Decimal Numbers
Paul D Anderson via Digitalmars-d-announce
digitalmars-d-announce at puremagic.com
Thu Jul 3 14:55:40 PDT 2014
A candidate implementation of decimal numbers (arbitrary-precision
floating-point numbers) is available for review at
https://github.com/andersonpd/eris/tree/master/eris/decimal. This
is a
substantial rework of an earlier implementation which was located
at
https://github.com/andersonpd/decimal.
This is a D language implementation of the General Decimal
Arithmetic
Specification (http://www.speleotrove.com/decimal/decarith.pdf),
which is
compliant with IEEE-754 and other standards as noted in the
specification.
The current implementation is not complete; there are a lot of
TODOs and NOTEs
scattered throughout the code, but all the arithmetic and
miscellaneous
operations listed in the spec are working, along with decimal
versions of most
of the functions and constants in std.math. I think it is far
enough along for
effective review.
Briefly, this software adds the capability of properly rounded
arbitrary-precision floating-point arithmetic to the D language.
All arithmetic
operations are governed by a "context", which specifies the
precision (number of
decimal digits) and rounding mode for the operations. This same
functionality
exists in most modern computer languages (for example,
java.math.BigDecimal).
Unlike Java, however, which uses function syntax for arithmetic
ops
(add(BigDecimal, BigDecimal), etc.), in D the same arithmetic
operators that
work for floats or doubles work for decimal numbers. (Of course!)
In this implementation decimal numbers having different contexts
are different
types. The types are specified using template parameters for the
precision,
maximum exponent value and rounding mode. This means that
Decimal!(9,99,Rounding.HALF_EVEN) is a different type than
Decimal!(19,199,Rounding.HALF_DOWN). They are largely
interoperable, however.
Different decimal types can be cast to and from each
other.
There are three standard decimal structs which fit into 32-, 64-
and 128-bits of
memory, with 7, 16 and 34 digit precision, respectively. These
are used for
compact storage; they are converted to their corresponding
decimal numbers for
calculation. They bear the same relation to decimal numbers as
Walter's
half-float type does to floats.
(http://www.drdobbs.com/cpp/implementing-half-floats-in-d/240146674).
Implementation of these still needs a little work, and will be
added to github
very shortly.
Major TODO items:
1) The current underlying integer type uses my own big integer
struct
(eris.integer.extended) rather than std.bigint. This was mainly
due to problems
with constness and CTFE of BigInts. These problems have since
been resolved, but
I didn't want to switch over to BigInts until everything was
working for fear of
introducing new bugs.
2) Integration of Decimal32, Decimal64 and Decimal128 structs are
not complete.
(See above.)
3) Conversion to and from floats, doubles and reals is currently
working but it
is slow. (Conversion is through strings: double to string to
decimal and vice
versa.)
4) Still incomplete implementations of some functions in
decimal.math: expm1,
acosh, atanh, possibly others.
5) More unit tests (always!).
More information about the Digitalmars-d-announce
mailing list