+<PI>i, where previously it
returned -inf.0. (log 0) now throws a numerical overflow exception,
where previously it returned -inf.0. (log 0.0) still returns -inf.0.
Analogous changes made to `log10'.
* test-suite/tests/numbers.test (log, log10): Add tests.
Signed-off-by: Ludovic Courtès <ludo@gnu.org>
* libguile/numbers.c (scm_c_truncate): Use ceil (x) instead of
-floor (-x).
(scm_truncate_number): Implement directly instead of by checking the
sign and using scm_floor or scm_ceiling. Use scm_truncate_quotient
for fractions. Make extensible, so that new number types implemented
in GOOPS will be able to do the job more efficiently, since it is
often easier to implement truncate than floor or ceiling.
(scm_round_number): Optimize fractions case by using
scm_round_quotient. Make extensible, so that new number types
implemented in GOOPS will be able to do the job efficiently.
(scm_floor, scm_ceiling): Optimize fractions case by using
scm_floor_quotient and scm_ceiling_quotient, respectively.
* test-suite/tests/numbers.test: Add test cases.
* libguile/numbers.c (scm_euclidean_quotient, scm_euclidean_divide,
scm_centered_quotient, scm_centered_divide): Fix bug in inum/inum
case, where (quotient most-negative-fixnum -1) would not be converted
to a bignum.
(scm_euclidean_quotient): Be more anal-retentive about calling
scm_remember_upto_here_1 after mpz_sgn, (even though mpz_sgn is
documented as being implemented as a macro and certainly won't
do any allocation). It's better to be safe than sorry here.
(scm_euclidean_quotient, scm_centered_quotient): In the bignum/inum
case, check if the divisor is 1, since this will allow us to avoid
allocating a new bignum.
(scm_euclidean_divide, scm_centered_quotient, scm_centered_divide):
When computing the intermediate truncated quotient (xx / yy) and
remainder, use (xx % yy) instead of (xx - qq * yy), on the theory that
the compiler is more likely to handle this case intelligently and
maybe combine the operations.
(scm_euclidean_divide): In the bignum/inum case, we know that the
remainder will fit in an fixnum, so don't bother allocating a bignum
for it.
(scm_euclidean_quotient, scm_euclidean_remainder,
scm_euclidean_divide, scm_centered_quotient, scm_centered_remainder,
scm_centered_divide): Minor stylistic changes.
* test-suite/tests/numbers.test: Rework testing framework for
number-theoretic division operators to be more efficient and
comprehensive in its testing of code paths and problem cases.
When parsing non-real complex numbers, apply exactness specifiers on
per-component basis, as is done in PLT Scheme. For complex numbers
written in rectangular form, exactness specifiers are applied to the
real and imaginary parts before calling scm_make_rectangular. For
complex numbers written in polar form, exactness specifiers are applied
to the magnitude and angle before calling scm_make_polar.
There are two kinds of exactness specifiers: forced and implicit. A
forced exactness specifier is a "#e" or "#i" prefix at the beginning of
the entire number, and applies to both components of a complex number.
"#e" causes each component to be made exact, and "#i" causes each
component to be made inexact. If no forced exactness specifier is
present, then the exactness of each component is determined
independently by the presence or absence of a decimal point or hash mark
within that component. If a decimal point or hash mark is present, the
component is made inexact, otherwise it is made exact.
After the exactness specifiers have been applied to each component, they
are passed to either scm_make_rectangular or scm_make_polar to produce
the final result. Note that this will result in a real number if the
imaginary part, magnitude, or angle is an exact 0.
Previously, both forced and implicit exactness specifiers applied to
the number as a whole _after_ calling scm_make_rectangular or
scm_make_polar.
For example, (string->number "#i5.0+0i") now does the equivalent of:
(make-rectangular (exact->inexact 5.0) (exact->inexact 0))
which yields 5.0+0.0i. Previously it did the equivalent of:
(exact->inexact (make-rectangular 5.0 0))
which yielded 5.0.
* libguile/numbers.c (mem2ureal): Receive a forced exactness specifier
(forced_x), create and maintain our own implicit exactness specifier
flag local to this component (implicit_x), and apply these exactness
specifiers within this function. Previously, we received a pointer to
an implicit exactness specifier flag from above, and the exactness
specifiers were applied from within scm_i_string_length.
(mem2complex): Receive a forced exactness specifier parameter and pass
it down to mem2ureal. Previously, we passed down a pointer to an
implicit exactness specifier flag instead.
(scm_i_string_to_number): No longer create an implicit exactness
specifier flag here, and do not apply exactness specifiers here. All
we do here now regarding exactness is to parse the "#e" or "#i" prefix
(if any) and pass this information down to mem2ureal via mem2complex
in the form of an explicit exactness specifier (forced_x).
(scm_c_make_polar): If the cosine and sine of the angle are both NaNs
and the magnitude is zero, return 0.0+0.0i instead of +nan.0+nan.0i.
This case happens when the angle is not finite.
* test-suite/tests/numbers.test (string->number): Move the test cases
for non-real complex numbers into a separate table in which the
expected real and imaginary parts are separate entries. Add several
new test cases.
Add the ability to represent non-real complex numbers whose imaginary
part is an _inexact_ zero (0.0 or -0.0), per R6RS. Previously, such
numbers were immediately changed into inexact reals.
* libguile/numbers.c: Remove from the list of `General assumptions' in
numbers.c that objects satisfying SCM_COMPLEXP() have a non-zero
complex component. This is no longer true. Also add a warning
about another unrelated assumption that is not entirely correct
(that floor(r) == r implies that mpz_set_d will DTRT; it won't
if r is infinite).
(icmplx2str): Always print the imaginary part, even if it is zero.
Also handle a negative zero imaginary part more gracefully. It
now prints 0.0-0.0i, where previously it would print 0.0+-0.0i.
(mem2ureal): Replace scm_from_double (0.0) with flo0.
(scm_c_make_rectangular): Always create non-real complex numbers.
Previously it would create inexact reals if the specified imaginary
part was zero.
(scm_make_rectangular): If the imaginary part is an _exact_ 0, return
the real part unchanged (possibly exact), otherwise return a non-real
complex number (possibly with an inexact zero imaginary part).
Previously, it would return an inexact real number whenever the
imaginary part was any kind of zero.
(scm_make_polar): If the magnitude is an exact 0, return an exact 0.
If the angle is an exact 0, return the magnitude unchanged (possibly
exact). Otherwise return a non-real complex number (possibly with an
inexact zero imaginary part). Previously, it would return a real
number whenever the imaginary part was any kind of zero.
(scm_imag_part): Return an exact 0 if applied to a real number.
Previously it would return an inexact zero if applied to an inexact
real number.
(scm_inexact_to_exact): Accept complex numbers with inexact zero
imaginary part. In that case, simply use the real part and ignore the
imaginary part. Essentially we coerce the inexact zero imaginary part
to an exact 0.
* test-suite/tests/numbers.test: Add many test cases, and modify
existing tests as needed to reflect these changes. Also add a new
internal predicate: `almost-real-nan?' which tests for a non-real
complex number with zero imaginary part whose real part is a NaN.
* doc/ref/api-data.texi (Complex Numbers): Update description of complex
numbers to reflect these changes: non-real complex numbers in Guile
need not have non-zero imaginary part. Also, each part of a complex
number may be any inexact real, not just rationals as was previously
stated. Explicitly mention that each part may be an infinity, a NaN,
or a signed zero.
(Complex Number Operations): Change the formal parameter names of
`make-polar' from `x' and `y' to `mag' and `ang'.
* NEWS: Add news entries.
* libguile/numbers.c (scm_abs): (abs -0.0) now returns 0.0. Previously
it returned -0.0. Also move the REALP case above the BIGP case,
and consider it SCM_LIKELY to be REALP if not INUMP.
(scm_difference): (- 0 0.0) now returns -0.0. Previously it returned
0.0. Also make sure that (- 0 0.0+0.0i) will return -0.0-0.0i.
* test-suite/tests/numbers.test (abs, -): Add test cases, and change
some tests to use `eqv?' instead of `=', in order to test exactness
and distinguish signed zeroes.
