1 $NetBSD: softfloat.txt,v 1.2 2006/11/24 19:46:58 christos Exp $
3 SoftFloat Release 2a General Documentation
9 -------------------------------------------------------------------------------
12 SoftFloat is a software implementation of floating-point that conforms to
13 the IEC/IEEE Standard for Binary Floating-Point Arithmetic. As many as four
14 formats are supported: single precision, double precision, extended double
15 precision, and quadruple precision. All operations required by the standard
16 are implemented, except for conversions to and from decimal.
18 This document gives information about the types defined and the routines
19 implemented by SoftFloat. It does not attempt to define or explain the
20 IEC/IEEE Floating-Point Standard. Details about the standard are available
24 -------------------------------------------------------------------------------
27 SoftFloat is written in C and is designed to work with other C code. The
28 SoftFloat header files assume an ISO/ANSI-style C compiler. No attempt
29 has been made to accommodate compilers that are not ISO-conformant. In
30 particular, the distributed header files will not be acceptable to any
31 compiler that does not recognize function prototypes.
33 Support for the extended double-precision and quadruple-precision formats
34 depends on a C compiler that implements 64-bit integer arithmetic. If the
35 largest integer format supported by the C compiler is 32 bits, SoftFloat is
36 limited to only single and double precisions. When that is the case, all
37 references in this document to the extended double precision, quadruple
38 precision, and 64-bit integers should be ignored.
41 -------------------------------------------------------------------------------
50 Extended Double-Precision Rounding Precision
51 Exceptions and Exception Flags
54 Standard Arithmetic Functions
56 Round-to-Integer Functions
58 Signaling NaN Test Functions
59 Raise-Exception Function
64 -------------------------------------------------------------------------------
67 SoftFloat was written by John R. Hauser. This work was made possible in
68 part by the International Computer Science Institute, located at Suite 600,
69 1947 Center Street, Berkeley, California 94704. Funding was partially
70 provided by the National Science Foundation under grant MIP-9311980. The
71 original version of this code was written as part of a project to build
72 a fixed-point vector processor in collaboration with the University of
73 California at Berkeley, overseen by Profs. Nelson Morgan and John Wawrzynek.
75 THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort
76 has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
77 TIMES RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO
78 PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
79 AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
82 -------------------------------------------------------------------------------
85 When 64-bit integers are supported by the compiler, the `softfloat.h' header
86 file defines four types: `float32' (single precision), `float64' (double
87 precision), `floatx80' (extended double precision), and `float128'
88 (quadruple precision). The `float32' and `float64' types are defined in
89 terms of 32-bit and 64-bit integer types, respectively, while the `float128'
90 type is defined as a structure of two 64-bit integers, taking into account
91 the byte order of the particular machine being used. The `floatx80' type
92 is defined as a structure containing one 16-bit and one 64-bit integer, with
93 the machine's byte order again determining the order of the `high' and `low'
96 When 64-bit integers are _not_ supported by the compiler, the `softfloat.h'
97 header file defines only two types: `float32' and `float64'. Because
98 ISO/ANSI C guarantees at least one built-in integer type of 32 bits,
99 the `float32' type is identified with an appropriate integer type. The
100 `float64' type is defined as a structure of two 32-bit integers, with the
101 machine's byte order determining the order of the fields.
103 In either case, the types in `softfloat.h' are defined such that if a system
104 implements the usual C `float' and `double' types according to the IEC/IEEE
105 Standard, then the `float32' and `float64' types should be indistinguishable
106 in memory from the native `float' and `double' types. (On the other hand,
107 when `float32' or `float64' values are placed in processor registers by
108 the compiler, the type of registers used may differ from those used for the
109 native `float' and `double' types.)
111 SoftFloat implements the following arithmetic operations:
113 -- Conversions among all the floating-point formats, and also between
114 integers (32-bit and 64-bit) and any of the floating-point formats.
116 -- The usual add, subtract, multiply, divide, and square root operations
117 for all floating-point formats.
119 -- For each format, the floating-point remainder operation defined by the
122 -- For each floating-point format, a ``round to integer'' operation that
123 rounds to the nearest integer value in the same format. (The floating-
124 point formats can hold integer values, of course.)
126 -- Comparisons between two values in the same floating-point format.
128 The only functions required by the IEC/IEEE Standard that are not provided
129 are conversions to and from decimal.
132 -------------------------------------------------------------------------------
135 All four rounding modes prescribed by the IEC/IEEE Standard are implemented
136 for all operations that require rounding. The rounding mode is selected
137 by the global variable `float_rounding_mode'. This variable may be set
138 to one of the values `float_round_nearest_even', `float_round_to_zero',
139 `float_round_down', or `float_round_up'. The rounding mode is initialized
143 -------------------------------------------------------------------------------
144 Extended Double-Precision Rounding Precision
146 For extended double precision (`floatx80') only, the rounding precision
147 of the standard arithmetic operations is controlled by the global variable
148 `floatx80_rounding_precision'. The operations affected are:
150 floatx80_add floatx80_sub floatx80_mul floatx80_div floatx80_sqrt
152 When `floatx80_rounding_precision' is set to its default value of 80, these
153 operations are rounded (as usual) to the full precision of the extended
154 double-precision format. Setting `floatx80_rounding_precision' to 32
155 or to 64 causes the operations listed to be rounded to reduced precision
156 equivalent to single precision (`float32') or to double precision
157 (`float64'), respectively. When rounding to reduced precision, additional
158 bits in the result significand beyond the rounding point are set to zero.
159 The consequences of setting `floatx80_rounding_precision' to a value other
160 than 32, 64, or 80 is not specified. Operations other than the ones listed
161 above are not affected by `floatx80_rounding_precision'.
164 -------------------------------------------------------------------------------
165 Exceptions and Exception Flags
167 All five exception flags required by the IEC/IEEE Standard are
168 implemented. Each flag is stored as a unique bit in the global variable
169 `float_exception_flags'. The positions of the exception flag bits within
170 this variable are determined by the bit masks `float_flag_inexact',
171 `float_flag_underflow', `float_flag_overflow', `float_flag_divbyzero', and
172 `float_flag_invalid'. The exception flags variable is initialized to all 0,
173 meaning no exceptions.
175 An individual exception flag can be cleared with the statement
177 float_exception_flags &= ~ float_flag_<exception>;
179 where `<exception>' is the appropriate name. To raise a floating-point
180 exception, the SoftFloat function `float_raise' should be used (see below).
182 In the terminology of the IEC/IEEE Standard, SoftFloat can detect tininess
183 for underflow either before or after rounding. The choice is made by
184 the global variable `float_detect_tininess', which can be set to either
185 `float_tininess_before_rounding' or `float_tininess_after_rounding'.
186 Detecting tininess after rounding is better because it results in fewer
187 spurious underflow signals. The other option is provided for compatibility
188 with some systems. Like most systems, SoftFloat always detects loss of
189 accuracy for underflow as an inexact result.
192 -------------------------------------------------------------------------------
195 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
198 All conversions among the floating-point formats are supported, as are all
199 conversions between a floating-point format and 32-bit and 64-bit signed
200 integers. The complete set of conversion functions is:
202 int32_to_float32 int64_to_float32
203 int32_to_float64 int64_to_float32
204 int32_to_floatx80 int64_to_floatx80
205 int32_to_float128 int64_to_float128
207 float32_to_int32 float32_to_int64
208 float32_to_int32 float64_to_int64
209 floatx80_to_int32 floatx80_to_int64
210 float128_to_int32 float128_to_int64
212 float32_to_float64 float32_to_floatx80 float32_to_float128
213 float64_to_float32 float64_to_floatx80 float64_to_float128
214 floatx80_to_float32 floatx80_to_float64 floatx80_to_float128
215 float128_to_float32 float128_to_float64 float128_to_floatx80
217 Each conversion function takes one operand of the appropriate type and
218 returns one result. Conversions from a smaller to a larger floating-point
219 format are always exact and so require no rounding. Conversions from 32-bit
220 integers to double precision and larger formats are also exact, and likewise
221 for conversions from 64-bit integers to extended double and quadruple
224 Conversions from floating-point to integer raise the invalid exception if
225 the source value cannot be rounded to a representable integer of the desired
226 size (32 or 64 bits). If the floating-point operand is a NaN, the largest
227 positive integer is returned. Otherwise, if the conversion overflows, the
228 largest integer with the same sign as the operand is returned.
230 On conversions to integer, if the floating-point operand is not already an
231 integer value, the operand is rounded according to the current rounding
232 mode as specified by `float_rounding_mode'. Because C (and perhaps other
233 languages) require that conversions to integers be rounded toward zero, the
234 following functions are provided for improved speed and convenience:
236 float32_to_int32_round_to_zero float32_to_int64_round_to_zero
237 float64_to_int32_round_to_zero float64_to_int64_round_to_zero
238 floatx80_to_int32_round_to_zero floatx80_to_int64_round_to_zero
239 float128_to_int32_round_to_zero float128_to_int64_round_to_zero
241 These variant functions ignore `float_rounding_mode' and always round toward
244 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
245 Standard Arithmetic Functions
247 The following standard arithmetic functions are provided:
249 float32_add float32_sub float32_mul float32_div float32_sqrt
250 float64_add float64_sub float64_mul float64_div float64_sqrt
251 floatx80_add floatx80_sub floatx80_mul floatx80_div floatx80_sqrt
252 float128_add float128_sub float128_mul float128_div float128_sqrt
254 Each function takes two operands, except for `sqrt' which takes only one.
255 The operands and result are all of the same type.
257 Rounding of the extended double-precision (`floatx80') functions is affected
258 by the `floatx80_rounding_precision' variable, as explained above in the
259 section _Extended_Double-Precision_Rounding_Precision_.
261 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
264 For each format, SoftFloat implements the remainder function according to
265 the IEC/IEEE Standard. The remainder functions are:
272 Each remainder function takes two operands. The operands and result are all
273 of the same type. Given operands x and y, the remainder functions return
274 the value x - n*y, where n is the integer closest to x/y. If x/y is exactly
275 halfway between two integers, n is the even integer closest to x/y. The
276 remainder functions are always exact and so require no rounding.
278 Depending on the relative magnitudes of the operands, the remainder
279 functions can take considerably longer to execute than the other SoftFloat
280 functions. This is inherent in the remainder operation itself and is not a
281 flaw in the SoftFloat implementation.
283 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
284 Round-to-Integer Functions
286 For each format, SoftFloat implements the round-to-integer function
287 specified by the IEC/IEEE Standard. The functions are:
291 floatx80_round_to_int
292 float128_round_to_int
294 Each function takes a single floating-point operand and returns a result of
295 the same type. (Note that the result is not an integer type.) The operand
296 is rounded to an exact integer according to the current rounding mode, and
297 the resulting integer value is returned in the same floating-point format.
299 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
302 The following floating-point comparison functions are provided:
304 float32_eq float32_le float32_lt
305 float64_eq float64_le float64_lt
306 floatx80_eq floatx80_le floatx80_lt
307 float128_eq float128_le float128_lt
309 Each function takes two operands of the same type and returns a 1 or 0
310 representing either _true_ or _false_. The abbreviation `eq' stands for
311 ``equal'' (=); `le' stands for ``less than or equal'' (<=); and `lt' stands
312 for ``less than'' (<).
314 The standard greater-than (>), greater-than-or-equal (>=), and not-equal
315 (!=) functions are easily obtained using the functions provided. The
316 not-equal function is just the logical complement of the equal function.
317 The greater-than-or-equal function is identical to the less-than-or-equal
318 function with the operands reversed; and the greater-than function can be
319 obtained from the less-than function in the same way.
321 The IEC/IEEE Standard specifies that the less-than-or-equal and less-than
322 functions raise the invalid exception if either input is any kind of NaN.
323 The equal functions, on the other hand, are defined not to raise the invalid
324 exception on quiet NaNs. For completeness, SoftFloat provides the following
325 additional functions:
327 float32_eq_signaling float32_le_quiet float32_lt_quiet
328 float64_eq_signaling float64_le_quiet float64_lt_quiet
329 floatx80_eq_signaling floatx80_le_quiet floatx80_lt_quiet
330 float128_eq_signaling float128_le_quiet float128_lt_quiet
332 The `signaling' equal functions are identical to the standard functions
333 except that the invalid exception is raised for any NaN input. Likewise,
334 the `quiet' comparison functions are identical to their counterparts except
335 that the invalid exception is not raised for quiet NaNs.
337 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
338 Signaling NaN Test Functions
340 The following functions test whether a floating-point value is a signaling
343 float32_is_signaling_nan
344 float64_is_signaling_nan
345 floatx80_is_signaling_nan
346 float128_is_signaling_nan
348 The functions take one operand and return 1 if the operand is a signaling
351 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
352 Raise-Exception Function
354 SoftFloat provides a function for raising floating-point exceptions:
358 The function takes a mask indicating the set of exceptions to raise. No
359 result is returned. In addition to setting the specified exception flags,
360 this function may cause a trap or abort appropriate for the current system.
362 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
365 -------------------------------------------------------------------------------
368 At the time of this writing, the most up-to-date information about
369 SoftFloat and the latest release can be found at the Web page `http://
370 HTTP.CS.Berkeley.EDU/~jhauser/arithmetic/SoftFloat.html'.