1/*
2 * Linux/PA-RISC Project (http://www.parisc-linux.org/)
3 *
4 * Floating-point emulation code
5 *  Copyright (C) 2001 Hewlett-Packard (Paul Bame) <bame@debian.org>
6 *
7 *    This program is free software; you can redistribute it and/or modify
8 *    it under the terms of the GNU General Public License as published by
9 *    the Free Software Foundation; either version 2, or (at your option)
10 *    any later version.
11 *
12 *    This program is distributed in the hope that it will be useful,
13 *    but WITHOUT ANY WARRANTY; without even the implied warranty of
14 *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15 *    GNU General Public License for more details.
16 *
17 *    You should have received a copy of the GNU General Public License
18 *    along with this program; if not, write to the Free Software
19 *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
20 */
21/*
22 * BEGIN_DESC
23 *
24 *  File:
25 *	@(#)	pa/spmath/dfsub.c		$Revision: 1.1.1.1 $
26 *
27 *  Purpose:
28 *	Double_subtract: subtract two double precision values.
29 *
30 *  External Interfaces:
31 *	dbl_fsub(leftptr, rightptr, dstptr, status)
32 *
33 *  Internal Interfaces:
34 *
35 *  Theory:
36 *	<<please update with a overview of the operation of this file>>
37 *
38 * END_DESC
39*/
40
41
42#include "float.h"
43#include "dbl_float.h"
44
45/*
46 * Double_subtract: subtract two double precision values.
47 */
48int
49dbl_fsub(
50	    dbl_floating_point *leftptr,
51	    dbl_floating_point *rightptr,
52	    dbl_floating_point *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 boolean 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		{
336		if(result_exponent==0)
337		    {
338		    /* Denormalized, exponent should be zero.  Left operand *
339		     * was normalized, so extent (guard, round) was zero    */
340		    goto underflow;
341		    }
342		else
343		    {
344		    /* No further normalization is needed. */
345		    Dbl_set_sign(resultp1,/*using*/sign_save);
346	    	    Ext_leftshiftby1(extent);
347		    goto round;
348		    }
349		}
350
351	    /* Check for denormalized, exponent should be zero.  Left    *
352	     * operand was normalized, so extent (guard, round) was zero */
353	    if(!(underflowtrap = Is_underflowtrap_enabled()) &&
354	       result_exponent==0) goto underflow;
355
356	    /* Shift extension to complete one bit of normalization and
357	     * update exponent. */
358	    Ext_leftshiftby1(extent);
359
360	    /* Discover first one bit to determine shift amount.  Use a
361	     * modified binary search.  We have already shifted the result
362	     * one position right and still not found a one so the remainder
363	     * of the extension must be zero and simplifies rounding. */
364	    /* Scan bytes */
365	    while(Dbl_iszero_hiddenhigh7mantissa(resultp1))
366		{
367		Dbl_leftshiftby8(resultp1,resultp2);
368		if((result_exponent -= 8) <= 0  && !underflowtrap)
369		    goto underflow;
370		}
371	    /* Now narrow it down to the nibble */
372	    if(Dbl_iszero_hiddenhigh3mantissa(resultp1))
373		{
374		/* The lower nibble contains the normalizing one */
375		Dbl_leftshiftby4(resultp1,resultp2);
376		if((result_exponent -= 4) <= 0 && !underflowtrap)
377		    goto underflow;
378		}
379	    /* Select case were first bit is set (already normalized)
380	     * otherwise select the proper shift. */
381	    if((jumpsize = Dbl_hiddenhigh3mantissa(resultp1)) > 7)
382		{
383		/* Already normalized */
384		if(result_exponent <= 0) goto underflow;
385		Dbl_set_sign(resultp1,/*using*/sign_save);
386		Dbl_set_exponent(resultp1,/*using*/result_exponent);
387		Dbl_copytoptr(resultp1,resultp2,dstptr);
388		return(NOEXCEPTION);
389		}
390	    Dbl_sethigh4bits(resultp1,/*using*/sign_save);
391	    switch(jumpsize)
392		{
393		case 1:
394		    {
395		    Dbl_leftshiftby3(resultp1,resultp2);
396		    result_exponent -= 3;
397		    break;
398		    }
399		case 2:
400		case 3:
401		    {
402		    Dbl_leftshiftby2(resultp1,resultp2);
403		    result_exponent -= 2;
404		    break;
405		    }
406		case 4:
407		case 5:
408		case 6:
409		case 7:
410		    {
411		    Dbl_leftshiftby1(resultp1,resultp2);
412		    result_exponent -= 1;
413		    break;
414		    }
415		}
416	    if(result_exponent > 0)
417		{
418		Dbl_set_exponent(resultp1,/*using*/result_exponent);
419		Dbl_copytoptr(resultp1,resultp2,dstptr);
420		return(NOEXCEPTION);		/* Sign bit is already set */
421		}
422	    /* Fixup potential underflows */
423	  underflow:
424	    if(Is_underflowtrap_enabled())
425		{
426		Dbl_set_sign(resultp1,sign_save);
427                Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
428		Dbl_copytoptr(resultp1,resultp2,dstptr);
429		/* inexact = FALSE */
430		return(UNDERFLOWEXCEPTION);
431		}
432	    /*
433	     * Since we cannot get an inexact denormalized result,
434	     * we can now return.
435	     */
436	    Dbl_fix_overshift(resultp1,resultp2,(1-result_exponent),extent);
437	    Dbl_clear_signexponent(resultp1);
438	    Dbl_set_sign(resultp1,sign_save);
439	    Dbl_copytoptr(resultp1,resultp2,dstptr);
440	    return(NOEXCEPTION);
441	    } /* end if(hidden...)... */
442	/* Fall through and round */
443	} /* end if(save >= 0)... */
444    else
445	{
446	/* Subtract magnitudes */
447	Dbl_addition(leftp1,leftp2,rightp1,rightp2,/*to*/resultp1,resultp2);
448	if(Dbl_isone_hiddenoverflow(resultp1))
449	    {
450	    /* Prenormalization required. */
451	    Dbl_rightshiftby1_withextent(resultp2,extent,extent);
452	    Dbl_arithrightshiftby1(resultp1,resultp2);
453	    result_exponent++;
454	    } /* end if hiddenoverflow... */
455	} /* end else ...subtract magnitudes... */
456
457    /* Round the result.  If the extension is all zeros,then the result is
458     * exact.  Otherwise round in the correct direction.  No underflow is
459     * possible. If a postnormalization is necessary, then the mantissa is
460     * all zeros so no shift is needed. */
461  round:
462    if(Ext_isnotzero(extent))
463	{
464	inexact = TRUE;
465	switch(Rounding_mode())
466	    {
467	    case ROUNDNEAREST: /* The default. */
468	    if(Ext_isone_sign(extent))
469		{
470		/* at least 1/2 ulp */
471		if(Ext_isnotzero_lower(extent)  ||
472		  Dbl_isone_lowmantissap2(resultp2))
473		    {
474		    /* either exactly half way and odd or more than 1/2ulp */
475		    Dbl_increment(resultp1,resultp2);
476		    }
477		}
478	    break;
479
480	    case ROUNDPLUS:
481	    if(Dbl_iszero_sign(resultp1))
482		{
483		/* Round up positive results */
484		Dbl_increment(resultp1,resultp2);
485		}
486	    break;
487
488	    case ROUNDMINUS:
489	    if(Dbl_isone_sign(resultp1))
490		{
491		/* Round down negative results */
492		Dbl_increment(resultp1,resultp2);
493		}
494
495	    case ROUNDZERO:;
496	    /* truncate is simple */
497	    } /* end switch... */
498	if(Dbl_isone_hiddenoverflow(resultp1)) result_exponent++;
499	}
500    if(result_exponent == DBL_INFINITY_EXPONENT)
501        {
502        /* Overflow */
503        if(Is_overflowtrap_enabled())
504	    {
505	    Dbl_setwrapped_exponent(resultp1,result_exponent,ovfl);
506	    Dbl_copytoptr(resultp1,resultp2,dstptr);
507	    if (inexact)
508	    if (Is_inexacttrap_enabled())
509		return(OVERFLOWEXCEPTION | INEXACTEXCEPTION);
510		else Set_inexactflag();
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()) return(INEXACTEXCEPTION);
524	else Set_inexactflag();
525    return(NOEXCEPTION);
526    }
527