xref: /netbsd-src/sys/arch/hppa/spmath/dfsub.c (revision d710132b4b8ce7f7cccaaf660cb16aa16b4077a0)
1 /*	$NetBSD: dfsub.c,v 1.1 2002/06/05 01:04:24 fredette 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 "../spmath/float.h"
45 #include "../spmath/dbl_float.h"
46 
47 /*
48  * Double_subtract: subtract two double precision values.
49  */
50 int
51 dbl_fsub(leftptr, rightptr, dstptr, status)
52     dbl_floating_point *leftptr, *rightptr, *dstptr;
53     unsigned int *status;
54     {
55     register unsigned int signless_upper_left, signless_upper_right, save;
56     register unsigned int leftp1, leftp2, rightp1, rightp2, extent;
57     register unsigned int resultp1 = 0, resultp2 = 0;
58 
59     register int result_exponent, right_exponent, diff_exponent;
60     register int sign_save, jumpsize;
61     register int inexact = FALSE, underflowtrap;
62 
63     /* Create local copies of the numbers */
64     Dbl_copyfromptr(leftptr,leftp1,leftp2);
65     Dbl_copyfromptr(rightptr,rightp1,rightp2);
66 
67     /* A zero "save" helps discover equal operands (for later),  *
68      * and is used in swapping operands (if needed).	     */
69     Dbl_xortointp1(leftp1,rightp1,/*to*/save);
70 
71     /*
72      * check first operand for NaN's or infinity
73      */
74     if ((result_exponent = Dbl_exponent(leftp1)) == DBL_INFINITY_EXPONENT)
75 	{
76 	if (Dbl_iszero_mantissa(leftp1,leftp2))
77 	    {
78 	    if (Dbl_isnotnan(rightp1,rightp2))
79 		{
80 		if (Dbl_isinfinity(rightp1,rightp2) && save==0)
81 		    {
82 		    /*
83 		     * invalid since operands are same signed infinity's
84 		     */
85 		    if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
86 		    Set_invalidflag();
87 		    Dbl_makequietnan(resultp1,resultp2);
88 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
89 		    return(NOEXCEPTION);
90 		    }
91 		/*
92 		 * return infinity
93 		 */
94 		Dbl_copytoptr(leftp1,leftp2,dstptr);
95 		return(NOEXCEPTION);
96 		}
97 	    }
98 	else
99 	    {
100 	    /*
101 	     * is NaN; signaling or quiet?
102 	     */
103 	    if (Dbl_isone_signaling(leftp1))
104 		{
105 		/* trap if INVALIDTRAP enabled */
106 		if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
107 		/* make NaN quiet */
108 		Set_invalidflag();
109 		Dbl_set_quiet(leftp1);
110 		}
111 	    /*
112 	     * is second operand a signaling NaN?
113 	     */
114 	    else if (Dbl_is_signalingnan(rightp1))
115 		{
116 		/* trap if INVALIDTRAP enabled */
117 		if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
118 		/* make NaN quiet */
119 		Set_invalidflag();
120 		Dbl_set_quiet(rightp1);
121 		Dbl_copytoptr(rightp1,rightp2,dstptr);
122 		return(NOEXCEPTION);
123 		}
124 	    /*
125 	     * return quiet NaN
126 	     */
127 	    Dbl_copytoptr(leftp1,leftp2,dstptr);
128 	    return(NOEXCEPTION);
129 	    }
130 	} /* End left NaN or Infinity processing */
131     /*
132      * check second operand for NaN's or infinity
133      */
134     if (Dbl_isinfinity_exponent(rightp1))
135 	{
136 	if (Dbl_iszero_mantissa(rightp1,rightp2))
137 	    {
138 	    /* return infinity */
139 	    Dbl_invert_sign(rightp1);
140 	    Dbl_copytoptr(rightp1,rightp2,dstptr);
141 	    return(NOEXCEPTION);
142 	    }
143 	/*
144 	 * is NaN; signaling or quiet?
145 	 */
146 	if (Dbl_isone_signaling(rightp1))
147 	    {
148 	    /* trap if INVALIDTRAP enabled */
149 	    if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
150 	    /* make NaN quiet */
151 	    Set_invalidflag();
152 	    Dbl_set_quiet(rightp1);
153 	    }
154 	/*
155 	 * return quiet NaN
156 	 */
157 	Dbl_copytoptr(rightp1,rightp2,dstptr);
158 	return(NOEXCEPTION);
159 	} /* End right NaN or Infinity processing */
160 
161     /* Invariant: Must be dealing with finite numbers */
162 
163     /* Compare operands by removing the sign */
164     Dbl_copytoint_exponentmantissap1(leftp1,signless_upper_left);
165     Dbl_copytoint_exponentmantissap1(rightp1,signless_upper_right);
166 
167     /* sign difference selects add or sub operation. */
168     if(Dbl_ismagnitudeless(leftp2,rightp2,signless_upper_left,signless_upper_right))
169 	{
170 	/* Set the left operand to the larger one by XOR swap *
171 	 *  First finish the first word using "save"	  */
172 	Dbl_xorfromintp1(save,rightp1,/*to*/rightp1);
173 	Dbl_xorfromintp1(save,leftp1,/*to*/leftp1);
174 	Dbl_swap_lower(leftp2,rightp2);
175 	result_exponent = Dbl_exponent(leftp1);
176 	Dbl_invert_sign(leftp1);
177 	}
178     /* Invariant:  left is not smaller than right. */
179 
180     if((right_exponent = Dbl_exponent(rightp1)) == 0)
181 	{
182 	/* Denormalized operands.  First look for zeroes */
183 	if(Dbl_iszero_mantissa(rightp1,rightp2))
184 	    {
185 	    /* right is zero */
186 	    if(Dbl_iszero_exponentmantissa(leftp1,leftp2))
187 		{
188 		/* Both operands are zeros */
189 		Dbl_invert_sign(rightp1);
190 		if(Is_rounding_mode(ROUNDMINUS))
191 		    {
192 		    Dbl_or_signs(leftp1,/*with*/rightp1);
193 		    }
194 		else
195 		    {
196 		    Dbl_and_signs(leftp1,/*with*/rightp1);
197 		    }
198 		}
199 	    else
200 		{
201 		/* Left is not a zero and must be the result.  Trapped
202 		 * underflows are signaled if left is denormalized.  Result
203 		 * is always exact. */
204 		if( (result_exponent == 0) && Is_underflowtrap_enabled() )
205 		    {
206 		    /* need to normalize results mantissa */
207 		    sign_save = Dbl_signextendedsign(leftp1);
208 		    Dbl_leftshiftby1(leftp1,leftp2);
209 		    Dbl_normalize(leftp1,leftp2,result_exponent);
210 		    Dbl_set_sign(leftp1,/*using*/sign_save);
211 		    Dbl_setwrapped_exponent(leftp1,result_exponent,unfl);
212 		    Dbl_copytoptr(leftp1,leftp2,dstptr);
213 		    /* inexact = FALSE */
214 		    return(UNDERFLOWEXCEPTION);
215 		    }
216 		}
217 	    Dbl_copytoptr(leftp1,leftp2,dstptr);
218 	    return(NOEXCEPTION);
219 	    }
220 
221 	/* Neither are zeroes */
222 	Dbl_clear_sign(rightp1);	/* Exponent is already cleared */
223 	if(result_exponent == 0 )
224 	    {
225 	    /* Both operands are denormalized.  The result must be exact
226 	     * and is simply calculated.  A sum could become normalized and a
227 	     * difference could cancel to a true zero. */
228 	    if( (/*signed*/int) save >= 0 )
229 		{
230 		Dbl_subtract(leftp1,leftp2,/*minus*/rightp1,rightp2,
231 		 /*into*/resultp1,resultp2);
232 		if(Dbl_iszero_mantissa(resultp1,resultp2))
233 		    {
234 		    if(Is_rounding_mode(ROUNDMINUS))
235 			{
236 			Dbl_setone_sign(resultp1);
237 			}
238 		    else
239 			{
240 			Dbl_setzero_sign(resultp1);
241 			}
242 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
243 		    return(NOEXCEPTION);
244 		    }
245 		}
246 	    else
247 		{
248 		Dbl_addition(leftp1,leftp2,rightp1,rightp2,
249 		 /*into*/resultp1,resultp2);
250 		if(Dbl_isone_hidden(resultp1))
251 		    {
252 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
253 		    return(NOEXCEPTION);
254 		    }
255 		}
256 	    if(Is_underflowtrap_enabled())
257 		{
258 		/* need to normalize result */
259 		sign_save = Dbl_signextendedsign(resultp1);
260 		Dbl_leftshiftby1(resultp1,resultp2);
261 		Dbl_normalize(resultp1,resultp2,result_exponent);
262 		Dbl_set_sign(resultp1,/*using*/sign_save);
263 		Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
264 		Dbl_copytoptr(resultp1,resultp2,dstptr);
265 		/* inexact = FALSE */
266 		return(UNDERFLOWEXCEPTION);
267 		}
268 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
269 	    return(NOEXCEPTION);
270 	    }
271 	right_exponent = 1;	/* Set exponent to reflect different bias
272 				 * with denomalized numbers. */
273 	}
274     else
275 	{
276 	Dbl_clear_signexponent_set_hidden(rightp1);
277 	}
278     Dbl_clear_exponent_set_hidden(leftp1);
279     diff_exponent = result_exponent - right_exponent;
280 
281     /*
282      * Special case alignment of operands that would force alignment
283      * beyond the extent of the extension.  A further optimization
284      * could special case this but only reduces the path length for this
285      * infrequent case.
286      */
287     if(diff_exponent > DBL_THRESHOLD)
288 	{
289 	diff_exponent = DBL_THRESHOLD;
290 	}
291 
292     /* Align right operand by shifting to right */
293     Dbl_right_align(/*operand*/rightp1,rightp2,/*shifted by*/diff_exponent,
294      /*and lower to*/extent);
295 
296     /* Treat sum and difference of the operands separately. */
297     if( (/*signed*/int) save >= 0 )
298 	{
299 	/*
300 	 * Difference of the two operands.  Their can be no overflow.  A
301 	 * borrow can occur out of the hidden bit and force a post
302 	 * normalization phase.
303 	 */
304 	Dbl_subtract_withextension(leftp1,leftp2,/*minus*/rightp1,rightp2,
305 	 /*with*/extent,/*into*/resultp1,resultp2);
306 	if(Dbl_iszero_hidden(resultp1))
307 	    {
308 	    /* Handle normalization */
309 	    /* A straight foward algorithm would now shift the result
310 	     * and extension left until the hidden bit becomes one.  Not
311 	     * all of the extension bits need participate in the shift.
312 	     * Only the two most significant bits (round and guard) are
313 	     * needed.  If only a single shift is needed then the guard
314 	     * bit becomes a significant low order bit and the extension
315 	     * must participate in the rounding.  If more than a single
316 	     * shift is needed, then all bits to the right of the guard
317 	     * bit are zeros, and the guard bit may or may not be zero. */
318 	    sign_save = Dbl_signextendedsign(resultp1);
319 	    Dbl_leftshiftby1_withextent(resultp1,resultp2,extent,resultp1,resultp2);
320 
321 	    /* Need to check for a zero result.  The sign and exponent
322 	     * fields have already been zeroed.  The more efficient test
323 	     * of the full object can be used.
324 	     */
325 	    if(Dbl_iszero(resultp1,resultp2))
326 		/* Must have been "x-x" or "x+(-x)". */
327 		{
328 		if(Is_rounding_mode(ROUNDMINUS)) Dbl_setone_sign(resultp1);
329 		Dbl_copytoptr(resultp1,resultp2,dstptr);
330 		return(NOEXCEPTION);
331 		}
332 	    result_exponent--;
333 	    /* Look to see if normalization is finished. */
334 	    if(Dbl_isone_hidden(resultp1)) {
335 		if(result_exponent==0) {
336 		    /* Denormalized, exponent should be zero.  Left operand *
337 		     * was normalized, so extent (guard, round) was zero    */
338 		    goto underflow;
339 		} else {
340 			/* No further normalization is needed. */
341 			Dbl_set_sign(resultp1,/*using*/sign_save);
342 			Ext_leftshiftby1(extent);
343 			goto round;
344 		}
345 	    }
346 
347 	    /* Check for denormalized, exponent should be zero.  Left    *
348 	     * operand was normalized, so extent (guard, round) was zero */
349 	    if(!(underflowtrap = Is_underflowtrap_enabled()) &&
350 	       result_exponent==0) goto underflow;
351 
352 	    /* Shift extension to complete one bit of normalization and
353 	     * update exponent. */
354 	    Ext_leftshiftby1(extent);
355 
356 	    /* Discover first one bit to determine shift amount.  Use a
357 	     * modified binary search.  We have already shifted the result
358 	     * one position right and still not found a one so the remainder
359 	     * of the extension must be zero and simplifies rounding. */
360 	    /* Scan bytes */
361 	    while(Dbl_iszero_hiddenhigh7mantissa(resultp1))
362 		{
363 		Dbl_leftshiftby8(resultp1,resultp2);
364 		if((result_exponent -= 8) <= 0  && !underflowtrap)
365 		    goto underflow;
366 		}
367 	    /* Now narrow it down to the nibble */
368 	    if(Dbl_iszero_hiddenhigh3mantissa(resultp1))
369 		{
370 		/* The lower nibble contains the normalizing one */
371 		Dbl_leftshiftby4(resultp1,resultp2);
372 		if((result_exponent -= 4) <= 0 && !underflowtrap)
373 		    goto underflow;
374 		}
375 	    /* Select case were first bit is set (already normalized)
376 	     * otherwise select the proper shift. */
377 	    if((jumpsize = Dbl_hiddenhigh3mantissa(resultp1)) > 7)
378 		{
379 		/* Already normalized */
380 		if(result_exponent <= 0) goto underflow;
381 		Dbl_set_sign(resultp1,/*using*/sign_save);
382 		Dbl_set_exponent(resultp1,/*using*/result_exponent);
383 		Dbl_copytoptr(resultp1,resultp2,dstptr);
384 		return(NOEXCEPTION);
385 		}
386 	    Dbl_sethigh4bits(resultp1,/*using*/sign_save);
387 	    switch(jumpsize)
388 		{
389 		case 1:
390 		    {
391 		    Dbl_leftshiftby3(resultp1,resultp2);
392 		    result_exponent -= 3;
393 		    break;
394 		    }
395 		case 2:
396 		case 3:
397 		    {
398 		    Dbl_leftshiftby2(resultp1,resultp2);
399 		    result_exponent -= 2;
400 		    break;
401 		    }
402 		case 4:
403 		case 5:
404 		case 6:
405 		case 7:
406 		    {
407 		    Dbl_leftshiftby1(resultp1,resultp2);
408 		    result_exponent -= 1;
409 		    break;
410 		    }
411 		}
412 	    if(result_exponent > 0)
413 		{
414 		Dbl_set_exponent(resultp1,/*using*/result_exponent);
415 		Dbl_copytoptr(resultp1,resultp2,dstptr);
416 		return(NOEXCEPTION);		/* Sign bit is already set */
417 		}
418 	    /* Fixup potential underflows */
419 	  underflow:
420 	    if(Is_underflowtrap_enabled())
421 		{
422 		Dbl_set_sign(resultp1,sign_save);
423 		Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
424 		Dbl_copytoptr(resultp1,resultp2,dstptr);
425 		/* inexact = FALSE */
426 		return(UNDERFLOWEXCEPTION);
427 		}
428 	    /*
429 	     * Since we cannot get an inexact denormalized result,
430 	     * we can now return.
431 	     */
432 	    Dbl_fix_overshift(resultp1,resultp2,(1-result_exponent),extent);
433 	    Dbl_clear_signexponent(resultp1);
434 	    Dbl_set_sign(resultp1,sign_save);
435 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
436 	    return(NOEXCEPTION);
437 	    } /* end if(hidden...)... */
438 	/* Fall through and round */
439 	} /* end if(save >= 0)... */
440     else
441 	{
442 	/* Subtract magnitudes */
443 	Dbl_addition(leftp1,leftp2,rightp1,rightp2,/*to*/resultp1,resultp2);
444 	if(Dbl_isone_hiddenoverflow(resultp1))
445 	    {
446 	    /* Prenormalization required. */
447 	    Dbl_rightshiftby1_withextent(resultp2,extent,extent);
448 	    Dbl_arithrightshiftby1(resultp1,resultp2);
449 	    result_exponent++;
450 	    } /* end if hiddenoverflow... */
451 	} /* end else ...subtract magnitudes... */
452 
453     /* Round the result.  If the extension is all zeros,then the result is
454      * exact.  Otherwise round in the correct direction.  No underflow is
455      * possible. If a postnormalization is necessary, then the mantissa is
456      * all zeros so no shift is needed. */
457   round:
458     if(Ext_isnotzero(extent))
459 	{
460 	inexact = TRUE;
461 	switch(Rounding_mode())
462 	    {
463 	    case ROUNDNEAREST: /* The default. */
464 	    if(Ext_isone_sign(extent))
465 		{
466 		/* at least 1/2 ulp */
467 		if(Ext_isnotzero_lower(extent)  ||
468 		  Dbl_isone_lowmantissap2(resultp2))
469 		    {
470 		    /* either exactly half way and odd or more than 1/2ulp */
471 		    Dbl_increment(resultp1,resultp2);
472 		    }
473 		}
474 	    break;
475 
476 	    case ROUNDPLUS:
477 	    if(Dbl_iszero_sign(resultp1))
478 		{
479 		/* Round up positive results */
480 		Dbl_increment(resultp1,resultp2);
481 		}
482 	    break;
483 
484 	    case ROUNDMINUS:
485 	    if(Dbl_isone_sign(resultp1))
486 		{
487 		/* Round down negative results */
488 		Dbl_increment(resultp1,resultp2);
489 		}
490 
491 	    case ROUNDZERO:;
492 	    /* truncate is simple */
493 	    } /* end switch... */
494 	if(Dbl_isone_hiddenoverflow(resultp1)) result_exponent++;
495 	}
496     if(result_exponent == DBL_INFINITY_EXPONENT)
497 	{
498 	/* Overflow */
499 	if(Is_overflowtrap_enabled())
500 	    {
501 	    Dbl_setwrapped_exponent(resultp1,result_exponent,ovfl);
502 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
503 	    if (inexact) {
504 		if (Is_inexacttrap_enabled())
505 		    return(OVERFLOWEXCEPTION | INEXACTEXCEPTION);
506 		else
507 		    Set_inexactflag();
508 	    }
509 	    return(OVERFLOWEXCEPTION);
510 	    }
511 	else
512 	    {
513 	    inexact = TRUE;
514 	    Set_overflowflag();
515 	    Dbl_setoverflow(resultp1,resultp2);
516 	    }
517 	}
518     else Dbl_set_exponent(resultp1,result_exponent);
519     Dbl_copytoptr(resultp1,resultp2,dstptr);
520     if(inexact) {
521 	if(Is_inexacttrap_enabled())
522 		return(INEXACTEXCEPTION);
523 	else
524 		Set_inexactflag();
525     }
526     return(NOEXCEPTION);
527     }
528