xref: /titanic_51/usr/src/lib/libm/common/m9x/fma.c (revision 177d5b5f8c0e969013441207a0a705ae66b08cf7)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #if defined(ELFOBJ)
31 #pragma weak fma = __fma
32 #endif
33 
34 #include "libm.h"
35 #include "fma.h"
36 #include "fenv_inlines.h"
37 
38 #if defined(__sparc)
39 
40 static const union {
41 	unsigned i[2];
42 	double d;
43 } C[] = {
44 	{ 0x3fe00000u, 0 },
45 	{ 0x40000000u, 0 },
46 	{ 0x43300000u, 0 },
47 	{ 0x41a00000u, 0 },
48 	{ 0x3e500000u, 0 },
49 	{ 0x3df00000u, 0 },
50 	{ 0x3bf00000u, 0 },
51 	{ 0x7fe00000u, 0 },
52 	{ 0x00100000u, 0 },
53 	{ 0x00100001u, 0 }
54 };
55 
56 #define	half	C[0].d
57 #define	two	C[1].d
58 #define	two52	C[2].d
59 #define	two27	C[3].d
60 #define	twom26	C[4].d
61 #define	twom32	C[5].d
62 #define	twom64	C[6].d
63 #define	huge	C[7].d
64 #define	tiny	C[8].d
65 #define	tiny2	C[9].d
66 
67 static const unsigned int fsr_rm = 0xc0000000u;
68 
69 /*
70  * fma for SPARC: 64-bit double precision, big-endian
71  */
72 double
73 __fma(double x, double y, double z) {
74 	union {
75 		unsigned i[2];
76 		double d;
77 	} xx, yy, zz;
78 	double xhi, yhi, xlo, ylo, t;
79 	unsigned int xy0, xy1, xy2, xy3, z0, z1, z2, z3, fsr, rm, sticky;
80 	int hx, hy, hz, ex, ey, ez, exy, sxy, sz, e, ibit;
81 	volatile double	dummy;
82 
83 	/* extract the high order words of the arguments */
84 	xx.d = x;
85 	yy.d = y;
86 	zz.d = z;
87 	hx = xx.i[0] & ~0x80000000;
88 	hy = yy.i[0] & ~0x80000000;
89 	hz = zz.i[0] & ~0x80000000;
90 
91 	/* dispense with inf, nan, and zero cases */
92 	if (hx >= 0x7ff00000 || hy >= 0x7ff00000 || (hx | xx.i[1]) == 0 ||
93 		(hy | yy.i[1]) == 0)	/* x or y is inf, nan, or zero */
94 		return (x * y + z);
95 
96 	if (hz >= 0x7ff00000)	/* z is inf or nan */
97 		return (x + z);	/* avoid spurious under/overflow in x * y */
98 
99 	if ((hz | zz.i[1]) == 0)	/* z is zero */
100 		/*
101 		 * x * y isn't zero but could underflow to zero,
102 		 * so don't add z, lest we perturb the sign
103 		 */
104 		return (x * y);
105 
106 	/*
107 	 * now x, y, and z are all finite and nonzero; save the fsr and
108 	 * set round-to-negative-infinity mode (and clear nonstandard
109 	 * mode before we try to scale subnormal operands)
110 	 */
111 	__fenv_getfsr32(&fsr);
112 	__fenv_setfsr32(&fsr_rm);
113 
114 	/* extract signs and exponents, and normalize subnormals */
115 	sxy = (xx.i[0] ^ yy.i[0]) & 0x80000000;
116 	sz = zz.i[0] & 0x80000000;
117 	ex = hx >> 20;
118 	if (!ex) {
119 		xx.d = x * two52;
120 		ex = ((xx.i[0] & ~0x80000000) >> 20) - 52;
121 	}
122 	ey = hy >> 20;
123 	if (!ey) {
124 		yy.d = y * two52;
125 		ey = ((yy.i[0] & ~0x80000000) >> 20) - 52;
126 	}
127 	ez = hz >> 20;
128 	if (!ez) {
129 		zz.d = z * two52;
130 		ez = ((zz.i[0] & ~0x80000000) >> 20) - 52;
131 	}
132 
133 	/* multiply x*y to 106 bits */
134 	exy = ex + ey - 0x3ff;
135 	xx.i[0] = (xx.i[0] & 0xfffff) | 0x3ff00000;
136 	yy.i[0] = (yy.i[0] & 0xfffff) | 0x3ff00000;
137 	x = xx.d;
138 	y = yy.d;
139 	xhi = ((x + twom26) + two27) - two27;
140 	yhi = ((y + twom26) + two27) - two27;
141 	xlo = x - xhi;
142 	ylo = y - yhi;
143 	x *= y;
144 	y = ((xhi * yhi - x) + xhi * ylo + xlo * yhi) + xlo * ylo;
145 	if (x >= two) {
146 		x *= half;
147 		y *= half;
148 		exy++;
149 	}
150 
151 	/* extract the significands */
152 	xx.d = x;
153 	xy0 = (xx.i[0] & 0xfffff) | 0x100000;
154 	xy1 = xx.i[1];
155 	yy.d = t = y + twom32;
156 	xy2 = yy.i[1];
157 	yy.d = (y - (t - twom32)) + twom64;
158 	xy3 = yy.i[1];
159 	z0 = (zz.i[0] & 0xfffff) | 0x100000;
160 	z1 = zz.i[1];
161 	z2 = z3 = 0;
162 
163 	/*
164 	 * now x*y is represented by sxy, exy, and xy[0-3], and z is
165 	 * represented likewise; swap if need be so |xy| <= |z|
166 	 */
167 	if (exy > ez || (exy == ez && (xy0 > z0 || (xy0 == z0 &&
168 		(xy1 > z1 || (xy1 == z1 && (xy2 | xy3) != 0)))))) {
169 		e = sxy; sxy = sz; sz = e;
170 		e = exy; exy = ez; ez = e;
171 		e = xy0; xy0 = z0; z0 = e;
172 		e = xy1; xy1 = z1; z1 = e;
173 		z2 = xy2; xy2 = 0;
174 		z3 = xy3; xy3 = 0;
175 	}
176 
177 	/* shift the significand of xy keeping a sticky bit */
178 	e = ez - exy;
179 	if (e > 116) {
180 		xy0 = xy1 = xy2 = 0;
181 		xy3 = 1;
182 	} else if (e >= 96) {
183 		sticky = xy3 | xy2 | xy1 | ((xy0 << 1) << (127 - e));
184 		xy3 = xy0 >> (e - 96);
185 		if (sticky)
186 			xy3 |= 1;
187 		xy0 = xy1 = xy2 = 0;
188 	} else if (e >= 64) {
189 		sticky = xy3 | xy2 | ((xy1 << 1) << (95 - e));
190 		xy3 = (xy1 >> (e - 64)) | ((xy0 << 1) << (95 - e));
191 		if (sticky)
192 			xy3 |= 1;
193 		xy2 = xy0 >> (e - 64);
194 		xy0 = xy1 = 0;
195 	} else if (e >= 32) {
196 		sticky = xy3 | ((xy2 << 1) << (63 - e));
197 		xy3 = (xy2 >> (e - 32)) | ((xy1 << 1) << (63 - e));
198 		if (sticky)
199 			xy3 |= 1;
200 		xy2 = (xy1 >> (e - 32)) | ((xy0 << 1) << (63 - e));
201 		xy1 = xy0 >> (e - 32);
202 		xy0 = 0;
203 	} else if (e) {
204 		sticky = (xy3 << 1) << (31 - e);
205 		xy3 = (xy3 >> e) | ((xy2 << 1) << (31 - e));
206 		if (sticky)
207 			xy3 |= 1;
208 		xy2 = (xy2 >> e) | ((xy1 << 1) << (31 - e));
209 		xy1 = (xy1 >> e) | ((xy0 << 1) << (31 - e));
210 		xy0 >>= e;
211 	}
212 
213 	/* if this is a magnitude subtract, negate the significand of xy */
214 	if (sxy ^ sz) {
215 		xy0 = ~xy0;
216 		xy1 = ~xy1;
217 		xy2 = ~xy2;
218 		xy3 = -xy3;
219 		if (xy3 == 0)
220 			if (++xy2 == 0)
221 				if (++xy1 == 0)
222 					xy0++;
223 	}
224 
225 	/* add, propagating carries */
226 	z3 += xy3;
227 	e = (z3 < xy3);
228 	z2 += xy2;
229 	if (e) {
230 		z2++;
231 		e = (z2 <= xy2);
232 	} else
233 		e = (z2 < xy2);
234 	z1 += xy1;
235 	if (e) {
236 		z1++;
237 		e = (z1 <= xy1);
238 	} else
239 		e = (z1 < xy1);
240 	z0 += xy0;
241 	if (e)
242 		z0++;
243 
244 	/* postnormalize and collect rounding information into z2 */
245 	if (ez < 1) {
246 		/* result is tiny; shift right until exponent is within range */
247 		e = 1 - ez;
248 		if (e > 56) {
249 			z2 = 1;	/* result can't be exactly zero */
250 			z0 = z1 = 0;
251 		} else if (e >= 32) {
252 			sticky = z3 | z2 | ((z1 << 1) << (63 - e));
253 			z2 = (z1 >> (e - 32)) | ((z0 << 1) << (63 - e));
254 			if (sticky)
255 				z2 |= 1;
256 			z1 = z0 >> (e - 32);
257 			z0 = 0;
258 		} else {
259 			sticky = z3 | (z2 << 1) << (31 - e);
260 			z2 = (z2 >> e) | ((z1 << 1) << (31 - e));
261 			if (sticky)
262 				z2 |= 1;
263 			z1 = (z1 >> e) | ((z0 << 1) << (31 - e));
264 			z0 >>= e;
265 		}
266 		ez = 1;
267 	} else if (z0 >= 0x200000) {
268 		/* carry out; shift right by one */
269 		sticky = (z2 & 1) | z3;
270 		z2 = (z2 >> 1) | (z1 << 31);
271 		if (sticky)
272 			z2 |= 1;
273 		z1 = (z1 >> 1) | (z0 << 31);
274 		z0 >>= 1;
275 		ez++;
276 	} else {
277 		if (z0 < 0x100000 && (z0 | z1 | z2 | z3) != 0) {
278 			/*
279 			 * borrow/cancellation; shift left as much as
280 			 * exponent allows
281 			 */
282 			while (!(z0 | (z1 & 0xffe00000)) && ez >= 33) {
283 				z0 = z1;
284 				z1 = z2;
285 				z2 = z3;
286 				z3 = 0;
287 				ez -= 32;
288 			}
289 			while (z0 < 0x100000 && ez > 1) {
290 				z0 = (z0 << 1) | (z1 >> 31);
291 				z1 = (z1 << 1) | (z2 >> 31);
292 				z2 = (z2 << 1) | (z3 >> 31);
293 				z3 <<= 1;
294 				ez--;
295 			}
296 		}
297 		if (z3)
298 			z2 |= 1;
299 	}
300 
301 	/* get the rounding mode and clear current exceptions */
302 	rm = fsr >> 30;
303 	fsr &= ~FSR_CEXC;
304 
305 	/* strip off the integer bit, if there is one */
306 	ibit = z0 & 0x100000;
307 	if (ibit)
308 		z0 -= 0x100000;
309 	else {
310 		ez = 0;
311 		if (!(z0 | z1 | z2)) { /* exact zero */
312 			zz.i[0] = rm == FSR_RM ? 0x80000000 : 0;
313 			zz.i[1] = 0;
314 			__fenv_setfsr32(&fsr);
315 			return (zz.d);
316 		}
317 	}
318 
319 	/*
320 	 * flip the sense of directed roundings if the result is negative;
321 	 * the logic below applies to a positive result
322 	 */
323 	if (sz)
324 		rm ^= rm >> 1;
325 
326 	/* round and raise exceptions */
327 	if (z2) {
328 		fsr |= FSR_NXC;
329 
330 		/* decide whether to round the fraction up */
331 		if (rm == FSR_RP || (rm == FSR_RN && (z2 > 0x80000000u ||
332 			(z2 == 0x80000000u && (z1 & 1))))) {
333 			/* round up and renormalize if necessary */
334 			if (++z1 == 0) {
335 				if (++z0 == 0x100000) {
336 					z0 = 0;
337 					ez++;
338 				}
339 			}
340 		}
341 	}
342 
343 	/* check for under/overflow */
344 	if (ez >= 0x7ff) {
345 		if (rm == FSR_RN || rm == FSR_RP) {
346 			zz.i[0] = sz | 0x7ff00000;
347 			zz.i[1] = 0;
348 		} else {
349 			zz.i[0] = sz | 0x7fefffff;
350 			zz.i[1] = 0xffffffff;
351 		}
352 		fsr |= FSR_OFC | FSR_NXC;
353 	} else {
354 		zz.i[0] = sz | (ez << 20) | z0;
355 		zz.i[1] = z1;
356 
357 		/*
358 		 * !ibit => exact result was tiny before rounding,
359 		 * z2 nonzero => result delivered is inexact
360 		 */
361 		if (!ibit) {
362 			if (z2)
363 				fsr |= FSR_UFC | FSR_NXC;
364 			else if (fsr & FSR_UFM)
365 				fsr |= FSR_UFC;
366 		}
367 	}
368 
369 	/* restore the fsr and emulate exceptions as needed */
370 	if ((fsr & FSR_CEXC) & (fsr >> 23)) {
371 		__fenv_setfsr32(&fsr);
372 		if (fsr & FSR_OFC) {
373 			dummy = huge;
374 			dummy *= huge;
375 		} else if (fsr & FSR_UFC) {
376 			dummy = tiny;
377 			if (fsr & FSR_NXC)
378 				dummy *= tiny;
379 			else
380 				dummy -= tiny2;
381 		} else {
382 			dummy = huge;
383 			dummy += tiny;
384 		}
385 	} else {
386 		fsr |= (fsr & 0x1f) << 5;
387 		__fenv_setfsr32(&fsr);
388 	}
389 	return (zz.d);
390 }
391 
392 #elif defined(__x86)
393 
394 #if defined(__amd64)
395 #define	NI	4
396 #else
397 #define	NI	3
398 #endif
399 
400 /*
401  *  fma for x86: 64-bit double precision, little-endian
402  */
403 double
404 __fma(double x, double y, double z) {
405 	union {
406 		unsigned i[NI];
407 		long double e;
408 	} xx, yy, zz;
409 	long double xe, ye, xhi, xlo, yhi, ylo;
410 	int ex, ey, ez;
411 	unsigned cwsw, oldcwsw, rm;
412 
413 	/* convert the operands to double extended */
414 	xx.e = (long double) x;
415 	yy.e = (long double) y;
416 	zz.e = (long double) z;
417 
418 	/* extract the exponents of the arguments */
419 	ex = xx.i[2] & 0x7fff;
420 	ey = yy.i[2] & 0x7fff;
421 	ez = zz.i[2] & 0x7fff;
422 
423 	/* dispense with inf, nan, and zero cases */
424 	if (ex == 0x7fff || ey == 0x7fff || ex == 0 || ey == 0)
425 		/* x or y is inf, nan, or zero */
426 		return ((double) (xx.e * yy.e + zz.e));
427 
428 	if (ez >= 0x7fff) /* z is inf or nan */
429 		return ((double) (xx.e + zz.e));
430 					/* avoid spurious inexact in x * y */
431 
432 	/*
433 	 * save the control and status words, mask all exceptions, and
434 	 * set rounding to 64-bit precision and to-nearest
435 	 */
436 	__fenv_getcwsw(&oldcwsw);
437 	cwsw = (oldcwsw & 0xf0c0ffff) | 0x033f0000;
438 	__fenv_setcwsw(&cwsw);
439 
440 	/* multiply x*y to 106 bits */
441 	xe = xx.e;
442 	xx.i[0] = 0;
443 	xhi = xx.e; /* hi 32 bits */
444 	xlo = xe - xhi; /* lo 21 bits */
445 	ye = yy.e;
446 	yy.i[0] = 0;
447 	yhi = yy.e;
448 	ylo = ye - yhi;
449 	xe = xe * ye;
450 	ye = ((xhi * yhi - xe) + xhi * ylo + xlo * yhi) + xlo * ylo;
451 
452 	/* distill the sum of xe, ye, and z */
453 	xhi = ye + zz.e;
454 	yhi = xhi - ye;
455 	xlo = (zz.e - yhi) + (ye - (xhi - yhi));
456 						/* now (xhi,xlo) = ye + z */
457 
458 	yhi = xe + xhi;
459 	ye = yhi - xe;
460 	ylo = (xhi - ye) + (xe - (yhi - ye));	/* now (yhi,ylo) = xe + xhi */
461 
462 	xhi = xlo + ylo;
463 	xe = xhi - xlo;
464 	xlo = (ylo - xe) + (xlo - (xhi - xe));	/* now (xhi,xlo) = xlo + ylo */
465 
466 	yy.e = yhi + xhi;
467 	ylo = (yhi - yy.e) + xhi;		/* now (yy.e,ylo) = xhi + yhi */
468 
469 	if (yy.i[1] != 0) {	/* yy.e is nonzero */
470 		/* perturb yy.e if its least significant 10 bits are zero */
471 		if (!(yy.i[0] & 0x3ff)) {
472 			xx.e = ylo + xlo;
473 			if (xx.i[1] != 0) {
474 				xx.i[2] = (xx.i[2] & 0x8000) |
475 					((yy.i[2] & 0x7fff) - 63);
476 				xx.i[1] = 0x80000000;
477 				xx.i[0] = 0;
478 				yy.e += xx.e;
479 			}
480 		}
481 	} else {
482 		/* set sign of zero result according to rounding direction */
483 		rm = oldcwsw & 0x0c000000;
484 		yy.i[2] = ((rm == FCW_RM)? 0x8000 : 0);
485 	}
486 
487 	/*
488 	 * restore the control and status words and convert the result
489 	 * to double
490 	 */
491 	__fenv_setcwsw(&oldcwsw);
492 	return ((double) yy.e);
493 }
494 
495 #else
496 #error Unknown architecture
497 #endif
498