xref: /netbsd-src/sys/arch/hppa/spmath/dfsub.c (revision b179d0ed445a7ffce6cb8fcf02d64f401be5616c)
1 /*	$NetBSD: dfsub.c,v 1.7 2023/07/31 20:48:04 andvar Exp $	*/
2 
3 /*	$OpenBSD: dfsub.c,v 1.4 2001/03/29 03:58:17 mickey Exp $	*/
4 
5 /*
6  * Copyright 1996 1995 by Open Software Foundation, Inc.
7  *              All Rights Reserved
8  *
9  * Permission to use, copy, modify, and distribute this software and
10  * its documentation for any purpose and without fee is hereby granted,
11  * provided that the above copyright notice appears in all copies and
12  * that both the copyright notice and this permission notice appear in
13  * supporting documentation.
14  *
15  * OSF DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE
16  * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
17  * FOR A PARTICULAR PURPOSE.
18  *
19  * IN NO EVENT SHALL OSF BE LIABLE FOR ANY SPECIAL, INDIRECT, OR
20  * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
21  * LOSS OF USE, DATA OR PROFITS, WHETHER IN ACTION OF CONTRACT,
22  * NEGLIGENCE, OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION
23  * WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
24  *
25  */
26 /*
27  * pmk1.1
28  */
29 /*
30  * (c) Copyright 1986 HEWLETT-PACKARD COMPANY
31  *
32  * To anyone who acknowledges that this file is provided "AS IS"
33  * without any express or implied warranty:
34  *     permission to use, copy, modify, and distribute this file
35  * for any purpose is hereby granted without fee, provided that
36  * the above copyright notice and this notice appears in all
37  * copies, and that the name of Hewlett-Packard Company not be
38  * used in advertising or publicity pertaining to distribution
39  * of the software without specific, written prior permission.
40  * Hewlett-Packard Company makes no representations about the
41  * suitability of this software for any purpose.
42  */
43 
44 #include <sys/cdefs.h>
45 __KERNEL_RCSID(0, "$NetBSD: dfsub.c,v 1.7 2023/07/31 20:48:04 andvar Exp $");
46 
47 #include "../spmath/float.h"
48 #include "../spmath/dbl_float.h"
49 
50 /*
51  * Double_subtract: subtract two double precision values.
52  */
53 int
dbl_fsub(dbl_floating_point * leftptr,dbl_floating_point * rightptr,dbl_floating_point * dstptr,unsigned int * status)54 dbl_fsub(dbl_floating_point *leftptr, dbl_floating_point *rightptr,
55     dbl_floating_point *dstptr, unsigned int *status)
56 {
57     register unsigned int signless_upper_left, signless_upper_right, save;
58     register unsigned int leftp1, leftp2, rightp1, rightp2, extent;
59     register unsigned int resultp1 = 0, resultp2 = 0;
60 
61     register int result_exponent, right_exponent, diff_exponent;
62     register int sign_save, jumpsize;
63     register int inexact = false, underflowtrap;
64 
65     /* Create local copies of the numbers */
66     Dbl_copyfromptr(leftptr,leftp1,leftp2);
67     Dbl_copyfromptr(rightptr,rightp1,rightp2);
68 
69     /* A zero "save" helps discover equal operands (for later),  *
70      * and is used in swapping operands (if needed).	     */
71     Dbl_xortointp1(leftp1,rightp1,/*to*/save);
72 
73     /*
74      * check first operand for NaN's or infinity
75      */
76     if ((result_exponent = Dbl_exponent(leftp1)) == DBL_INFINITY_EXPONENT)
77 	{
78 	if (Dbl_iszero_mantissa(leftp1,leftp2))
79 	    {
80 	    if (Dbl_isnotnan(rightp1,rightp2))
81 		{
82 		if (Dbl_isinfinity(rightp1,rightp2) && save==0)
83 		    {
84 		    /*
85 		     * invalid since operands are same signed infinity's
86 		     */
87 		    if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
88 		    Set_invalidflag();
89 		    Dbl_makequietnan(resultp1,resultp2);
90 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
91 		    return(NOEXCEPTION);
92 		    }
93 		/*
94 		 * return infinity
95 		 */
96 		Dbl_copytoptr(leftp1,leftp2,dstptr);
97 		return(NOEXCEPTION);
98 		}
99 	    }
100 	else
101 	    {
102 	    /*
103 	     * is NaN; signaling or quiet?
104 	     */
105 	    if (Dbl_isone_signaling(leftp1))
106 		{
107 		/* trap if INVALIDTRAP enabled */
108 		if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
109 		/* make NaN quiet */
110 		Set_invalidflag();
111 		Dbl_set_quiet(leftp1);
112 		}
113 	    /*
114 	     * is second operand a signaling NaN?
115 	     */
116 	    else if (Dbl_is_signalingnan(rightp1))
117 		{
118 		/* trap if INVALIDTRAP enabled */
119 		if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
120 		/* make NaN quiet */
121 		Set_invalidflag();
122 		Dbl_set_quiet(rightp1);
123 		Dbl_copytoptr(rightp1,rightp2,dstptr);
124 		return(NOEXCEPTION);
125 		}
126 	    /*
127 	     * return quiet NaN
128 	     */
129 	    Dbl_copytoptr(leftp1,leftp2,dstptr);
130 	    return(NOEXCEPTION);
131 	    }
132 	} /* End left NaN or Infinity processing */
133     /*
134      * check second operand for NaN's or infinity
135      */
136     if (Dbl_isinfinity_exponent(rightp1))
137 	{
138 	if (Dbl_iszero_mantissa(rightp1,rightp2))
139 	    {
140 	    /* return infinity */
141 	    Dbl_invert_sign(rightp1);
142 	    Dbl_copytoptr(rightp1,rightp2,dstptr);
143 	    return(NOEXCEPTION);
144 	    }
145 	/*
146 	 * is NaN; signaling or quiet?
147 	 */
148 	if (Dbl_isone_signaling(rightp1))
149 	    {
150 	    /* trap if INVALIDTRAP enabled */
151 	    if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
152 	    /* make NaN quiet */
153 	    Set_invalidflag();
154 	    Dbl_set_quiet(rightp1);
155 	    }
156 	/*
157 	 * return quiet NaN
158 	 */
159 	Dbl_copytoptr(rightp1,rightp2,dstptr);
160 	return(NOEXCEPTION);
161 	} /* End right NaN or Infinity processing */
162 
163     /* Invariant: Must be dealing with finite numbers */
164 
165     /* Compare operands by removing the sign */
166     Dbl_copytoint_exponentmantissap1(leftp1,signless_upper_left);
167     Dbl_copytoint_exponentmantissap1(rightp1,signless_upper_right);
168 
169     /* sign difference selects add or sub operation. */
170     if(Dbl_ismagnitudeless(leftp2,rightp2,signless_upper_left,signless_upper_right))
171 	{
172 	/* Set the left operand to the larger one by XOR swap *
173 	 *  First finish the first word using "save"	  */
174 	Dbl_xorfromintp1(save,rightp1,/*to*/rightp1);
175 	Dbl_xorfromintp1(save,leftp1,/*to*/leftp1);
176 	Dbl_swap_lower(leftp2,rightp2);
177 	result_exponent = Dbl_exponent(leftp1);
178 	Dbl_invert_sign(leftp1);
179 	}
180     /* Invariant:  left is not smaller than right. */
181 
182     if((right_exponent = Dbl_exponent(rightp1)) == 0)
183 	{
184 	/* Denormalized operands.  First look for zeroes */
185 	if(Dbl_iszero_mantissa(rightp1,rightp2))
186 	    {
187 	    /* right is zero */
188 	    if(Dbl_iszero_exponentmantissa(leftp1,leftp2))
189 		{
190 		/* Both operands are zeros */
191 		Dbl_invert_sign(rightp1);
192 		if(Is_rounding_mode(ROUNDMINUS))
193 		    {
194 		    Dbl_or_signs(leftp1,/*with*/rightp1);
195 		    }
196 		else
197 		    {
198 		    Dbl_and_signs(leftp1,/*with*/rightp1);
199 		    }
200 		}
201 	    else
202 		{
203 		/* Left is not a zero and must be the result.  Trapped
204 		 * underflows are signaled if left is denormalized.  Result
205 		 * is always exact. */
206 		if( (result_exponent == 0) && Is_underflowtrap_enabled() )
207 		    {
208 		    /* need to normalize results mantissa */
209 		    sign_save = Dbl_signextendedsign(leftp1);
210 		    Dbl_leftshiftby1(leftp1,leftp2);
211 		    Dbl_normalize(leftp1,leftp2,result_exponent);
212 		    Dbl_set_sign(leftp1,/*using*/sign_save);
213 		    Dbl_setwrapped_exponent(leftp1,result_exponent,unfl);
214 		    Dbl_copytoptr(leftp1,leftp2,dstptr);
215 		    /* inexact = false */
216 		    return(UNDERFLOWEXCEPTION);
217 		    }
218 		}
219 	    Dbl_copytoptr(leftp1,leftp2,dstptr);
220 	    return(NOEXCEPTION);
221 	    }
222 
223 	/* Neither are zeroes */
224 	Dbl_clear_sign(rightp1);	/* Exponent is already cleared */
225 	if(result_exponent == 0 )
226 	    {
227 	    /* Both operands are denormalized.  The result must be exact
228 	     * and is simply calculated.  A sum could become normalized and a
229 	     * difference could cancel to a true zero. */
230 	    if( (/*signed*/int) save >= 0 )
231 		{
232 		Dbl_subtract(leftp1,leftp2,/*minus*/rightp1,rightp2,
233 		 /*into*/resultp1,resultp2);
234 		if(Dbl_iszero_mantissa(resultp1,resultp2))
235 		    {
236 		    if(Is_rounding_mode(ROUNDMINUS))
237 			{
238 			Dbl_setone_sign(resultp1);
239 			}
240 		    else
241 			{
242 			Dbl_setzero_sign(resultp1);
243 			}
244 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
245 		    return(NOEXCEPTION);
246 		    }
247 		}
248 	    else
249 		{
250 		Dbl_addition(leftp1,leftp2,rightp1,rightp2,
251 		 /*into*/resultp1,resultp2);
252 		if(Dbl_isone_hidden(resultp1))
253 		    {
254 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
255 		    return(NOEXCEPTION);
256 		    }
257 		}
258 	    if(Is_underflowtrap_enabled())
259 		{
260 		/* need to normalize result */
261 		sign_save = Dbl_signextendedsign(resultp1);
262 		Dbl_leftshiftby1(resultp1,resultp2);
263 		Dbl_normalize(resultp1,resultp2,result_exponent);
264 		Dbl_set_sign(resultp1,/*using*/sign_save);
265 		Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
266 		Dbl_copytoptr(resultp1,resultp2,dstptr);
267 		/* inexact = false */
268 		return(UNDERFLOWEXCEPTION);
269 		}
270 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
271 	    return(NOEXCEPTION);
272 	    }
273 	right_exponent = 1;	/* Set exponent to reflect different bias
274 				 * with denormalized numbers. */
275 	}
276     else
277 	{
278 	Dbl_clear_signexponent_set_hidden(rightp1);
279 	}
280     Dbl_clear_exponent_set_hidden(leftp1);
281     diff_exponent = result_exponent - right_exponent;
282 
283     /*
284      * Special case alignment of operands that would force alignment
285      * beyond the extent of the extension.  A further optimization
286      * could special case this but only reduces the path length for this
287      * infrequent case.
288      */
289     if(diff_exponent > DBL_THRESHOLD)
290 	{
291 	diff_exponent = DBL_THRESHOLD;
292 	}
293 
294     /* Align right operand by shifting to right */
295     Dbl_right_align(/*operand*/rightp1,rightp2,/*shifted by*/diff_exponent,
296      /*and lower to*/extent);
297 
298     /* Treat sum and difference of the operands separately. */
299     if( (/*signed*/int) save >= 0 )
300 	{
301 	/*
302 	 * Difference of the two operands.  Their can be no overflow.  A
303 	 * borrow can occur out of the hidden bit and force a post
304 	 * normalization phase.
305 	 */
306 	Dbl_subtract_withextension(leftp1,leftp2,/*minus*/rightp1,rightp2,
307 	 /*with*/extent,/*into*/resultp1,resultp2);
308 	if(Dbl_iszero_hidden(resultp1))
309 	    {
310 	    /* Handle normalization */
311 	    /* A straight forward algorithm would now shift the result
312 	     * and extension left until the hidden bit becomes one.  Not
313 	     * all of the extension bits need participate in the shift.
314 	     * Only the two most significant bits (round and guard) are
315 	     * needed.  If only a single shift is needed then the guard
316 	     * bit becomes a significant low order bit and the extension
317 	     * must participate in the rounding.  If more than a single
318 	     * shift is needed, then all bits to the right of the guard
319 	     * bit are zeros, and the guard bit may or may not be zero. */
320 	    sign_save = Dbl_signextendedsign(resultp1);
321 	    Dbl_leftshiftby1_withextent(resultp1,resultp2,extent,resultp1,resultp2);
322 
323 	    /* Need to check for a zero result.  The sign and exponent
324 	     * fields have already been zeroed.  The more efficient test
325 	     * of the full object can be used.
326 	     */
327 	    if(Dbl_iszero(resultp1,resultp2))
328 		/* Must have been "x-x" or "x+(-x)". */
329 		{
330 		if(Is_rounding_mode(ROUNDMINUS)) Dbl_setone_sign(resultp1);
331 		Dbl_copytoptr(resultp1,resultp2,dstptr);
332 		return(NOEXCEPTION);
333 		}
334 	    result_exponent--;
335 	    /* Look to see if normalization is finished. */
336 	    if(Dbl_isone_hidden(resultp1)) {
337 		if(result_exponent==0) {
338 		    /* Denormalized, exponent should be zero.  Left operand *
339 		     * was normalized, so extent (guard, round) was zero    */
340 		    goto underflow;
341 		} else {
342 			/* No further normalization is needed. */
343 			Dbl_set_sign(resultp1,/*using*/sign_save);
344 			Ext_leftshiftby1(extent);
345 			goto round;
346 		}
347 	    }
348 
349 	    /* Check for denormalized, exponent should be zero.  Left    *
350 	     * operand was normalized, so extent (guard, round) was zero */
351 	    if(!(underflowtrap = Is_underflowtrap_enabled()) &&
352 	       result_exponent==0) goto underflow;
353 
354 	    /* Shift extension to complete one bit of normalization and
355 	     * update exponent. */
356 	    Ext_leftshiftby1(extent);
357 
358 	    /* Discover first one bit to determine shift amount.  Use a
359 	     * modified binary search.  We have already shifted the result
360 	     * one position right and still not found a one so the remainder
361 	     * of the extension must be zero and simplifies rounding. */
362 	    /* Scan bytes */
363 	    while(Dbl_iszero_hiddenhigh7mantissa(resultp1))
364 		{
365 		Dbl_leftshiftby8(resultp1,resultp2);
366 		if((result_exponent -= 8) <= 0  && !underflowtrap)
367 		    goto underflow;
368 		}
369 	    /* Now narrow it down to the nibble */
370 	    if(Dbl_iszero_hiddenhigh3mantissa(resultp1))
371 		{
372 		/* The lower nibble contains the normalizing one */
373 		Dbl_leftshiftby4(resultp1,resultp2);
374 		if((result_exponent -= 4) <= 0 && !underflowtrap)
375 		    goto underflow;
376 		}
377 	    /* Select case were first bit is set (already normalized)
378 	     * otherwise select the proper shift. */
379 	    if((jumpsize = Dbl_hiddenhigh3mantissa(resultp1)) > 7)
380 		{
381 		/* Already normalized */
382 		if(result_exponent <= 0) goto underflow;
383 		Dbl_set_sign(resultp1,/*using*/sign_save);
384 		Dbl_set_exponent(resultp1,/*using*/result_exponent);
385 		Dbl_copytoptr(resultp1,resultp2,dstptr);
386 		return(NOEXCEPTION);
387 		}
388 	    Dbl_sethigh4bits(resultp1,/*using*/sign_save);
389 	    switch(jumpsize)
390 		{
391 		case 1:
392 		    {
393 		    Dbl_leftshiftby3(resultp1,resultp2);
394 		    result_exponent -= 3;
395 		    break;
396 		    }
397 		case 2:
398 		case 3:
399 		    {
400 		    Dbl_leftshiftby2(resultp1,resultp2);
401 		    result_exponent -= 2;
402 		    break;
403 		    }
404 		case 4:
405 		case 5:
406 		case 6:
407 		case 7:
408 		    {
409 		    Dbl_leftshiftby1(resultp1,resultp2);
410 		    result_exponent -= 1;
411 		    break;
412 		    }
413 		}
414 	    if(result_exponent > 0)
415 		{
416 		Dbl_set_exponent(resultp1,/*using*/result_exponent);
417 		Dbl_copytoptr(resultp1,resultp2,dstptr);
418 		return(NOEXCEPTION);		/* Sign bit is already set */
419 		}
420 	    /* Fixup potential underflows */
421 	  underflow:
422 	    if(Is_underflowtrap_enabled())
423 		{
424 		Dbl_set_sign(resultp1,sign_save);
425 		Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
426 		Dbl_copytoptr(resultp1,resultp2,dstptr);
427 		/* inexact = false */
428 		return(UNDERFLOWEXCEPTION);
429 		}
430 	    /*
431 	     * Since we cannot get an inexact denormalized result,
432 	     * we can now return.
433 	     */
434 	    Dbl_fix_overshift(resultp1,resultp2,(1-result_exponent),extent);
435 	    Dbl_clear_signexponent(resultp1);
436 	    Dbl_set_sign(resultp1,sign_save);
437 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
438 	    return(NOEXCEPTION);
439 	    } /* end if(hidden...)... */
440 	/* Fall through and round */
441 	} /* end if(save >= 0)... */
442     else
443 	{
444 	/* Subtract magnitudes */
445 	Dbl_addition(leftp1,leftp2,rightp1,rightp2,/*to*/resultp1,resultp2);
446 	if(Dbl_isone_hiddenoverflow(resultp1))
447 	    {
448 	    /* Prenormalization required. */
449 	    Dbl_rightshiftby1_withextent(resultp2,extent,extent);
450 	    Dbl_arithrightshiftby1(resultp1,resultp2);
451 	    result_exponent++;
452 	    } /* end if hiddenoverflow... */
453 	} /* end else ...subtract magnitudes... */
454 
455     /* Round the result.  If the extension is all zeros,then the result is
456      * exact.  Otherwise round in the correct direction.  No underflow is
457      * possible. If a postnormalization is necessary, then the mantissa is
458      * all zeros so no shift is needed. */
459   round:
460     if(Ext_isnotzero(extent))
461 	{
462 	inexact = true;
463 	switch(Rounding_mode())
464 	    {
465 	    case ROUNDNEAREST: /* The default. */
466 	    if(Ext_isone_sign(extent))
467 		{
468 		/* at least 1/2 ulp */
469 		if(Ext_isnotzero_lower(extent)  ||
470 		  Dbl_isone_lowmantissap2(resultp2))
471 		    {
472 		    /* either exactly half way and odd or more than 1/2ulp */
473 		    Dbl_increment(resultp1,resultp2);
474 		    }
475 		}
476 	    break;
477 
478 	    case ROUNDPLUS:
479 	    if(Dbl_iszero_sign(resultp1))
480 		{
481 		/* Round up positive results */
482 		Dbl_increment(resultp1,resultp2);
483 		}
484 	    break;
485 
486 	    case ROUNDMINUS:
487 	    if(Dbl_isone_sign(resultp1))
488 		{
489 		/* Round down negative results */
490 		Dbl_increment(resultp1,resultp2);
491 		}
492 
493 	    case ROUNDZERO:;
494 	    /* truncate is simple */
495 	    } /* end switch... */
496 	if(Dbl_isone_hiddenoverflow(resultp1)) result_exponent++;
497 	}
498     if(result_exponent == DBL_INFINITY_EXPONENT)
499 	{
500 	/* Overflow */
501 	if(Is_overflowtrap_enabled())
502 	    {
503 	    Dbl_setwrapped_exponent(resultp1,result_exponent,ovfl);
504 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
505 	    if (inexact) {
506 		if (Is_inexacttrap_enabled())
507 		    return(OVERFLOWEXCEPTION | INEXACTEXCEPTION);
508 		else
509 		    Set_inexactflag();
510 	    }
511 	    return(OVERFLOWEXCEPTION);
512 	    }
513 	else
514 	    {
515 	    inexact = true;
516 	    Set_overflowflag();
517 	    Dbl_setoverflow(resultp1,resultp2);
518 	    }
519 	}
520     else Dbl_set_exponent(resultp1,result_exponent);
521     Dbl_copytoptr(resultp1,resultp2,dstptr);
522     if(inexact) {
523 	if(Is_inexacttrap_enabled())
524 		return(INEXACTEXCEPTION);
525 	else
526 		Set_inexactflag();
527     }
528     return(NOEXCEPTION);
529     }
530