| |
| <HTML> |
| |
| <HEAD> |
| <TITLE>Berkeley SoftFloat Library Interface</TITLE> |
| </HEAD> |
| |
| <BODY> |
| |
| <H1>Berkeley SoftFloat Release 3e: Library Interface</H1> |
| |
| <P> |
| John R. Hauser<BR> |
| 2018 January 20<BR> |
| </P> |
| |
| |
| <H2>Contents</H2> |
| |
| <BLOCKQUOTE> |
| <TABLE BORDER=0 CELLSPACING=0 CELLPADDING=0> |
| <COL WIDTH=25> |
| <COL WIDTH=*> |
| <TR><TD COLSPAN=2>1. Introduction</TD></TR> |
| <TR><TD COLSPAN=2>2. Limitations</TD></TR> |
| <TR><TD COLSPAN=2>3. Acknowledgments and License</TD></TR> |
| <TR><TD COLSPAN=2>4. Types and Functions</TD></TR> |
| <TR><TD></TD><TD>4.1. Boolean and Integer Types</TD></TR> |
| <TR><TD></TD><TD>4.2. Floating-Point Types</TD></TR> |
| <TR><TD></TD><TD>4.3. Supported Floating-Point Functions</TD></TR> |
| <TR> |
| <TD></TD> |
| <TD>4.4. Non-canonical Representations in <CODE>extFloat80_t</CODE></TD> |
| </TR> |
| <TR><TD></TD><TD>4.5. Conventions for Passing Arguments and Results</TD></TR> |
| <TR><TD COLSPAN=2>5. Reserved Names</TD></TR> |
| <TR><TD COLSPAN=2>6. Mode Variables</TD></TR> |
| <TR><TD></TD><TD>6.1. Rounding Mode</TD></TR> |
| <TR><TD></TD><TD>6.2. Underflow Detection</TD></TR> |
| <TR> |
| <TD></TD> |
| <TD>6.3. Rounding Precision for the <NOBR>80-Bit</NOBR> Extended Format</TD> |
| </TR> |
| <TR><TD COLSPAN=2>7. Exceptions and Exception Flags</TD></TR> |
| <TR><TD COLSPAN=2>8. Function Details</TD></TR> |
| <TR><TD></TD><TD>8.1. Conversions from Integer to Floating-Point</TD></TR> |
| <TR><TD></TD><TD>8.2. Conversions from Floating-Point to Integer</TD></TR> |
| <TR><TD></TD><TD>8.3. Conversions Among Floating-Point Types</TD></TR> |
| <TR><TD></TD><TD>8.4. Basic Arithmetic Functions</TD></TR> |
| <TR><TD></TD><TD>8.5. Fused Multiply-Add Functions</TD></TR> |
| <TR><TD></TD><TD>8.6. Remainder Functions</TD></TR> |
| <TR><TD></TD><TD>8.7. Round-to-Integer Functions</TD></TR> |
| <TR><TD></TD><TD>8.8. Comparison Functions</TD></TR> |
| <TR><TD></TD><TD>8.9. Signaling NaN Test Functions</TD></TR> |
| <TR><TD></TD><TD>8.10. Raise-Exception Function</TD></TR> |
| <TR><TD COLSPAN=2>9. Changes from SoftFloat <NOBR>Release 2</NOBR></TD></TR> |
| <TR><TD></TD><TD>9.1. Name Changes</TD></TR> |
| <TR><TD></TD><TD>9.2. Changes to Function Arguments</TD></TR> |
| <TR><TD></TD><TD>9.3. Added Capabilities</TD></TR> |
| <TR><TD></TD><TD>9.4. Better Compatibility with the C Language</TD></TR> |
| <TR><TD></TD><TD>9.5. New Organization as a Library</TD></TR> |
| <TR><TD></TD><TD>9.6. Optimization Gains (and Losses)</TD></TR> |
| <TR><TD COLSPAN=2>10. Future Directions</TD></TR> |
| <TR><TD COLSPAN=2>11. Contact Information</TD></TR> |
| </TABLE> |
| </BLOCKQUOTE> |
| |
| |
| <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 five binary formats: <NOBR>16-bit</NOBR> |
| half-precision, <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 3e</NOBR>. |
| This release modifies the behavior of the rarely used <I>odd</I> rounding mode |
| (<I>round to odd</I>, also known as <I>jamming</I>), and also adds some new |
| specialization and optimization examples for those compiling SoftFloat. |
| </P> |
| |
| <P> |
| The previous <NOBR>Release 3d</NOBR> fixed bugs that were found in the square |
| root functions for the <NOBR>64-bit</NOBR>, <NOBR>80-bit</NOBR>, and |
| <NOBR>128-bit</NOBR> floating-point formats. |
| (Thanks to Alexei Sibidanov at the University of Victoria for reporting an |
| incorrect result.) |
| The bugs affected all prior <NOBR>Release-3</NOBR> versions of SoftFloat |
| <NOBR>through 3c</NOBR>. |
| The flaw in the <NOBR>64-bit</NOBR> floating-point square root function was of |
| very minor impact, causing a <NOBR>1-ulp</NOBR> error (<NOBR>1 unit</NOBR> in |
| the last place) a few times out of a billion. |
| The bugs in the <NOBR>80-bit</NOBR> and <NOBR>128-bit</NOBR> square root |
| functions were more serious. |
| Although incorrect results again occurred only a few times out of a billion, |
| when they did occur a large portion of the less-significant bits could be |
| wrong. |
| </P> |
| |
| <P> |
| Among earlier releases, 3b was notable for adding support for the |
| <NOBR>16-bit</NOBR> half-precision format. |
| For more about the evolution of SoftFloat releases, see |
| <A HREF="SoftFloat-history.html"><NOBR><CODE>SoftFloat-history.html</CODE></NOBR></A>. |
| </P> |
| |
| <P> |
| The functional interface of SoftFloat <NOBR>Release 3</NOBR> and later differs |
| in many details from the releases that came before. |
| 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 |
| starting 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 3e</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 was a completely new implementation |
| supplanting earlier releases. |
| The project to create <NOBR>Release 3</NOBR> (now <NOBR>through 3e</NOBR>) 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> |
| <COL> |
| <COL WIDTH=10> |
| <COL> |
| <TR> |
| <TD VALIGN=TOP><NOBR>Par Lab:</NOBR></TD> |
| <TD></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 VALIGN=TOP><NOBR>ASPIRE Lab:</NOBR></TD> |
| <TD></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 3e</NOBR> as well |
| as to each source file individually. |
| </P> |
| |
| <P> |
| Copyright 2011, 2012, 2013, 2014, 2015, 2016, 2017, 2018 The Regents of the |
| University of California. |
| All rights reserved. |
| </P> |
| |
| <P> |
| Redistribution and use in source and binary forms, with or without |
| modification, are permitted provided that the following conditions are met: |
| <OL> |
| |
| <LI> |
| <P> |
| Redistributions of source code must retain the above copyright notice, this |
| list of conditions, and the following disclaimer. |
| </P> |
| |
| <LI> |
| <P> |
| Redistributions in binary form must reproduce the above copyright notice, this |
| list of conditions, and the following disclaimer in the documentation and/or |
| other materials provided with the distribution. |
| </P> |
| |
| <LI> |
| <P> |
| Neither the name of the University nor the names of its contributors may be |
| used to endorse or promote products derived from this software without specific |
| prior written permission. |
| </P> |
| |
| </OL> |
| </P> |
| |
| <P> |
| THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS “AS IS”, |
| AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, ARE |
| DISCLAIMED. |
| IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, |
| INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, |
| BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
| DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
| LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE |
| OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF |
| ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| </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><stdbool.h></CODE> and <CODE><stdint.h></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 latest 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><stdbool.h></CODE> and on these type names from |
| <CODE><stdint.h></CODE>: |
| <BLOCKQUOTE> |
| <PRE> |
| uint16_t |
| uint32_t |
| uint64_t |
| int32_t |
| int64_t |
| uint_fast8_t |
| uint_fast32_t |
| uint_fast64_t |
| int_fast32_t |
| int_fast64_t |
| </PRE> |
| </BLOCKQUOTE> |
| </P> |
| |
| |
| <H3>4.2. Floating-Point Types</H3> |
| |
| <P> |
| The <CODE>softfloat.h</CODE> header defines five floating-point types: |
| <BLOCKQUOTE> |
| <TABLE CELLSPACING=0 CELLPADDING=0> |
| <TR> |
| <TD><CODE>float16_t</CODE></TD> |
| <TD><NOBR>16-bit</NOBR> half-precision binary format</TD> |
| </TR> |
| <TR> |
| <TD><CODE>float32_t</CODE></TD> |
| <TD><NOBR>32-bit</NOBR> single-precision binary format</TD> |
| </TR> |
| <TR> |
| <TD><CODE>float64_t</CODE></TD> |
| <TD><NOBR>64-bit</NOBR> double-precision binary format</TD> |
| </TR> |
| <TR> |
| <TD><CODE>extFloat80_t </CODE></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><NOBR>128-bit</NOBR> quadruple-precision binary format</TD> |
| </TR> |
| </TABLE> |
| </BLOCKQUOTE> |
| The non-extended types are each exactly the size specified: |
| <NOBR>16 bits</NOBR> for <CODE>float16_t</CODE>, <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> |
| <PRE> |
| uint16_t signExp; |
| uint64_t signif; |
| </PRE> |
| </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 3e</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 |
| finite floating-point numbers are encodable in this type in multiple equivalent |
| forms. |
| Of these multiple encodings, there is always a unique one with the least |
| encoded exponent value, and this encoding is considered the <I>canonical</I> |
| representation of the floating-point number. |
| Any other equivalent representations (having a higher encoded exponent value) |
| are <I>non-canonical</I>. |
| For a value in the subnormal range (including zero), the canonical |
| representation always has an encoded exponent of zero and a leading significand |
| bit <NOBR>of 0</NOBR>. |
| For finite values outside the subnormal range, the canonical representation |
| always has an encoded exponent that is nonzero and a leading significand bit |
| <NOBR>of 1</NOBR>. |
| </P> |
| |
| <P> |
| For an infinity or NaN, the leading significand bit is similarly expected to |
| <NOBR>be 1</NOBR>. |
| An infinity or NaN with a leading significand bit <NOBR>of 0</NOBR> is again |
| considered non-canonical. |
| Hence, altogether, to be canonical, a value of type <CODE>extFloat80_t</CODE> |
| must have a leading significand bit <NOBR>of 1</NOBR>, unless the value is |
| subnormal or zero, in which case the leading significand bit and the encoded |
| exponent must both be zero. |
| </P> |
| |
| <P> |
| SoftFloat’s 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>’ |
| 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 global 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. |
| For some SoftFloat ports, these variables may be <I>per-thread</I> (declared |
| <CODE>thread_local</CODE>), meaning that different execution threads have their |
| own separate copies of the variables. |
| </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. |
| Some ports of SoftFloat may also implement the <I>round-to-odd</I> mode. |
| </P> |
| |
| <P> |
| 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 CELLSPACING=0 CELLPADDING=0> |
| <TR> |
| <TD><CODE>softfloat_round_near_even</CODE></TD> |
| <TD>round to nearest, with ties to even</TD> |
| </TR> |
| <TR> |
| <TD><CODE>softfloat_round_near_maxMag </CODE></TD> |
| <TD>round to nearest, with ties to maximum magnitude (away from zero)</TD> |
| </TR> |
| <TR> |
| <TD><CODE>softfloat_round_minMag</CODE></TD> |
| <TD>round to minimum magnitude (toward zero)</TD> |
| </TR> |
| <TR> |
| <TD><CODE>softfloat_round_min</CODE></TD> |
| <TD>round to minimum (down)</TD> |
| </TR> |
| <TR> |
| <TD><CODE>softfloat_round_max</CODE></TD> |
| <TD>round to maximum (up)</TD> |
| </TR> |
| <TR> |
| <TD><CODE>softfloat_round_odd</CODE></TD> |
| <TD>round to odd (jamming), if supported by the SoftFloat port</TD> |
| </TR> |
| </TABLE> |
| </BLOCKQUOTE> |
| Variable <CODE>softfloat_roundingMode</CODE> is initialized to |
| <CODE>softfloat_round_near_even</CODE>. |
| </P> |
| |
| <P> |
| When <CODE>softfloat_round_odd</CODE> is the rounding mode for a function that |
| rounds to an integer value (either conversion to an integer format or a |
| ‘<CODE>roundToInt</CODE>’ function), if the input is not already an |
| integer, the rounded result is the closest <EM>odd</EM> integer. |
| For other operations, this rounding mode acts as though the floating-point |
| result is first rounded to minimum magnitude, the same as |
| <CODE>softfloat_round_minMag</CODE>, and then, if the result is inexact, the |
| least-significant bit of the result is set <NOBR>to 1</NOBR>. |
| Rounding to odd is also known as <EM>jamming</EM>. |
| </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 usually 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 the <NOBR>80-Bit</NOBR> 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> |
| For some SoftFloat ports, <CODE>softfloat_exceptionFlags</CODE> may be |
| <I>per-thread</I> (declared <CODE>thread_local</CODE>), meaning that different |
| execution threads have their own separate instances of it. |
| </P> |
| |
| <P> |
| An individual exception flag can be cleared with the statement |
| <BLOCKQUOTE> |
| <CODE>softfloat_exceptionFlags &= ~softfloat_flag_<<I>exception</I>>;</CODE> |
| </BLOCKQUOTE> |
| where <CODE><<I>exception</I>></CODE> is the appropriate name. |
| To raise a floating-point exception, function <CODE>softfloat_raiseFlags</CODE> |
| should normally be used. |
| </P> |
| |
| <P> |
| When SoftFloat detects an exception other than <I>inexact</I>, it calls |
| <CODE>softfloat_raiseFlags</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_raiseFlags</CODE>. |
| A program may also supply its own <CODE>softfloat_raiseFlags</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_raiseFlags</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><<I>float</I>></CODE> appears in function names as |
| a substitute for one of these abbreviations: |
| <BLOCKQUOTE> |
| <TABLE CELLSPACING=0 CELLPADDING=0> |
| <TR> |
| <TD><CODE>f16</CODE></TD> |
| <TD>indicates <CODE>float16_t</CODE>, passed by value</TD> |
| </TR> |
| <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 </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_<<I>float</I>></CODE><BR> |
| <CODE>ui64_to_<<I>float</I>></CODE><BR> |
| <CODE>i32_to_<<I>float</I>></CODE><BR> |
| <CODE>i64_to_<<I>float</I>></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: |
| <BLOCKQUOTE> |
| <PRE> |
| float64_t i32_to_f64( int32_t <I>a</I> ); |
| </PRE> |
| <PRE> |
| void i32_to_f128M( int32_t <I>a</I>, float128_t *<I>destPtr</I> ); |
| </PRE> |
| </BLOCKQUOTE> |
| </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><<I>float</I>>_to_ui32</CODE><BR> |
| <CODE><<I>float</I>>_to_ui64</CODE><BR> |
| <CODE><<I>float</I>>_to_i32</CODE><BR> |
| <CODE><<I>float</I>>_to_i64</CODE> |
| </BLOCKQUOTE> |
| The functions have signatures as follows, depending on whether the |
| floating-point input is passed by value or via pointers: |
| <BLOCKQUOTE> |
| <PRE> |
| int_fast32_t f64_to_i32( float64_t <I>a</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> ); |
| </PRE> |
| <PRE> |
| int_fast32_t |
| f128M_to_i32( const float128_t *<I>aPtr</I>, uint_fast8_t <I>roundingMode</I>, bool <I>exact</I> ); |
| </PRE> |
| </BLOCKQUOTE> |
| </P> |
| |
| <P> |
| 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> |
| A conversion from floating-point to integer format raises 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 circumstances, the integer result returned is determined by the |
| particular port of SoftFloat, although typically this value will be either the |
| maximum or minimum value of the integer format. |
| The functions that convert to integer types never raise the floating-point |
| <I>overflow</I> exception. |
| </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><<I>float</I>>_to_ui32_r_minMag</CODE><BR> |
| <CODE><<I>float</I>>_to_ui64_r_minMag</CODE><BR> |
| <CODE><<I>float</I>>_to_i32_r_minMag</CODE><BR> |
| <CODE><<I>float</I>>_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: |
| <BLOCKQUOTE> |
| <PRE> |
| int_fast32_t f64_to_i32_r_minMag( float64_t <I>a</I>, bool <I>exact</I> ); |
| </PRE> |
| <PRE> |
| int_fast32_t f128M_to_i32_r_minMag( const float128_t *<I>aPtr</I>, bool <I>exact</I> ); |
| </PRE> |
| </BLOCKQUOTE> |
| </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><<I>float</I>>_to_<<I>float</I>></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: |
| <BLOCKQUOTE> |
| <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> |
| </BLOCKQUOTE> |
| </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><<I>float</I>>_add</CODE><BR> |
| <CODE><<I>float</I>>_sub</CODE><BR> |
| <CODE><<I>float</I>>_mul</CODE><BR> |
| <CODE><<I>float</I>>_div</CODE><BR> |
| <CODE><<I>float</I>>_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: |
| <BLOCKQUOTE> |
| <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> |
| <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> |
| </BLOCKQUOTE> |
| 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 the <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><<I>float</I>>_mulAdd</CODE> |
| </BLOCKQUOTE> |
| Unlike other operations, fused multiple-add is not supported for the |
| <NOBR>80-bit</NOBR> double-extended-precision format, |
| <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: |
| <BLOCKQUOTE> |
| <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> |
| </BLOCKQUOTE> |
| The functions compute |
| <NOBR>(<CODE><I>a</I></CODE> × <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 quiet 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><<I>float</I>>_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: |
| <BLOCKQUOTE> |
| <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> |
| </BLOCKQUOTE> |
| 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> |
| − <I>n</I> × <CODE><I>b</I></CODE></NOBR>, |
| where <I>n</I> is the integer closest to |
| <NOBR><CODE><I>a</I></CODE> ÷ <CODE><I>b</I></CODE></NOBR>. |
| If <NOBR><CODE><I>a</I></CODE> ÷ <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> ÷ <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 an inherent characteristic of 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><<I>float</I>>_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: |
| <BLOCKQUOTE> |
| <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> |
| </BLOCKQUOTE> |
| 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> |
| |
| <P> |
| 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. |
| </P> |
| |
| <H3>8.8. Comparison Functions</H3> |
| |
| <P> |
| For each format, the following floating-point comparison functions are |
| provided: |
| <BLOCKQUOTE> |
| <CODE><<I>float</I>>_eq</CODE><BR> |
| <CODE><<I>float</I>>_le</CODE><BR> |
| <CODE><<I>float</I>>_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” (≤); |
| and <CODE>lt</CODE> stands for “less than” (<). |
| Depending on whether the floating-point operands are passed by value or via |
| pointers, the comparison functions have signatures of these forms: |
| <BLOCKQUOTE> |
| <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> |
| </BLOCKQUOTE> |
| </P> |
| |
| <P> |
| The usual greater-than (>), greater-than-or-equal (≥), and not-equal |
| (≠) 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><<I>float</I>>_eq_signaling</CODE><BR> |
| <CODE><<I>float</I>>_le_quiet</CODE><BR> |
| <CODE><<I>float</I>>_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><<I>float</I>>_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 |
| <BLOCKQUOTE> |
| <PRE> |
| bool f64_isSignalingNaN( float64_t <I>a</I> ); |
| </PRE> |
| <PRE> |
| bool f128M_isSignalingNaN( const float128_t *<I>aPtr</I> ); |
| </PRE> |
| </BLOCKQUOTE> |
| </P> |
| |
| <H3>8.10. Raise-Exception Function</H3> |
| |
| <P> |
| SoftFloat provides a single function for raising floating-point exceptions: |
| <BLOCKQUOTE> |
| <PRE> |
| void softfloat_raiseFlags( uint_fast8_t <I>exceptions</I> ); |
| </PRE> |
| </BLOCKQUOTE> |
| 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_raiseFlags</CODE> |
| function may cause a trap or abort appropriate for the current system. |
| </P> |
| |
| |
| <H2>9. Changes from SoftFloat <NOBR>Release 2</NOBR></H2> |
| |
| <P> |
| Apart from a change in the legal use license, <NOBR>Release 3</NOBR> of |
| SoftFloat introduced numerous technical differences compared to earlier |
| releases. |
| </P> |
| |
| <H3>9.1. Name Changes</H3> |
| |
| <P> |
| The most obvious and pervasive difference 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 </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_raiseFlags</CODE></TD> |
| </TR> |
| </TABLE> |
| </BLOCKQUOTE> |
| </P> |
| |
| <P> |
| Furthermore, <NOBR>Release 3</NOBR> adopted the following new abbreviations for |
| function names: |
| <BLOCKQUOTE> |
| <TABLE> |
| <TR> |
| <TD>used in names in Release 2:<CODE> </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 have been a few other changes to function names: |
| <BLOCKQUOTE> |
| <TABLE> |
| <TR> |
| <TD>used in names in Release 2:<CODE> </CODE></TD> |
| <TD>used in names in Release 3:<CODE> </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 (<NOBR>section 8.2</NOBR>)</TD> |
| </TR> |
| <TR> |
| <TD><CODE>round_to_int</CODE></TD> |
| <TD><CODE>roundToInt</CODE></TD> |
| <TD>round-to-integer functions (<NOBR>section 8.7</NOBR>)</TD> |
| </TR> |
| <TR> |
| <TD><CODE>is_signaling_nan </CODE></TD> |
| <TD><CODE>isSignalingNaN</CODE></TD> |
| <TD>signaling NaN test functions (<NOBR>section 8.9</NOBR>)</TD> |
| </TR> |
| </TABLE> |
| </BLOCKQUOTE> |
| </P> |
| |
| <H3>9.2. Changes to Function Arguments</H3> |
| |
| <P> |
| Besides simple name changes, some operations were given a different interface |
| in <NOBR>Release 3</NOBR> than they had in <NOBR>Release 2</NOBR>: |
| <UL> |
| |
| <LI> |
| <P> |
| Since <NOBR>Release 3</NOBR>, integer arguments and results of functions have |
| standard types from header <CODE><stdint.h></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> and later pass Booleans as |
| standard type <CODE>bool</CODE> from <CODE><stdbool.h></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> and |
| later 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 |
| since <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> or later 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> |
| With <NOBR>Release 3</NOBR>, some new features have been added that were not |
| present in <NOBR>Release 2</NOBR>: |
| <UL> |
| |
| <LI> |
| <P> |
| 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> |
| As of <NOBR>Release 3b</NOBR>, <NOBR>16-bit</NOBR> half-precision, |
| <CODE>float16_t</CODE>, is supported. |
| </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> |
| Fused multiply-add functions have been added for all floating-point formats |
| except <NOBR>80-bit</NOBR> double-extended-precision, |
| <CODE>extFloat80_t</CODE>. |
| </P> |
| |
| <P> |
| <LI> |
| New rounding modes are supported: |
| <CODE>softfloat_round_near_maxMag</CODE> (round to nearest, with ties to |
| maximum magnitude, away from zero), and, as of <NOBR>Release 3c</NOBR>, |
| optional <CODE>softfloat_round_odd</CODE> (round to odd, also known as |
| jamming). |
| </P> |
| |
| </UL> |
| </P> |
| |
| <H3>9.4. Better Compatibility with the C Language</H3> |
| |
| <P> |
| <NOBR>Release 3</NOBR> of SoftFloat was 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> |
| Starting 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 have been 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 usually 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> |
| 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"><NOBR><CODE>http://www.jhauser.us/arithmetic/SoftFloat.html</CODE></NOBR></A>. |
| </P> |
| |
| |
| </BODY> |
| |
