
<HTML>

<HEAD>
<TITLE>Berkeley SoftFloat Library Interface</TITLE>
</HEAD>

<BODY>

<H1>Berkeley SoftFloat Release 3: Library Interface</H1>

<P>
John R. Hauser<BR>
2014 ______<BR>
</P>

<P>
*** CONTENT DONE.
</P>

<P>
*** REPLACE QUOTATION MARKS.
<BR>
*** REPLACE APOSTROPHES.
<BR>
*** REPLACE EM DASH.
</P>


<H2>Contents</H2>

<P>
*** CHECK.<BR>
*** FIX FORMATTING.
</P>

<PRE>
    Introduction
    Limitations
    Acknowledgments and License
    Types and Functions
        Boolean and Integer Types
        Floating-Point Types
        Supported Floating-Point Functions
        Non-canonical Representations in extFloat80_t
        Conventions for Passing Arguments and Results
    Reserved Names
    Mode Variables
        Rounding Mode
        Underflow Detection
        Rounding Precision for 80-Bit Extended Format
    Exceptions and Exception Flags
    Function Details
        Conversions from Integer to Floating-Point
        Conversions from Floating-Point to Integer
        Conversions Among Floating-Point Types
        Basic Arithmetic Functions
        Fused Multiply-Add Functions
        Remainder Functions
        Round-to-Integer Functions
        Comparison Functions
        Signaling NaN Test Functions
        Raise-Exception Function
    Changes from SoftFloat Release 2
        Name Changes
        Changes to Function Arguments
        Added Capabilities
        Better Compatibility with the C Language
        New Organization as a Library
        Optimization Gains (and Losses)
    Future Directions
    Contact Information
</PRE>


<H2>1. Introduction</H2>

<P>
Berkeley SoftFloat is a software implementation of binary floating-point that
conforms to the IEEE Standard for Floating-Point Arithmetic.
The current release supports four binary formats:  <NOBR>32-bit</NOBR>
single-precision, <NOBR>64-bit</NOBR> double-precision, <NOBR>80-bit</NOBR>
double-extended-precision, and <NOBR>128-bit</NOBR> quadruple-precision.
The following functions are supported for each format:
<UL>
<LI>
addition, subtraction, multiplication, division, and square root;
<LI>
fused multiply-add as defined by the IEEE Standard, except for
<NOBR>80-bit</NOBR> double-extended-precision;
<LI>
remainder as defined by the IEEE Standard;
<LI>
round to integral value;
<LI>
comparisons;
<LI>
conversions to/from other supported formats; and
<LI>
conversions to/from <NOBR>32-bit</NOBR> and <NOBR>64-bit</NOBR> integers,
signed and unsigned.
</UL>
All operations required by the original 1985 version of the IEEE Floating-Point
Standard are implemented, except for conversions to and from decimal.
</P>

<P>
This document gives information about the types defined and the routines
implemented by SoftFloat.
It does not attempt to define or explain the IEEE Floating-Point Standard.
Information about the standard is available elsewhere.
</P>

<P>
The current version of SoftFloat is <NOBR>Release 3</NOBR>.
The functional interface of this version differs in many details from that of
earlier SoftFloat releases.
For specifics of these differences, see <NOBR>section 9</NOBR> below,
<I>Changes from SoftFloat <NOBR>Release 2</NOBR></I>.
</P>


<H2>2. Limitations</H2>

<P>
SoftFloat assumes the computer has an addressable byte size of 8 or
<NOBR>16 bits</NOBR>.
(Nearly all computers in use today have <NOBR>8-bit</NOBR> bytes.)
</P>

<P>
SoftFloat is written in C and is designed to work with other C code.
The C compiler used must conform at a minimum to the 1989 ANSI standard for the
C language (same as the 1990 ISO standard) and must in addition support basic
arithmetic on <NOBR>64-bit</NOBR> integers.
Earlier releases of SoftFloat included implementations of <NOBR>32-bit</NOBR>
single-precision and <NOBR>64-bit</NOBR> double-precision floating-point that
did not require <NOBR>64-bit</NOBR> integers, but this option is not supported
with <NOBR>Release 3</NOBR>.
Since 1999, ISO standards for C have mandated compiler support for
<NOBR>64-bit</NOBR> integers.
A compiler conforming to the 1999 C Standard or later is recommended but not
strictly required.
</P>

<P>
Most operations not required by the original 1985 version of the IEEE
Floating-Point Standard but added in the 2008 version are not yet supported in
SoftFloat <NOBR>Release 3</NOBR>.
</P>


<H2>3. Acknowledgments and License</H2>

<P>
The SoftFloat package was written by me, <NOBR>John R.</NOBR> Hauser.
<NOBR>Release 3</NOBR> of SoftFloat is a completely new implementation
supplanting earlier releases.
This project was done in the employ of the University of California, Berkeley,
within the Department of Electrical Engineering and Computer Sciences, first
for the Parallel Computing Laboratory (Par Lab) and then for the ASPIRE Lab.
The work was officially overseen by Prof. Krste Asanovic, with funding provided
by these sources:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD><NOBR>Par Lab:</NOBR></TD>
<TD>
Microsoft (Award #024263), Intel (Award #024894), and U.C. Discovery
(Award #DIG07-10227), with additional support from Par Lab affiliates Nokia,
NVIDIA, Oracle, and Samsung.
</TD>
</TR>
<TR>
<TD><NOBR>ASPIRE Lab:</NOBR></TD>
<TD>
DARPA PERFECT program (Award #HR0011-12-2-0016), with additional support from
ASPIRE industrial sponsor Intel and ASPIRE affiliates Google, Nokia, NVIDIA,
Oracle, and Samsung.
</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
</P>

<P>
The following applies to the whole of SoftFloat <NOBR>Release 3</NOBR> as well
as to each source file individually.
</P>

<P>
Copyright 2011, 2012, 2013, 2014 The Regents of the University of California
(Regents).
All Rights Reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
</P>

<P>
Redistributions of source code must retain the above copyright notice, this
list of conditions, and the following two paragraphs of disclaimer.
Redistributions in binary form must reproduce the above copyright notice, this
list of conditions, and the following two paragraphs of disclaimer in the
documentation and/or other materials provided with the distribution.
Neither the name of the Regents nor the names of its contributors may be used
to endorse or promote products derived from this software without specific
prior written permission.
</P>

<P>
IN NO EVENT SHALL REGENTS BE LIABLE TO ANY PARTY FOR DIRECT, INDIRECT, SPECIAL,
INCIDENTAL, OR CONSEQUENTIAL DAMAGES, INCLUDING LOST PROFITS, ARISING OUT OF
THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN IF REGENTS HAS BEEN
ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
</P>

<P>
REGENTS SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
THE SOFTWARE AND ACCOMPANYING DOCUMENTATION, IF ANY, PROVIDED HEREUNDER IS
PROVIDED "<NOBR>AS IS</NOBR>".
REGENTS HAS NO OBLIGATION TO PROVIDE MAINTENANCE, SUPPORT, UPDATES,
ENHANCEMENTS, OR MODIFICATIONS.
</P>


<H2>4. Types and Functions</H2>

<P>
The types and functions of SoftFloat are declared in header file
<CODE>softfloat.h</CODE>.
</P>

<H3>4.1. Boolean and Integer Types</H3>

<P>
Header file <CODE>softfloat.h</CODE> depends on standard headers
<CODE>&lt;stdbool.h&gt;</CODE> and <CODE>&lt;stdint.h&gt;</CODE> to define type
<CODE>bool</CODE> and several integer types.
These standard headers have been part of the ISO C Standard Library since 1999.
With any recent compiler, they are likely to be supported, even if the compiler
does not claim complete conformance to the ISO C Standard.
For older or nonstandard compilers, a port of SoftFloat may have substitutes
for these headers.
Header <CODE>softfloat.h</CODE> depends only on the name <CODE>bool</CODE> from
<CODE>&lt;stdbool.h&gt;</CODE> and on these type names from
<CODE>&lt;stdint.h&gt;</CODE>:
<PRE>
     uint16_t
     uint32_t
     uint64_t
     int32_t
     int64_t
     uint_fast8_t
     uint_fast32_t
     uint_fast64_t
</PRE>
</P>


<H3>4.2. Floating-Point Types</H3>

<P>
The <CODE>softfloat.h</CODE> header defines four floating-point types:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD><CODE>float32_t</CODE></TD>
<TD>&nbsp;</TD>
<TD><NOBR>32-bit</NOBR> single-precision binary format</TD>
</TR>
<TR>
<TD><CODE>float64_t</CODE></TD>
<TD>&nbsp;</TD>
<TD><NOBR>64-bit</NOBR> double-precision binary format</TD>
</TR>
<TR>
<TD><CODE>extFloat80_t</CODE></TD>
<TD>&nbsp;</TD>
<TD><NOBR>80-bit</NOBR> double-extended-precision binary format (old Intel or
Motorola format)</TD>
</TR>
<TR>
<TD><CODE>float128_t</CODE></TD>
<TD>&nbsp;</TD>
<TD><NOBR>128-bit</NOBR> quadruple-precision binary format</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
The non-extended types are each exactly the size specified:
<NOBR>32 bits</NOBR> for <CODE>float32_t</CODE>, <NOBR>64 bits</NOBR> for
<CODE>float64_t</CODE>, and <NOBR>128 bits</NOBR> for <CODE>float128_t</CODE>.
Aside from these size requirements, the definitions of all these types may
differ for different ports of SoftFloat to specific systems.
A given port of SoftFloat may or may not define some of the floating-point
types as aliases for the C standard types <CODE>float</CODE>,
<CODE>double</CODE>, and <CODE>long</CODE> <CODE>double</CODE>.
</P>

<P>
Header file <CODE>softfloat.h</CODE> also defines a structure,
<CODE>struct</CODE> <CODE>extFloat80M</CODE>, for the representation of
<NOBR>80-bit</NOBR> double-extended-precision floating-point values in memory.
This structure is the same size as type <CODE>extFloat80_t</CODE> and contains
at least these two fields (not necessarily in this order):
<BLOCKQUOTE>
<TABLE>
<TR><TD><CODE>uint16_t signExp;</CODE></TD></TR>
<TR><TD><CODE>uint64_t signif;</CODE></TD></TR>
</TABLE>
</BLOCKQUOTE>
Field <CODE>signExp</CODE> contains the sign and exponent of the floating-point
value, with the sign in the most significant bit (<NOBR>bit 15</NOBR>) and the
encoded exponent in the other <NOBR>15 bits</NOBR>.
Field <CODE>signif</CODE> is the complete <NOBR>64-bit</NOBR> significand of
the floating-point value.
(In the usual encoding for <NOBR>80-bit</NOBR> extended floating-point, the
leading <NOBR>1 bit</NOBR> of normalized numbers is not implicit but is stored
in the most significant bit of the significand.)
</P>

<H3>4.3. Supported Floating-Point Functions</H3>

<P>
SoftFloat implements these arithmetic operations for its floating-point types:
<UL>
<LI>
conversions between any two floating-point formats;
<LI>
for each floating-point format, conversions to and from signed and unsigned
<NOBR>32-bit</NOBR> and <NOBR>64-bit</NOBR> integers;
<LI>
for each format, the usual addition, subtraction, multiplication, division, and
square root operations;
<LI>
for each format except <CODE>extFloat80_t</CODE>, the fused multiply-add
operation defined by the IEEE Standard;
<LI>
for each format, the floating-point remainder operation defined by the IEEE
Standard;
<LI>
for each format, a ``round to integer'' operation that rounds to the nearest
integer value in the same format; and
<LI>
comparisons between two values in the same floating-point format.
</UL>
</P>

<P>
The following operations required by the 2008 IEEE Floating-Point Standard are
not supported in SoftFloat <NOBR>Release 3</NOBR>:
<UL>
<LI>
<B>nextUp</B>, <B>nextDown</B>, <B>minNum</B>, <B>maxNum</B>, <B>minNumMag</B>,
<B>maxNumMag</B>, <B>scaleB</B>, and <B>logB</B>;
<LI>
conversions between floating-point formats and decimal or hexadecimal character
sequences;
<LI>
all ``quiet-computation'' operations (<B>copy</B>, <B>negate</B>, <B>abs</B>,
and <B>copySign</B>, which all involve only simple copying and/or manipulation
of the floating-point sign bit); and
<LI>
all ``non-computational'' operations other than <B>isSignaling</B> (which is
supported).
</UL>
</P>

<H3>4.4. Non-canonical Representations in <CODE>extFloat80_t</CODE></H3>

<P>
Because the <NOBR>80-bit</NOBR> double-extended-precision format,
<CODE>extFloat80_t</CODE>, stores an explicit leading significand bit, many
floating-point numbers are encodable in this type in equivalent normalized and
denormalized forms.
Zeros and values in the subnormal range have each only a single possible
encoding, for which the leading significand bit must <NOBR>be 0</NOBR>.
For other finite values (outside the subnormal range), a unique normalized
representation, with leading significand bit set <NOBR>to 1</NOBR>, always
exists, and is considered the <I>canonical</I> representation of the value.
Any equivalent denormalized representations (having leading significand bit
<NOBR>of 0</NOBR>) are <I>non-canonical</I>.
Similarly, the leading significand bit is expected to <NOBR>be 1</NOBR> for
infinities and NaNs as well;
any infinity or NaN with a leading significand bit <NOBR>of 0</NOBR> is again
considered non-canonical.
In short, for an <CODE>extFloat80_t</CODE> representation to be canonical, the
leading significand bit must <NOBR>be 1</NOBR> unless it is required to
<NOBR>be 0</NOBR> because the encoded value is zero or a subnormal.
</P>

<P>
For <NOBR>Release 3</NOBR> of SoftFloat, functions are not guaranteed to
operate as expected when inputs of type <CODE>extFloat80_t</CODE> are
non-canonical.
Assuming all of a function's <CODE>extFloat80_t</CODE> inputs (if any) are
canonical, function outputs of type <CODE>extFloat80_t</CODE> will always be
canonical.
</P>

<H3>4.5. Conventions for Passing Arguments and Results</H3>

<P>
Values that are at most <NOBR>64 bits</NOBR> in size (i.e., not the
<NOBR>80-bit</NOBR> or <NOBR>128-bit</NOBR> floating-point formats) are in all
cases passed as function arguments by value.
Likewise, when an output of a function is no more than <NOBR>64 bits</NOBR>, it
is always returned directly as the function result.
Thus, for example, the SoftFloat function for adding two <NOBR>64-bit</NOBR>
floating-point values has this simple signature:
<BLOCKQUOTE>
<CODE>float64_t f64_add( float64_t, float64_t );</CODE>
</BLOCKQUOTE>
</P>

<P>
The story is more complex when function inputs and outputs are
<NOBR>80-bit</NOBR> and <NOBR>128-bit</NOBR> floating-point.
For these types, SoftFloat always provides a function that passes these larger
values into or out of the function indirectly, via pointers.
For example, for adding two <NOBR>128-bit</NOBR> floating-point values,
SoftFloat supplies this function:
<BLOCKQUOTE>
<CODE>void f128M_add( const float128_t *, const float128_t *, float128_t * );</CODE>
</BLOCKQUOTE>
The first two arguments point to the values to be added, and the last argument
points to the location where the sum will be stored.
The <CODE>M</CODE> in the name <CODE>f128M_add</CODE> is mnemonic for the fact
that the <NOBR>128-bit</NOBR> inputs and outputs are ``in memory'', pointed to
by pointer arguments.
</P>

<P>
All ports of SoftFloat implement these <I>pass-by-pointer</I> functions for
types <CODE>extFloat80_t</CODE> and <CODE>float128_t</CODE>.
At the same time, SoftFloat ports may also implement alternate versions of
these same functions that pass <CODE>extFloat80_t</CODE> and
<CODE>float128_t</CODE> by value, like the smaller formats.
Thus, besides the function with name <CODE>f128M_add</CODE> shown above, a
SoftFloat port may also supply an equivalent function with this signature:
<BLOCKQUOTE>
<CODE>float128_t f128_add( float128_t, float128_t );</CODE>
</BLOCKQUOTE>
</P>

<P>
As a general rule, on computers where the machine word size is
<NOBR>32 bits</NOBR> or smaller, only the pass-by-pointer versions of functions
(e.g., <CODE>f128M_add</CODE>) are provided for types <CODE>extFloat80_t</CODE>
and <CODE>float128_t</CODE>, because passing such large types directly can have
significant extra cost.
On computers where the word size is <NOBR>64 bits</NOBR> or larger, both
function versions (<CODE>f128M_add</CODE> and <CODE>f128_add</CODE>) are
provided, because the cost of passing by value is then more reasonable.
Applications that must be portable accross both classes of computers must use
the pointer-based functions, as these are always implemented.
However, if it is known that SoftFloat includes the by-value functions for all
platforms of interest, programmers can use whichever version they prefer.
</P>


<H2>5. Reserved Names</H2>

<P>
In addition to the variables and functions documented here, SoftFloat defines
some symbol names for its own private use.
These private names always begin with the prefix `<CODE>softfloat_</CODE>'.
When a program includes header <CODE>softfloat.h</CODE> or links with the
SoftFloat library, all names with prefix `<CODE>softfloat_</CODE>' and not
documented for an application's use are reserved for possible use by SoftFloat.
Applications that use SoftFloat should not define their own names with this
prefix, and should reference only such names as are documented.
</P>


<H2>6. Mode Variables</H2>

<P>
The following variables control rounding mode, underflow detection, and the
<NOBR>80-bit</NOBR> extended format's rounding precision:
<BLOCKQUOTE>
<CODE>softfloat_roundingMode</CODE><BR>
<CODE>softfloat_detectTininess</CODE><BR>
<CODE>extF80_roundingPrecision</CODE>
</BLOCKQUOTE>
These mode variables are covered in the next several subsections.
</P>

<H3>6.1. Rounding Mode</H3>

<P>
All five rounding modes defined by the 2008 IEEE Floating-Point Standard are
implemented for all operations that require rounding.
The rounding mode is selected by the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t softfloat_roundingMode;</CODE>
</BLOCKQUOTE>
This variable may be set to one of the values
<BLOCKQUOTE>
<TABLE>
<TR>
<TD><CODE>softfloat_round_near_even</CODE></TD>
<TD>&nbsp;</TD>
<TD>round to nearest, with ties to even</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_near_maxMag</CODE></TD>
<TD>&nbsp;</TD>
<TD>round to nearest, with ties to maximum magnitude (away from zero)</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_minMag</CODE></TD>
<TD>&nbsp;</TD>
<TD>round to minimum magnitude (toward zero)</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_min</CODE></TD>
<TD>&nbsp;</TD>
<TD>round to minimum (down)</TD>
</TR>
<TR>
<TD><CODE>softfloat_round_max</CODE></TD>
<TD>&nbsp;</TD>
<TD>round to maximum (up)</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
Variable <CODE>softfloat_roundingMode</CODE> is initialized to
<CODE>softfloat_round_near_even</CODE>.
</P>

<H3>6.2. Underflow Detection</H3>

<P>
In the terminology of the IEEE Standard, SoftFloat can detect tininess for
underflow either before or after rounding.
The choice is made by the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t softfloat_detectTininess;</CODE>
</BLOCKQUOTE>
which can be set to either
<BLOCKQUOTE>
<CODE>softfloat_tininess_beforeRounding</CODE><BR>
<CODE>softfloat_tininess_afterRounding</CODE>
</BLOCKQUOTE>
Detecting tininess after rounding is better because it results in fewer
spurious underflow signals.
The other option is provided for compatibility with some systems.
Like most systems (and as required by the newer 2008 IEEE Standard), SoftFloat
always detects loss of accuracy for underflow as an inexact result.
</P>

<H3>6.3. Rounding Precision for 80-Bit Extended Format</H3>

<P>
For <CODE>extFloat80_t</CODE> only, the rounding precision of the basic
arithmetic operations is controlled by the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t extF80_roundingPrecision;</CODE>
</BLOCKQUOTE>
The operations affected are:
<BLOCKQUOTE>
<CODE>extF80_add</CODE><BR>
<CODE>extF80_sub</CODE><BR>
<CODE>extF80_mul</CODE><BR>
<CODE>extF80_div</CODE><BR>
<CODE>extF80_sqrt</CODE>
</BLOCKQUOTE>
When <CODE>extF80_roundingPrecision</CODE> is set to its default value of 80,
these operations are rounded to the full precision of the <NOBR>80-bit</NOBR>
double-extended-precision format, like occurs for other formats.
Setting <CODE>extF80_roundingPrecision</CODE> to 32 or to 64 causes the
operations listed to be rounded to <NOBR>32-bit</NOBR> precision (equivalent to
<CODE>float32_t</CODE>) or to <NOBR>64-bit</NOBR> precision (equivalent to
<CODE>float64_t</CODE>), respectively.
When rounding to reduced precision, additional bits in the result significand
beyond the rounding point are set to zero.
The consequences of setting <CODE>extF80_roundingPrecision</CODE> to a value
other than 32, 64, or 80 is not specified.
Operations other than the ones listed above are not affected by
<CODE>extF80_roundingPrecision</CODE>.
</P>


<H2>7. Exceptions and Exception Flags</H2>

<P>
All five exception flags required by the IEEE Floating-Point Standard are
implemented.
Each flag is stored as a separate bit in the global variable
<BLOCKQUOTE>
<CODE>uint_fast8_t softfloat_exceptionFlags;</CODE>
</BLOCKQUOTE>
The positions of the exception flag bits within this variable are determined by
the bit masks
<BLOCKQUOTE>
<CODE>softfloat_flag_inexact</CODE><BR>
<CODE>softfloat_flag_underflow</CODE><BR>
<CODE>softfloat_flag_overflow</CODE><BR>
<CODE>softfloat_flag_infinite</CODE><BR>
<CODE>softfloat_flag_invalid</CODE>
</BLOCKQUOTE>
Variable <CODE>softfloat_exceptionFlags</CODE> is initialized to all zeros,
meaning no exceptions.
</P>

<P>
An individual exception flag can be cleared with the statement
<BLOCKQUOTE>
<CODE>softfloat_exceptionFlags &= ~softfloat_flag_&lt;<I>exception</I>&gt;;</CODE>
</BLOCKQUOTE>
where <CODE>&lt;<I>exception</I>&gt;</CODE> is the appropriate name.
To raise a floating-point exception, function <CODE>softfloat_raise</CODE>
should normally be used.
</P>

<P>
When SoftFloat detects an exception other than <I>inexact</I>, it calls
<CODE>softfloat_raise</CODE>.
The default version of this function simply raises the corresponding exception
flags.
Particular ports of SoftFloat may support alternate behavior, such as exception
traps, by modifying the default <CODE>softfloat_raise</CODE>.
A program may also supply its own <CODE>softfloat_raise</CODE> function to
override the one from the SoftFloat library.
</P>

<P>
Because inexact results occur frequently under most circumstances (and thus are
hardly exceptional), SoftFloat does not ordinarily call
<CODE>softfloat_raise</CODE> for <I>inexact</I> exceptions.
It does always raise the <I>inexact</I> exception flag as required.
</P>


<H2>8. Function Details</H2>

<P>
In this section, <CODE>&lt;<I>float</I>&gt;</CODE> appears in function names as
a substitute for one of these abbreviations:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD><CODE>f32</CODE></TD>
<TD>indicates <CODE>float32_t</CODE>, passed by value</TD>
</TR>
<TR>
<TD><CODE>f64</CODE></TD>
<TD>indicates <CODE>float64_t</CODE>, passed by value</TD>
</TR>
<TR>
<TD><CODE>extF80M&nbsp;&nbsp;&nbsp;</CODE></TD>
<TD>indicates <CODE>extFloat80_t</CODE>, passed indirectly via pointers</TD>
</TR>
<TR>
<TD><CODE>extF80</CODE></TD>
<TD>indicates <CODE>extFloat80_t</CODE>, passed by value</TD>
</TR>
<TR>
<TD><CODE>f128M</CODE></TD>
<TD>indicates <CODE>float128_t</CODE>, passed indirectly via pointers</TD>
</TR>
<TR>
<TD><CODE>f128</CODE></TD>
<TD>indicates <CODE>float128_t</CODE>, passed by value</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
The circumstances under which values of floating-point types
<CODE>extFloat80_t</CODE> and <CODE>float128_t</CODE> may be passed either by
value or indirectly via pointers was discussed earlier in
<NOBR>section 4.5</NOBR>, <I>Conventions for Passing Arguments and Results</I>.
</P>

<H3>8.1. Conversions from Integer to Floating-Point</H3>

<P>
All conversions from a <NOBR>32-bit</NOBR> or <NOBR>64-bit</NOBR> integer,
signed or unsigned, to a floating-point format are supported.
Functions performing these conversions have these names:
<BLOCKQUOTE>
<CODE>ui32_to_&lt;<I>float</I>&gt;</CODE><BR>
<CODE>ui64_to_&lt;<I>float</I>&gt;</CODE><BR>
<CODE>i32_to_&lt;<I>float</I>&gt;</CODE><BR>
<CODE>i64_to_&lt;<I>float</I>&gt;</CODE>
</BLOCKQUOTE>
Conversions from <NOBR>32-bit</NOBR> integers to <NOBR>64-bit</NOBR>
double-precision and larger formats are always exact, and likewise conversions
from <NOBR>64-bit</NOBR> integers to <NOBR>80-bit</NOBR>
double-extended-precision and <NOBR>128-bit</NOBR> quadruple-precision are also
always exact.
</P>

<P>
Each conversion function takes one input of the appropriate type and generates
one output.
The following illustrates the signatures of these functions in cases when the
floating-point result is passed either by value or via pointers:
<PRE>
     float64_t i32_to_f64( int32_t <I>a</I> );

     void i32_to_f128M( int32_t <I>a</I>, float128_t *<I>destPtr</I> );
</PRE>
</P>

<H3>8.2. Conversions from Floating-Point to Integer</H3>

<P>
Conversions from a floating-point format to a <NOBR>32-bit</NOBR> or
<NOBR>64-bit</NOBR> integer, signed or unsigned, are supported with these
functions:
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_to_ui32</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_to_ui64</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_to_i32</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_to_i64</CODE>
</BLOCKQUOTE>
The functions have signatures as follows, depending on whether the
floating-point input is passed by value or via pointers:
<PRE>
     int32_t f64_to_i32( float64_t <I>a</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> );

     int32_t
      f128M_to_i32( const float128_t *<I>aPtr</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> );
</PRE>
The <CODE><I>roundingMode</I></CODE> argument specifies the rounding mode for
the conversion.
The variable that usually indicates rounding mode,
<CODE>softfloat_roundingMode</CODE>, is ignored.
Argument <CODE><I>exact</I></CODE> determines whether the <I>inexact</I>
exception flag is raised if the conversion is not exact.
If <CODE><I>exact</I></CODE> is <CODE>true</CODE>, the <I>inexact</I> flag may
be raised;
otherwise, it will not be, even if the conversion is inexact.
</P>

<P>
Conversions from floating-point to integer raise the <I>invalid</I> exception
if the source value cannot be rounded to a representable integer of the desired
size (32 or 64 bits).
In such a circumstance, if the floating-point input is a NaN or if the
conversion is to an unsigned integer type, the largest positive integer is
returned;
otherwise, the largest integer with the same sign as the input is returned.
The functions that convert to integer types never raise the <I>overflow</I>
exception.
</P>

<P>
Note that, when converting to an unsigned integer type, if the <I>invalid</I>
exception is raised because the input floating-point value would round to a
negative integer, the value returned is the <EM>maximum positive unsigned
integer</EM>.
Zero is not returned when the <I>invalid</I> exception is raised, even when
zero is the closest integer to the original floating-point value.
</P>

<P>
Because languages such <NOBR>as C</NOBR> require that conversions to integers
be rounded toward zero, the following functions are provided for improved speed
and convenience:
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_to_ui32_r_minMag</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_to_ui64_r_minMag</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_to_i32_r_minMag</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_to_i64_r_minMag</CODE>
</BLOCKQUOTE>
These functions round only toward zero (to minimum magnitude).
The signatures for these functions are the same as above without the redundant
<CODE><I>roundingMode</I></CODE> argument:
<PRE>
     int32_t f64_to_i32_r_minMag( float64_t <I>a</I>, bool <I>exact</I> );
</PRE>
<PRE>
     int32_t f128M_to_i32_r_minMag( const float128_t *<I>aPtr</I>, bool <I>exact</I> );
</PRE>
</P>

<H3>8.3. Conversions Among Floating-Point Types</H3>

<P>
Conversions between floating-point formats are done by functions with these
names:
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_to_&lt;<I>float</I>&gt;</CODE>
</BLOCKQUOTE>
All combinations of source and result type are supported where the source and
result are different formats.
There are four different styles of signature for these functions, depending on
whether the input and the output floating-point values are passed by value or
via pointers:
<PRE>
     float32_t f64_to_f32( float64_t <I>a</I> );
</PRE>
<PRE>
     float32_t f128M_to_f32( const float128_t *<I>aPtr</I> );
</PRE>
<PRE>
     void f32_to_f128M( float32_t <I>a</I>, float128_t *<I>destPtr</I> );
</PRE>
<PRE>
     void extF80M_to_f128M( const extFloat80_t *<I>aPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
</P>

<P>
Conversions from a smaller to a larger floating-point format are always exact
and so require no rounding.
</P>

<H3>8.4. Basic Arithmetic Functions</H3>

<P>
The following basic arithmetic functions are provided:
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_add</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_sub</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_mul</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_div</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_sqrt</CODE>
</BLOCKQUOTE>
Each floating-point operation takes two operands, except for <CODE>sqrt</CODE>
(square root) which takes only one.
The operands and result are all of the same floating-point format.
Signatures for these functions take the following forms:
<PRE>
     float64_t f64_add( float64_t <I>a</I>, float64_t <I>b</I> );
</PRE>
<PRE>
     void
      f128M_add(
          const float128_t *<I>aPtr</I>, const float128_t *<I>bPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
</P>
<P>
<PRE>
     float64_t f64_sqrt( float64_t <I>a</I> );
</PRE>
<PRE>
     void f128M_sqrt( const float128_t *<I>aPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
When floating-point values are passed indirectly through pointers, arguments
<CODE><I>aPtr</I></CODE> and <CODE><I>bPtr</I></CODE> point to the input
operands, and the last argument, <CODE><I>destPtr</I></CODE>, points to the
location where the result is stored.
</P>

<P>
Rounding of the <NOBR>80-bit</NOBR> double-extended-precision
(<CODE>extFloat80_t</CODE>) functions is affected by variable
<CODE>extF80_roundingPrecision</CODE>, as explained earlier in
<NOBR>section 6.3</NOBR>,
<I>Rounding Precision for <NOBR>80-Bit</NOBR> Extended Format</I>.
</P>

<H3>8.5. Fused Multiply-Add Functions</H3>

<P>
The 2008 version of the IEEE Floating-Point Standard defines a <I>fused
multiply-add</I> operation that does a combined multiplication and addition
with only a single rounding.
SoftFloat implements fused multiply-add with functions
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_mulAdd</CODE>
</BLOCKQUOTE>
Unlike other operations, fused multiple-add is supported only for the
non-extended formats, <CODE>float32_t</CODE>, <CODE>float64_t</CODE>, and
<CODE>float128_t</CODE>.
No fused multiple-add function is currently provided for the
<NOBR>80-bit</NOBR> double-extended-precision type, <CODE>extFloat80_t</CODE>.
</P>

<P>
Depending on whether floating-point values are passed by value or via pointers,
the fused multiply-add functions have signatures of these forms:
<PRE>
     float64_t f64_mulAdd( float64_t <I>a</I>, float64_t <I>b</I>, float64_t <I>c</I> );
</PRE>
<PRE>
     void
      f128M_mulAdd(
          const float128_t *<I>aPtr</I>,
          const float128_t *<I>bPtr</I>,
          const float128_t *<I>cPtr</I>,
          float128_t *<I>destPtr</I>
      );
</PRE>
The functions compute
<NOBR>(<CODE><I>a</I></CODE> &times; <CODE><I>b</I></CODE>)
 + <CODE><I>c</I></CODE></NOBR>
with a single rounding.
When floating-point values are passed indirectly through pointers, arguments
<CODE><I>aPtr</I></CODE>, <CODE><I>bPtr</I></CODE>, and
<CODE><I>cPtr</I></CODE> point to operands <CODE><I>a</I></CODE>,
<CODE><I>b</I></CODE>, and <CODE><I>c</I></CODE> respectively, and
<CODE><I>destPtr</I></CODE> points to the location where the result is stored.
</P>

<P>
If one of the multiplication operands <CODE><I>a</I></CODE> and
<CODE><I>b</I></CODE> is infinite and the other is zero, these functions raise
the invalid exception even if operand <CODE><I>c</I></CODE> is a NaN.
</P>

<H3>8.6. Remainder Functions</H3>

<P>
For each format, SoftFloat implements the remainder operation defined by the
IEEE Floating-Point Standard.
The remainder functions have names
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_rem</CODE>
</BLOCKQUOTE>
Each remainder operation takes two floating-point operands of the same format
and returns a result in the same format.
Depending on whether floating-point values are passed by value or via pointers,
the remainder functions have signatures of these forms:
<PRE>
     float64_t f64_rem( float64_t <I>a</I>, float64_t <I>b</I> );
</PRE>
<PRE>
     void
      f128M_rem(
          const float128_t *<I>aPtr</I>, const float128_t *<I>bPtr</I>, float128_t *<I>destPtr</I> );
</PRE>
When floating-point values are passed indirectly through pointers, arguments
<CODE><I>aPtr</I></CODE> and <CODE><I>bPtr</I></CODE> point to operands
<CODE><I>a</I></CODE> and <CODE><I>b</I></CODE> respectively, and
<CODE><I>destPtr</I></CODE> points to the location where the result is stored.
</P>

<P>
The IEEE Standard remainder operation computes the value
<NOBR><CODE><I>a</I></CODE>
 &minus; <I>n</I> &times; <CODE><I>b</I></CODE></NOBR>,
where <I>n</I> is the integer closest to
<NOBR><CODE><I>a</I></CODE> &divide; <CODE><I>b</I></CODE></NOBR>.
If <NOBR><CODE><I>a</I></CODE> &divide; <CODE><I>b</I></CODE></NOBR> is exactly
halfway between two integers, <I>n</I> is the <EM>even</EM> integer closest to
<NOBR><CODE><I>a</I></CODE> &divide; <CODE><I>b</I></CODE></NOBR>.
The IEEE Standard's remainder operation is always exact and so requires no
rounding.
</P>

<P>
Depending on the relative magnitudes of the operands, the remainder
functions can take considerably longer to execute than the other SoftFloat
functions.
This is inherent in the remainder operation itself and is not a flaw in the
SoftFloat implementation.
</P>

<H3>8.7. Round-to-Integer Functions</H3>

<P>
For each format, SoftFloat implements the round-to-integer operation specified
by the IEEE Floating-Point Standard.
These functions are named
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_roundToInt</CODE>
</BLOCKQUOTE>
Each round-to-integer operation takes a single floating-point operand.
This operand is rounded to an integer according to a specified rounding mode,
and the resulting integer value is returned in the same floating-point format.
(Note that the result is not an integer type.)
</P>

<P>
The signatures of the round-to-integer functions are similar to those for
conversions to an integer type:
<PRE>
     float64_t f64_roundToInt( float64_t <I>a</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> );
</PRE>
<PRE>
     void
      f128M_roundToInt(
          const float128_t *<I>aPtr</I>,
          uint_fast8_t <I>roundingMode</I>,
          bool <I>exact</I>,
          float128_t *<I>destPtr</I>
      );
</PRE>
The <CODE><I>roundingMode</I></CODE> argument specifies the rounding mode to
apply.
The variable that usually indicates rounding mode,
<CODE>softfloat_roundingMode</CODE>, is ignored.
Argument <CODE><I>exact</I></CODE> determines whether the <I>inexact</I>
exception flag is raised if the conversion is not exact.
If <CODE><I>exact</I></CODE> is <CODE>true</CODE>, the <I>inexact</I> flag may
be raised;
otherwise, it will not be, even if the conversion is inexact.
When floating-point values are passed indirectly through pointers,
<CODE><I>aPtr</I></CODE> points to the input operand and
<CODE><I>destPtr</I></CODE> points to the location where the result is stored.
</P>

<H3>8.8. Comparison Functions</H3>

<P>
For each format, the following floating-point comparison functions are
provided:
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_eq</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_le</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_lt</CODE>
</BLOCKQUOTE>
Each comparison takes two operands of the same type and returns a Boolean.
The abbreviation <CODE>eq</CODE> stands for ``equal'' (=);
<CODE>le</CODE> stands for ``less than or equal'' (&le;);
and <CODE>lt</CODE> stands for ``less than'' (&lt;).
Depending on whether the floating-point operands are passed by value or via
pointers, the comparison functions have signatures of these forms:
<PRE>
     bool f64_eq( float64_t <I>a</I>, float64_t <I>b</I> );
</PRE>
<PRE>
     bool f128M_eq( const float128_t *<I>aPtr</I>, const float128_t *<I>bPtr</I> );
</PRE>
</P>

<P>
The usual greater-than (&gt;), greater-than-or-equal (&ge;), and not-equal
(&ne;) comparisons are easily obtained from the functions provided.
The not-equal function is just the logical complement of the equal function.
The greater-than-or-equal function is identical to the less-than-or-equal
function with the arguments in reverse order, and likewise the greater-than
function is identical to the less-than function with the arguments reversed.
</P>

<P>
The IEEE Floating-Point Standard specifies that the less-than-or-equal and
less-than comparisons by default raise the <I>invalid</I> exception if either
operand is any kind of NaN.
Equality comparisons, on the other hand, are defined by default to raise the
<I>invalid</I> exception only for signaling NaNs, not quiet NaNs.
For completeness, SoftFloat provides these complementary functions:
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_eq_signaling</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_le_quiet</CODE><BR>
<CODE>&lt;<I>float</I>&gt;_lt_quiet</CODE>
</BLOCKQUOTE>
The <CODE>signaling</CODE> equality comparisons are identical to the default
equality comparisons except that the <I>invalid</I> exception is raised for any
NaN input, not just for signaling NaNs.
Similarly, the <CODE>quiet</CODE> comparison functions are identical to their
default counterparts except that the <I>invalid</I> exception is not raised for
quiet NaNs.
</P>

<H3>8.9. Signaling NaN Test Functions</H3>

<P>
Functions for testing whether a floating-point value is a signaling NaN are
provided with these names:
<BLOCKQUOTE>
<CODE>&lt;<I>float</I>&gt;_isSignalingNaN</CODE>
</BLOCKQUOTE>
The functions take one floating-point operand and return a Boolean indicating
whether the operand is a signaling NaN.
Accordingly, the functions have the forms
<PRE>
     bool f64_isSignalingNaN( float64_t <I>a</I> );
</PRE>
<PRE>
     bool f128M_isSignalingNaN( const float128_t *<I>aPtr</I> );
</PRE>
</P>

<H3>8.10. Raise-Exception Function</H3>

<P>
SoftFloat provides a single function for raising floating-point exceptions:
<PRE>
     void softfloat_raise( uint_fast8_t <I>exceptions</I> );
</PRE>
The <CODE><I>exceptions</I></CODE> argument is a mask indicating the set of
exceptions to raise.
(See earlier section 7, <I>Exceptions and Exception Flags</I>.)
In addition to setting the specified exception flags in variable
<CODE>softfloat_exceptionFlags</CODE>, the <CODE>softfloat_raise</CODE>
function may cause a trap or abort appropriate for the current system.
</P>


<H2>9. Changes from SoftFloat <NOBR>Release 2</NOBR></H2>

<H3>9.1. Name Changes</H3>

<P>
One of the biggest differences compared to <NOBR>Release 2</NOBR> is that the
names of most functions and variables have changed, even when the behavior has
not.
First, the floating-point types, the mode variables, the exception flags
variable, the function to raise exceptions, and various associated constants
have been renamed as follows:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD>old name, Release 2:</TD>
<TD>new name, Release 3:</TD>
</TR>
<TR>
<TD><CODE>float32</CODE></TD>
<TD><CODE>float32_t</CODE></TD>
</TR>
<TR>
<TD><CODE>float64</CODE></TD>
<TD><CODE>float64_t</CODE></TD>
</TR>
<TR>
<TD><CODE>floatx80</CODE></TD>
<TD><CODE>extFloat80_t</CODE></TD>
</TR>
<TR>
<TD><CODE>float128</CODE></TD>
<TD><CODE>float128_t</CODE></TD>
</TR>
<TR>
<TD><CODE>float_rounding_mode</CODE></TD>
<TD><CODE>softfloat_roundingMode</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_nearest_even</CODE></TD>
<TD><CODE>softfloat_round_near_even</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_to_zero</CODE></TD>
<TD><CODE>softfloat_round_minMag</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_down</CODE></TD>
<TD><CODE>softfloat_round_min</CODE></TD>
</TR>
<TR>
<TD><CODE>float_round_up</CODE></TD>
<TD><CODE>softfloat_round_max</CODE></TD>
</TR>
<TR>
<TD><CODE>float_detect_tininess</CODE></TD>
<TD><CODE>softfloat_detectTininess</CODE></TD>
</TR>
<TR>
<TD><CODE>float_tininess_before_rounding&nbsp;&nbsp;&nbsp;&nbsp;</CODE></TD>
<TD><CODE>softfloat_tininess_beforeRounding</CODE></TD>
</TR>
<TR>
<TD><CODE>float_tininess_after_rounding</CODE></TD>
<TD><CODE>softfloat_tininess_afterRounding</CODE></TD>
</TR>
<TR>
<TD><CODE>floatx80_rounding_precision</CODE></TD>
<TD><CODE>extF80_roundingPrecision</CODE></TD>
</TR>
<TR>
<TD><CODE>float_exception_flags</CODE></TD>
<TD><CODE>softfloat_exceptionFlags</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_inexact</CODE></TD>
<TD><CODE>softfloat_flag_inexact</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_underflow</CODE></TD>
<TD><CODE>softfloat_flag_underflow</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_overflow</CODE></TD>
<TD><CODE>softfloat_flag_overflow</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_divbyzero</CODE></TD>
<TD><CODE>softfloat_flag_infinite</CODE></TD>
</TR>
<TR>
<TD><CODE>float_flag_invalid</CODE></TD>
<TD><CODE>softfloat_flag_invalid</CODE></TD>
</TR>
<TR>
<TD><CODE>float_raise</CODE></TD>
<TD><CODE>softfloat_raise</CODE></TD>
</TR>
</TABLE>
</BLOCKQUOTE>
</P>

<P>
Furthermore, <NOBR>Release 3</NOBR> has adopted the following new abbreviations
for function names:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD>used in names in Release 2:<CODE>&nbsp;&nbsp;&nbsp;&nbsp;</CODE></TD>
<TD>used in names in Release 3:</TD>
</TR>
<TR> <TD><CODE>int32</CODE></TD>    <TD><CODE>i32</CODE></TD>    </TR>
<TR> <TD><CODE>int64</CODE></TD>    <TD><CODE>i64</CODE></TD>    </TR>
<TR> <TD><CODE>float32</CODE></TD>  <TD><CODE>f32</CODE></TD>    </TR>
<TR> <TD><CODE>float64</CODE></TD>  <TD><CODE>f64</CODE></TD>    </TR>
<TR> <TD><CODE>floatx80</CODE></TD> <TD><CODE>extF80</CODE></TD> </TR>
<TR> <TD><CODE>float128</CODE></TD> <TD><CODE>f128</CODE></TD>   </TR>
</TABLE>
</BLOCKQUOTE>
Thus, for example, the function to add two <NOBR>32-bit</NOBR> floating-point
numbers, previously called <CODE>float32_add</CODE> in <NOBR>Release 2</NOBR>,
is now <CODE>f32_add</CODE>.
Lastly, there are a few other changes to function names:
<BLOCKQUOTE>
<TABLE>
<TR>
<TD>used in names in Release 2:<CODE>&nbsp;&nbsp;&nbsp;</CODE></TD>
<TD>used in names in Release 3:<CODE>&nbsp;&nbsp;&nbsp;</CODE></TD>
<TD>relevant functions:</TD>
</TR>
<TR>
<TD><CODE>_round_to_zero</CODE></TD>
<TD><CODE>_r_minMag</CODE></TD>
<TD>conversions from floating-point to integer, section 8.2</TD>
</TR>
<TR>
<TD><CODE>round_to_int</CODE></TD>
<TD><CODE>roundToInt</CODE></TD>
<TD>round-to-integer functions, section 8.7</TD>
</TR>
<TR>
<TD><CODE>is_signaling_nan&nbsp;&nbsp;&nbsp;&nbsp;</CODE></TD>
<TD><CODE>isSignalingNaN</CODE></TD>
<TD>signaling NaN test functions, section 8.9</TD>
</TR>
</TABLE>
</BLOCKQUOTE>
</P>

<H3>9.2. Changes to Function Arguments</H3>

<P>
Besides simple name changes, some operations have a different interface in
<NOBR>Release 3</NOBR> than they had in <NOBR>Release 2</NOBR>:
<UL>

<LI>
<P>
In <NOBR>Release 3</NOBR>, integer arguments and results of functions have
standard types from header <CODE>&lt;stdint.h&gt;</CODE>, such as
<CODE>uint32_t</CODE>, whereas previously their types could be defined
differently for each port of SoftFloat, usually using traditional C types such
as <CODE>unsigned</CODE> <CODE>int</CODE>.
Likewise, functions in <NOBR>Release 3</NOBR> pass Booleans as standard type
<CODE>bool</CODE> from <CODE>&lt;stdbool.h&gt;</CODE>, whereas previously these
were again passed as a port-specific type (usually <CODE>int</CODE>).
</P>

<LI>
<P>
As explained earlier in <NOBR>section 4.5</NOBR>, <I>Conventions for Passing
Arguments and Results</I>, SoftFloat functions in <NOBR>Release 3</NOBR> may
pass <NOBR>80-bit</NOBR> and <NOBR>128-bit</NOBR> floating-point values through
pointers, meaning that functions take pointer arguments and then read or write
floating-point values at the locations indicated by the pointers.
In <NOBR>Release 2</NOBR>, floating-point arguments and results were always
passed by value, regardless of their size.
</P>

<LI>
<P>
Functions that round to an integer have additional
<CODE><I>roundingMode</I></CODE> and <CODE><I>exact</I></CODE> arguments that
they did not have in <NOBR>Release 2</NOBR>.
Refer to sections 8.2 <NOBR>and 8.7</NOBR> for descriptions of these functions
in <NOBR>Release 3</NOBR>.
For <NOBR>Release 2</NOBR>, the rounding mode, when needed, was taken from the
same global variable that affects the basic arithmetic operations (now called
<CODE>softfloat_roundingMode</CODE> but previously known as
<CODE>float_rounding_mode</CODE>).
Also, for <NOBR>Release 2</NOBR>, if the original floating-point input was not
an exact integer value, and if the <I>invalid</I> exception was not raised by
the function, the <I>inexact</I> exception was always raised.
<NOBR>Release 2</NOBR> had no option to suppress raising <I>inexact</I> in this
case.
Applications using SoftFloat <NOBR>Release 3</NOBR> can get the same effect as
<NOBR>Release 2</NOBR> by passing variable <CODE>softfloat_roundingMode</CODE>
for argument <CODE><I>roundingMode</I></CODE> and <CODE>true</CODE> for
argument <CODE><I>exact</I></CODE>.
</P>

</UL>
</P>

<H3>9.3. Added Capabilities</H3>

<P>
<NOBR>Release 3</NOBR> adds some new features not present in
<NOBR>Release 2</NOBR>:
<UL>

<LI>
<P>
With <NOBR>Release 3</NOBR>, a port of SoftFloat can now define any of the
floating-point types <CODE>float32_t</CODE>, <CODE>float64_t</CODE>,
<CODE>extFloat80_t</CODE>, and <CODE>float128_t</CODE> as aliases for C's
standard floating-point types <CODE>float</CODE>, <CODE>double</CODE>, and
<CODE>long</CODE> <CODE>double</CODE>, using either <CODE>#define</CODE> or
<CODE>typedef</CODE>.
This potential convenience was not supported under <NOBR>Release 2</NOBR>.
</P>

<P>
(Note, however, that there may be a performance cost to defining SoftFloat's
floating-point types this way, depending on the platform and the applications
using SoftFloat.
Ports of SoftFloat may choose to forgo the convenience in favor of better
speed.)
</P>

<P>
<LI>
Functions have been added for converting between the floating-point types and
unsigned integers.
<NOBR>Release 2</NOBR> supported only signed integers, not unsigned.
</P>

<P>
<LI>
A new, fifth rounding mode, <CODE>softfloat_round_near_maxMag</CODE> (round to
nearest, with ties to maximum magnitude, away from zero) is supported for all
cases involving rounding.
</P>

<P>
<LI>
Fused multiply-add functions have been added for the non-extended formats,
<CODE>float32_t</CODE>, <CODE>float64_t</CODE>, and <CODE>float128_t</CODE>.
</P>

</UL>
</P>

<H3>9.4. Better Compatibility with the C Language</H3>

<P>
<NOBR>Release 3</NOBR> of SoftFloat is written to conform better to the ISO C
Standard's rules for portability.
For example, older releases of SoftFloat employed type conversions in ways
that, while commonly practiced, are not fully defined by the C Standard.
Such problematic type conversions have generally been replaced by the use of
unions, the behavior around which is more strictly regulated these days.
</P>

<H3>9.5. New Organization as a Library</H3>

<P>
With <NOBR>Release 3</NOBR>, SoftFloat now builds as a library.
Previously, SoftFloat compiled into a single, monolithic object file containing
all the SoftFloat functions, with the consequence that a program linking with
SoftFloat would get every SoftFloat function in its binary file even if only a
few functions were actually used.
With SoftFloat in the form of a library, a program that is linked by a standard
linker will include only those functions of SoftFloat that it needs and no
others.
</P>

<H3>9.6. Optimization Gains (and Losses)</H3>

<P>
Individual SoftFloat functions are variously improved in <NOBR>Release 3</NOBR>
compared to earlier releases.
In particular, better, faster algorithms have been deployed for the operations
of division, square root, and remainder.
For functions operating on the larger <NOBR>80-bit</NOBR> and
<NOBR>128-bit</NOBR> formats, <CODE>extFloat80_t</CODE> and
<CODE>float128_t</CODE>, code size has also generally been reduced.
</P>

<P>
However, because <NOBR>Release 2</NOBR> compiled all of SoftFloat together as a
single object file, compilers could make optimizations across function calls
when one SoftFloat function calls another.
Now that the functions of SoftFloat are compiled separately and only afterward
linked together into a program, there is not the same opportunity to optimize
across function calls.
Some loss of speed has been observed due to this change.
</P>


<H2>10. Future Directions</H2>

<P>
The following improvements are anticipated for future releases of SoftFloat:
<UL>
<LI>
support for the common <NOBR>16-bit</NOBR> ``half-precision'' floating-point
format;
<LI>
more functions from the 2008 version of the IEEE Floating-Point Standard;
<LI>
consistent, defined behavior for non-canonical representations of extended
format <CODE>extFloat80_t</CODE> (discussed in <NOBR>section 4.4</NOBR>,
<I>Non-canonical Representations in <CODE>extFloat80_t</CODE></I>).

</UL>
</P>


<H2>11. Contact Information</H2>

<P>
At the time of this writing, the most up-to-date information about SoftFloat
and the latest release can be found at the Web page
<A HREF="http://www.jhauser.us/arithmetic/SoftFloat.html"><CODE>http://www.jhauser.us/arithmetic/SoftFloat.html</CODE></A>.
</P>


</BODY>