[LLVMdev] Soft floating point support

[Neil Booth]

This patch supplies software IEEE floating point support.

The comment from the patch reproduced below says all there is
to say.

This patch contains the prior "cleanup" patch; please don't apply
that one.

Please let me know of any bugs.  It is tested reasonably well,
but until I put together random tests it's hard to have 100%
confidence.

Neil.

/*  A self-contained host- and target-independent arbitrary-precision
    floating-point software implementation using bignum integer
    arithmetic, as provided by static functions in the APInt class.
    The library will work with bignum integers whose parts are any
    unsigned type at least 16 bits wide.  64 bits is recommended.

    Written for clarity rather than speed, in particular with a view
    to use in the front-end of a cross compiler so that target
    arithmetic can be correctly performed on the host.  Performance
    should nonetheless be reasonable, particularly for its intended
    use.  It may be useful as a base implementation for a run-time
    library during development of a faster target-specific one.

    All 5 rounding modes in the IEEE-754R draft are handled correctly
    for all implemented operations.  Currently implemented operations
    are add, subtract, multiply, divide, fused-multiply-add,
    conversion-to-float, conversion-to-integer and
    conversion-from-integer.  New rounding modes (e.g. away from zero)
    can be added with three or four lines of code.  The library reads
    and correctly rounds hexadecimal floating point numbers as per
    C99; syntax is required to have been validated by the caller.
    Conversion from decimal is not currently implemented.

    Four formats are built-in: IEEE single precision, double
    precision, quadruple precision, and x87 80-bit extended double
    (when operating with full extended precision).  Adding a new
    format that obeys IEEE semantics only requires adding two lines of
    code: a declaration and definition of the format.

    All operations return the status of that operation as an exception
    bit-mask, so multiple operations can be done consecutively with
    their results or-ed together.  The returned status can be useful
    for compiler diagnostics; e.g., inexact, underflow and overflow
    can be easily diagnosed on constant folding, and compiler
    optimizers can determine what exceptions would be raised by
    folding operations and optimize, or perhaps not optimize,
    accordingly.

    At present, underflow tininess is detected after rounding; it
    should be straight forward to add support for the before-rounding
    case too.

    Non-zero finite numbers are represented internally as a sign bit,
    a 16-bit signed exponent, and the significand as an array of
    integer parts.  After normalization of a number of precision P the
    exponent is within the range of the format, and if the number is
    not denormal the P-th bit of the significand is set as an explicit
    integer bit.  For denormals the most significant bit is shifted
    right so that the exponent is maintained at the format's minimum,
    so that the smallest denormal has just the least significant bit
    of the significand set.  The sign of zeroes and infinities is
    significant; the exponent and significand of such numbers is
    indeterminate and meaningless.  For QNaNs the sign bit, as well as
    the exponent and significand are indeterminate and meaningless.

    TODO
    ====

    Some features that may or may not be worth adding:

    Conversions to and from decimal strings (hard).

    Conversions to hexadecimal string.

    Read and write IEEE-format in-memory representations.

    Optional ability to detect underflow tininess before rounding.

    New formats: x87 in single and double precision mode (IEEE apart
    from extended exponent range) and IBM two-double extended
    precision (hard).

    New operations: sqrt, copysign, nextafter, nexttoward.
*/