3 * Encoding and decoding floating-point values
5 * (c) 2023 Straylight/Edgeware
8 /*----- Licensing notice --------------------------------------------------*
10 * This file is part of the mLib utilities library.
12 * mLib is free software: you can redistribute it and/or modify it under
13 * the terms of the GNU Library General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or (at
15 * your option) any later version.
17 * mLib is distributed in the hope that it will be useful, but WITHOUT
18 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
19 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU Library General Public
20 * License for more details.
22 * You should have received a copy of the GNU Library General Public
23 * License along with mLib. If not, write to the Free Software
24 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307,
28 /*----- Header files ------------------------------------------------------*/
36 /*----- Formatting primitives ---------------------------------------------*/
38 /* We use the IEEE 754 `binary64' format. Briefly:
40 * * The top bit is the sign %$s$%: 0 encodes %$s = +1$%, and 1 encodes
41 * %$s = -1$%.. The format is signed-magnitude, so everything else is
42 * the same for positive and negative numbers.
44 * * The next eleven bits are the biased exponent %$e$%.
46 * * The remaining 52 bits are the significand %$m$%.
48 * If %$0 < e < 2047$% then the encoding represents the normal number
49 * %$s \cdot (1 + m/2^{52}) \cdot 2^{e-1023}$%.
51 * If %$e = 0$% and %$m = 0$% then the encoding represents positive or
54 * If %$e = 0$% and %$m \ne 0$% then the encoding represents a subnormal
55 * number %$s \cdot m/2^{52} \cdot 2^{-1022}$%.
57 * If %$e = 2047$% and %$m = 0$% then the encoding represents positive or
60 * If %$e = 2047$% and %$m \ne 0$% then the encoding represents a NaN. If
61 * the most significant bit of %$m$% is set then this is a quiet NaN;
62 * otherwise it's a signalling NaN.
65 /* --- @f64_to_k64@ --- *
67 * Arguments: @double x@ = a floating-point number
69 * Returns: A 64-bit encoding of @x@.
71 * Use: Encodes @x@ as a `binary64' value. See `buf_putf64' for the
75 static kludge64 f64_to_k64(double x)
81 /* Some machinery before we start. */
84 # define NANP(x) isnan(x)
86 # define NANP(x) (!((x) == (x)))
90 # define INFP(x) isinf(x)
92 # define INFP(x) ((x) > DBL_MAX || (x) < -DBL_MAX)
96 # define NEGP(x) signbit(x)
98 # define NEGP(x) ((x) < 0) /* incorrect for negative zero! */
103 hi = 0x7ff80000; lo = 0;
104 } else if (INFP(x)) {
105 /* Positive or negative infinity. */
106 hi = NEGP(x) ? 0xfff00000 : 0x7ff00000; lo = 0;
108 /* Positive or negative zero. */
109 hi = NEGP(x) ? 0x80000000 : 0; lo = 0;
111 /* A normal or subnormal number. Now we have to do some actual work. */
113 /* Let's get the sign dealt with so we don't have to worry about it any
116 if (!NEGP(x)) hi = 0;
117 else { x = -x; hi = 0x80000000; }
119 /* Now we start on the value. The first thing to do is to split off the
120 * exponent. Our number will be %$m \cdot 2^e$%, with %$1/2 \le m < 1$%.
124 /* If our number is too big, we'll round it to infinity. This will
125 * happen if %$x \ge 2^{1024}$%, i.e., if %$e > 1024$%.
128 { hi |= 0x7ff00000; lo = 0; }
130 /* Our number is sufficiently small that we can represent it at least
131 * approximately (though maybe we'll have to flush it to zero). The
132 * next step, then, is to pull the significand bits out.
135 /* Determine the correct exponent to store. We're not going to bias it
136 * yet, but this is where we deal with subnormal numbers. Our number
137 * is normal if %$x \ge 2^{-1022}$%, i.e., %$e > -1022$%. In this
138 * case, there's an implicit bit which we'll clear. Otherwise, if it's
139 * subnormal, we'll scale our floating-point number so that the
140 * significand will look right when we extract it, and adjust the
141 * exponent so that, when we're finally done, it will have the correct
144 if (e > -1022) m -= 0.5;
145 else { m = ldexp(m, 1021 + e); e = -1022; }
147 /* Now we pull out the 53 bits of the significand. This will, in
148 * general, leave a tail which we address through rounding. Scale it
149 * up so that we end up with %$0 \le m' < 2$%; then we round up if
150 * %$m > 1$%, or if %$m = 1$% and the low bit of the significand is
153 t = ldexp(m, 21); m -= ldexp(t, -21);
154 lo = ldexp(m, 53); m -= ldexp(lo, -53);
157 /* Round the number if necessary. */
158 if (lo&1 ? m >= 1.0 : m > 1)
159 { lo = U32(lo + 1); if (!lo) t++; }
161 /* Now we just put the pieces together. Note that our %$e$% is one
162 * greater than it should be, because our implicit bit should have
163 * been the unit bit not the 1/2 bit.
165 hi |= ((uint32)(e + 1022) << 20) | t;
169 /* Convert to external format and go home. */
170 SET64(k, hi, lo); return (k);
177 /* --- @k64_to_f64@ --- *
179 * Arguments: @double *x_out@ = where to put the result
180 * @kludge64 k@ = a 64-bit encoding of a floating-point value
182 * Returns: Zero on success, @-1@ on failure.
184 * Use: Decodes @k@ as a `binary64' value. See `buf_getf64' for the
188 static int k64_to_f64(double *x_out, kludge64 k)
193 /* We're using the IEEE 754 `binary64' format: see `float_to_k64' above. */
195 /* Pick the encoded number apart. */
196 hi = HI64(k); lo = LO64(k);
197 s = (hi >> 31)&1; e = (hi >> 20)&0x07ff; t = hi&0x000fffff;
199 /* Deal with various special cases. */
201 /* Maximum exponent indicates (positive or negative) infinity or NaN. */
204 /* It's a NaN. We're not going to be picky about which one. If we
205 * can't represent it then we'll just have to fail.
214 /* It's an infinity. If we don't have one of those to hand, then pick
215 * something really big.
219 x = s ? -INFINITY : INFINITY;
221 x = s ? -DBL_MAX : DBL_MAX;
225 /* It's a finite number, though maybe it's weird in some way. */
228 /* Minimum exponent indicates zero or a subnormal number. The
229 * subnormal exponent is a sentinel value that shouldn't be taken
230 * literally, so we should fix that. If the number is actually zero
231 * then the exponent won't matter much so don't bother checking.
236 /* It's a normal number. In which case there's an implicit bit which
243 /* All that remains is to stuff the significant and exponent into a
244 * floating point number. We'll have to do this in pieces, and we'll
245 * lean on the floating-point machinery to do rounding correctly.
247 x = ldexp(t, e - 1043) + ldexp(lo, e - 1075);
251 /* And we're done. */
252 *x_out = x; return (0);
255 /*----- External functions ------------------------------------------------*/
257 /* --- @buf_putf64{,b,l} --- *
259 * Arguments: @buf *b@ = a buffer to write to
260 * @double x@ = a number to write
262 * Returns: Zero on success, @-1@ on failure (and the buffer is broken).
264 * On C89, this function can't detect negative zero so these
265 * will be silently written as positive zero.
267 * This function doesn't distinguish NaNs. Any NaN is written
268 * as a quiet NaN with all payload bits zero.
270 * A finite value with too large a magnitude to be represented
271 * is rounded to the appropriate infinity. Other finite values
272 * are rounded as necessary, in the usual IEEE 754 round-to-
273 * nearest-or-even way.
276 int buf_putf64(buf *b, double x)
277 { return (buf_putk64(b, f64_to_k64(x))); }
278 int buf_putf64b(buf *b, double x)
279 { return (buf_putk64b(b, f64_to_k64(x))); }
280 int buf_putf64l(buf *b, double x)
281 { return (buf_putk64l(b, f64_to_k64(x))); }
283 /* --- @buf_getf64{,b,l} --- *
285 * Arguments: @buf *b@ = a buffer to read from
286 * @double *x_out@ = where to put the result
288 * Returns: Zero on success, @-1@ on failure (and the buffer is broken).
290 * If the system supports NaNs, then any encoded NaN is returned
291 * as the value of @NAN@ in @<math.h>@; otherwise, this function
294 * In general, values are rounded to the nearest available
295 * value, in the way that the system usually rounds. If the
296 * system doesn't support infinities, then any encoded infinity
297 * is reported as the largest-possible-magnitude finite value
301 int buf_getf64(buf *b, double *x_out)
305 if (buf_getk64(b, &k)) return (-1);
306 if (k64_to_f64(x_out, k)) { b->f |= BF_BROKEN; return (-1); }
309 int buf_getf64b(buf *b, double *x_out)
313 if (buf_getk64b(b, &k)) return (-1);
314 if (k64_to_f64(x_out, k)) { b->f |= BF_BROKEN; return (-1); }
317 int buf_getf64l(buf *b, double *x_out)
321 if (buf_getk64l(b, &k)) return (-1);
322 if (k64_to_f64(x_out, k)) { b->f |= BF_BROKEN; return (-1); }
326 /*----- That's all, folks -------------------------------------------------*/