xref: /linux/arch/parisc/math-emu/dfsub.c (revision e5a52fd2b8cdb700b3c07b030e050a49ef3156b9)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Linux/PA-RISC Project (http://www.parisc-linux.org/)
4  *
5  * Floating-point emulation code
6  *  Copyright (C) 2001 Hewlett-Packard (Paul Bame) <bame@debian.org>
7  */
8 /*
9  * BEGIN_DESC
10  *
11  *  File:
12  *	@(#)	pa/spmath/dfsub.c		$Revision: 1.1 $
13  *
14  *  Purpose:
15  *	Double_subtract: subtract two double precision values.
16  *
17  *  External Interfaces:
18  *	dbl_fsub(leftptr, rightptr, dstptr, status)
19  *
20  *  Internal Interfaces:
21  *
22  *  Theory:
23  *	<<please update with a overview of the operation of this file>>
24  *
25  * END_DESC
26 */
27 
28 
29 #include "float.h"
30 #include "dbl_float.h"
31 
32 /*
33  * Double_subtract: subtract two double precision values.
34  */
35 int
36 dbl_fsub(
37 	    dbl_floating_point *leftptr,
38 	    dbl_floating_point *rightptr,
39 	    dbl_floating_point *dstptr,
40 	    unsigned int *status)
41     {
42     register unsigned int signless_upper_left, signless_upper_right, save;
43     register unsigned int leftp1, leftp2, rightp1, rightp2, extent;
44     register unsigned int resultp1 = 0, resultp2 = 0;
45 
46     register int result_exponent, right_exponent, diff_exponent;
47     register int sign_save, jumpsize;
48     register boolean inexact = FALSE, underflowtrap;
49 
50     /* Create local copies of the numbers */
51     Dbl_copyfromptr(leftptr,leftp1,leftp2);
52     Dbl_copyfromptr(rightptr,rightp1,rightp2);
53 
54     /* A zero "save" helps discover equal operands (for later),  *
55      * and is used in swapping operands (if needed).             */
56     Dbl_xortointp1(leftp1,rightp1,/*to*/save);
57 
58     /*
59      * check first operand for NaN's or infinity
60      */
61     if ((result_exponent = Dbl_exponent(leftp1)) == DBL_INFINITY_EXPONENT)
62 	{
63 	if (Dbl_iszero_mantissa(leftp1,leftp2))
64 	    {
65 	    if (Dbl_isnotnan(rightp1,rightp2))
66 		{
67 		if (Dbl_isinfinity(rightp1,rightp2) && save==0)
68 		    {
69 		    /*
70 		     * invalid since operands are same signed infinity's
71 		     */
72 		    if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
73                     Set_invalidflag();
74                     Dbl_makequietnan(resultp1,resultp2);
75 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
76 		    return(NOEXCEPTION);
77 		    }
78 		/*
79 	 	 * return infinity
80 	 	 */
81 		Dbl_copytoptr(leftp1,leftp2,dstptr);
82 		return(NOEXCEPTION);
83 		}
84 	    }
85 	else
86 	    {
87             /*
88              * is NaN; signaling or quiet?
89              */
90             if (Dbl_isone_signaling(leftp1))
91 		{
92                	/* trap if INVALIDTRAP enabled */
93 		if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
94         	/* make NaN quiet */
95         	Set_invalidflag();
96         	Dbl_set_quiet(leftp1);
97         	}
98 	    /*
99 	     * is second operand a signaling NaN?
100 	     */
101 	    else if (Dbl_is_signalingnan(rightp1))
102 		{
103         	/* trap if INVALIDTRAP enabled */
104                	if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
105 		/* make NaN quiet */
106 		Set_invalidflag();
107 		Dbl_set_quiet(rightp1);
108 		Dbl_copytoptr(rightp1,rightp2,dstptr);
109 		return(NOEXCEPTION);
110 		}
111 	    /*
112  	     * return quiet NaN
113  	     */
114 	    Dbl_copytoptr(leftp1,leftp2,dstptr);
115  	    return(NOEXCEPTION);
116 	    }
117 	} /* End left NaN or Infinity processing */
118     /*
119      * check second operand for NaN's or infinity
120      */
121     if (Dbl_isinfinity_exponent(rightp1))
122 	{
123 	if (Dbl_iszero_mantissa(rightp1,rightp2))
124 	    {
125 	    /* return infinity */
126 	    Dbl_invert_sign(rightp1);
127 	    Dbl_copytoptr(rightp1,rightp2,dstptr);
128 	    return(NOEXCEPTION);
129 	    }
130         /*
131          * is NaN; signaling or quiet?
132          */
133         if (Dbl_isone_signaling(rightp1))
134 	    {
135             /* trap if INVALIDTRAP enabled */
136 	    if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION);
137 	    /* make NaN quiet */
138 	    Set_invalidflag();
139 	    Dbl_set_quiet(rightp1);
140 	    }
141 	/*
142 	 * return quiet NaN
143  	 */
144 	Dbl_copytoptr(rightp1,rightp2,dstptr);
145 	return(NOEXCEPTION);
146     	} /* End right NaN or Infinity processing */
147 
148     /* Invariant: Must be dealing with finite numbers */
149 
150     /* Compare operands by removing the sign */
151     Dbl_copytoint_exponentmantissap1(leftp1,signless_upper_left);
152     Dbl_copytoint_exponentmantissap1(rightp1,signless_upper_right);
153 
154     /* sign difference selects add or sub operation. */
155     if(Dbl_ismagnitudeless(leftp2,rightp2,signless_upper_left,signless_upper_right))
156 	{
157 	/* Set the left operand to the larger one by XOR swap *
158 	 *  First finish the first word using "save"          */
159 	Dbl_xorfromintp1(save,rightp1,/*to*/rightp1);
160 	Dbl_xorfromintp1(save,leftp1,/*to*/leftp1);
161      	Dbl_swap_lower(leftp2,rightp2);
162 	result_exponent = Dbl_exponent(leftp1);
163 	Dbl_invert_sign(leftp1);
164 	}
165     /* Invariant:  left is not smaller than right. */
166 
167     if((right_exponent = Dbl_exponent(rightp1)) == 0)
168         {
169 	/* Denormalized operands.  First look for zeroes */
170 	if(Dbl_iszero_mantissa(rightp1,rightp2))
171 	    {
172 	    /* right is zero */
173 	    if(Dbl_iszero_exponentmantissa(leftp1,leftp2))
174 		{
175 		/* Both operands are zeros */
176 		Dbl_invert_sign(rightp1);
177 		if(Is_rounding_mode(ROUNDMINUS))
178 		    {
179 		    Dbl_or_signs(leftp1,/*with*/rightp1);
180 		    }
181 		else
182 		    {
183 		    Dbl_and_signs(leftp1,/*with*/rightp1);
184 		    }
185 		}
186 	    else
187 		{
188 		/* Left is not a zero and must be the result.  Trapped
189 		 * underflows are signaled if left is denormalized.  Result
190 		 * is always exact. */
191 		if( (result_exponent == 0) && Is_underflowtrap_enabled() )
192 		    {
193 		    /* need to normalize results mantissa */
194 	    	    sign_save = Dbl_signextendedsign(leftp1);
195 		    Dbl_leftshiftby1(leftp1,leftp2);
196 		    Dbl_normalize(leftp1,leftp2,result_exponent);
197 		    Dbl_set_sign(leftp1,/*using*/sign_save);
198                     Dbl_setwrapped_exponent(leftp1,result_exponent,unfl);
199 		    Dbl_copytoptr(leftp1,leftp2,dstptr);
200 		    /* inexact = FALSE */
201 		    return(UNDERFLOWEXCEPTION);
202 		    }
203 		}
204 	    Dbl_copytoptr(leftp1,leftp2,dstptr);
205 	    return(NOEXCEPTION);
206 	    }
207 
208 	/* Neither are zeroes */
209 	Dbl_clear_sign(rightp1);	/* Exponent is already cleared */
210 	if(result_exponent == 0 )
211 	    {
212 	    /* Both operands are denormalized.  The result must be exact
213 	     * and is simply calculated.  A sum could become normalized and a
214 	     * difference could cancel to a true zero. */
215 	    if( (/*signed*/int) save >= 0 )
216 		{
217 		Dbl_subtract(leftp1,leftp2,/*minus*/rightp1,rightp2,
218 		 /*into*/resultp1,resultp2);
219 		if(Dbl_iszero_mantissa(resultp1,resultp2))
220 		    {
221 		    if(Is_rounding_mode(ROUNDMINUS))
222 			{
223 			Dbl_setone_sign(resultp1);
224 			}
225 		    else
226 			{
227 			Dbl_setzero_sign(resultp1);
228 			}
229 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
230 		    return(NOEXCEPTION);
231 		    }
232 		}
233 	    else
234 		{
235 		Dbl_addition(leftp1,leftp2,rightp1,rightp2,
236 		 /*into*/resultp1,resultp2);
237 		if(Dbl_isone_hidden(resultp1))
238 		    {
239 		    Dbl_copytoptr(resultp1,resultp2,dstptr);
240 		    return(NOEXCEPTION);
241 		    }
242 		}
243 	    if(Is_underflowtrap_enabled())
244 		{
245 		/* need to normalize result */
246 	    	sign_save = Dbl_signextendedsign(resultp1);
247 		Dbl_leftshiftby1(resultp1,resultp2);
248 		Dbl_normalize(resultp1,resultp2,result_exponent);
249 		Dbl_set_sign(resultp1,/*using*/sign_save);
250                 Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
251 		Dbl_copytoptr(resultp1,resultp2,dstptr);
252 		/* inexact = FALSE */
253 		return(UNDERFLOWEXCEPTION);
254 		}
255 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
256 	    return(NOEXCEPTION);
257 	    }
258 	right_exponent = 1;	/* Set exponent to reflect different bias
259 				 * with denomalized numbers. */
260 	}
261     else
262 	{
263 	Dbl_clear_signexponent_set_hidden(rightp1);
264 	}
265     Dbl_clear_exponent_set_hidden(leftp1);
266     diff_exponent = result_exponent - right_exponent;
267 
268     /*
269      * Special case alignment of operands that would force alignment
270      * beyond the extent of the extension.  A further optimization
271      * could special case this but only reduces the path length for this
272      * infrequent case.
273      */
274     if(diff_exponent > DBL_THRESHOLD)
275 	{
276 	diff_exponent = DBL_THRESHOLD;
277 	}
278 
279     /* Align right operand by shifting to right */
280     Dbl_right_align(/*operand*/rightp1,rightp2,/*shifted by*/diff_exponent,
281      /*and lower to*/extent);
282 
283     /* Treat sum and difference of the operands separately. */
284     if( (/*signed*/int) save >= 0 )
285 	{
286 	/*
287 	 * Difference of the two operands.  Their can be no overflow.  A
288 	 * borrow can occur out of the hidden bit and force a post
289 	 * normalization phase.
290 	 */
291 	Dbl_subtract_withextension(leftp1,leftp2,/*minus*/rightp1,rightp2,
292 	 /*with*/extent,/*into*/resultp1,resultp2);
293 	if(Dbl_iszero_hidden(resultp1))
294 	    {
295 	    /* Handle normalization */
296 	    /* A straight forward algorithm would now shift the result
297 	     * and extension left until the hidden bit becomes one.  Not
298 	     * all of the extension bits need participate in the shift.
299 	     * Only the two most significant bits (round and guard) are
300 	     * needed.  If only a single shift is needed then the guard
301 	     * bit becomes a significant low order bit and the extension
302 	     * must participate in the rounding.  If more than a single
303 	     * shift is needed, then all bits to the right of the guard
304 	     * bit are zeros, and the guard bit may or may not be zero. */
305 	    sign_save = Dbl_signextendedsign(resultp1);
306             Dbl_leftshiftby1_withextent(resultp1,resultp2,extent,resultp1,resultp2);
307 
308             /* Need to check for a zero result.  The sign and exponent
309 	     * fields have already been zeroed.  The more efficient test
310 	     * of the full object can be used.
311 	     */
312     	    if(Dbl_iszero(resultp1,resultp2))
313 		/* Must have been "x-x" or "x+(-x)". */
314 		{
315 		if(Is_rounding_mode(ROUNDMINUS)) Dbl_setone_sign(resultp1);
316 		Dbl_copytoptr(resultp1,resultp2,dstptr);
317 		return(NOEXCEPTION);
318 		}
319 	    result_exponent--;
320 	    /* Look to see if normalization is finished. */
321 	    if(Dbl_isone_hidden(resultp1))
322 		{
323 		if(result_exponent==0)
324 		    {
325 		    /* Denormalized, exponent should be zero.  Left operand *
326 		     * was normalized, so extent (guard, round) was zero    */
327 		    goto underflow;
328 		    }
329 		else
330 		    {
331 		    /* No further normalization is needed. */
332 		    Dbl_set_sign(resultp1,/*using*/sign_save);
333 	    	    Ext_leftshiftby1(extent);
334 		    goto round;
335 		    }
336 		}
337 
338 	    /* Check for denormalized, exponent should be zero.  Left    *
339 	     * operand was normalized, so extent (guard, round) was zero */
340 	    if(!(underflowtrap = Is_underflowtrap_enabled()) &&
341 	       result_exponent==0) goto underflow;
342 
343 	    /* Shift extension to complete one bit of normalization and
344 	     * update exponent. */
345 	    Ext_leftshiftby1(extent);
346 
347 	    /* Discover first one bit to determine shift amount.  Use a
348 	     * modified binary search.  We have already shifted the result
349 	     * one position right and still not found a one so the remainder
350 	     * of the extension must be zero and simplifies rounding. */
351 	    /* Scan bytes */
352 	    while(Dbl_iszero_hiddenhigh7mantissa(resultp1))
353 		{
354 		Dbl_leftshiftby8(resultp1,resultp2);
355 		if((result_exponent -= 8) <= 0  && !underflowtrap)
356 		    goto underflow;
357 		}
358 	    /* Now narrow it down to the nibble */
359 	    if(Dbl_iszero_hiddenhigh3mantissa(resultp1))
360 		{
361 		/* The lower nibble contains the normalizing one */
362 		Dbl_leftshiftby4(resultp1,resultp2);
363 		if((result_exponent -= 4) <= 0 && !underflowtrap)
364 		    goto underflow;
365 		}
366 	    /* Select case were first bit is set (already normalized)
367 	     * otherwise select the proper shift. */
368 	    if((jumpsize = Dbl_hiddenhigh3mantissa(resultp1)) > 7)
369 		{
370 		/* Already normalized */
371 		if(result_exponent <= 0) goto underflow;
372 		Dbl_set_sign(resultp1,/*using*/sign_save);
373 		Dbl_set_exponent(resultp1,/*using*/result_exponent);
374 		Dbl_copytoptr(resultp1,resultp2,dstptr);
375 		return(NOEXCEPTION);
376 		}
377 	    Dbl_sethigh4bits(resultp1,/*using*/sign_save);
378 	    switch(jumpsize)
379 		{
380 		case 1:
381 		    {
382 		    Dbl_leftshiftby3(resultp1,resultp2);
383 		    result_exponent -= 3;
384 		    break;
385 		    }
386 		case 2:
387 		case 3:
388 		    {
389 		    Dbl_leftshiftby2(resultp1,resultp2);
390 		    result_exponent -= 2;
391 		    break;
392 		    }
393 		case 4:
394 		case 5:
395 		case 6:
396 		case 7:
397 		    {
398 		    Dbl_leftshiftby1(resultp1,resultp2);
399 		    result_exponent -= 1;
400 		    break;
401 		    }
402 		}
403 	    if(result_exponent > 0)
404 		{
405 		Dbl_set_exponent(resultp1,/*using*/result_exponent);
406 		Dbl_copytoptr(resultp1,resultp2,dstptr);
407 		return(NOEXCEPTION);		/* Sign bit is already set */
408 		}
409 	    /* Fixup potential underflows */
410 	  underflow:
411 	    if(Is_underflowtrap_enabled())
412 		{
413 		Dbl_set_sign(resultp1,sign_save);
414                 Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
415 		Dbl_copytoptr(resultp1,resultp2,dstptr);
416 		/* inexact = FALSE */
417 		return(UNDERFLOWEXCEPTION);
418 		}
419 	    /*
420 	     * Since we cannot get an inexact denormalized result,
421 	     * we can now return.
422 	     */
423 	    Dbl_fix_overshift(resultp1,resultp2,(1-result_exponent),extent);
424 	    Dbl_clear_signexponent(resultp1);
425 	    Dbl_set_sign(resultp1,sign_save);
426 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
427 	    return(NOEXCEPTION);
428 	    } /* end if(hidden...)... */
429 	/* Fall through and round */
430 	} /* end if(save >= 0)... */
431     else
432 	{
433 	/* Subtract magnitudes */
434 	Dbl_addition(leftp1,leftp2,rightp1,rightp2,/*to*/resultp1,resultp2);
435 	if(Dbl_isone_hiddenoverflow(resultp1))
436 	    {
437 	    /* Prenormalization required. */
438 	    Dbl_rightshiftby1_withextent(resultp2,extent,extent);
439 	    Dbl_arithrightshiftby1(resultp1,resultp2);
440 	    result_exponent++;
441 	    } /* end if hiddenoverflow... */
442 	} /* end else ...subtract magnitudes... */
443 
444     /* Round the result.  If the extension is all zeros,then the result is
445      * exact.  Otherwise round in the correct direction.  No underflow is
446      * possible. If a postnormalization is necessary, then the mantissa is
447      * all zeros so no shift is needed. */
448   round:
449     if(Ext_isnotzero(extent))
450 	{
451 	inexact = TRUE;
452 	switch(Rounding_mode())
453 	    {
454 	    case ROUNDNEAREST: /* The default. */
455 	    if(Ext_isone_sign(extent))
456 		{
457 		/* at least 1/2 ulp */
458 		if(Ext_isnotzero_lower(extent)  ||
459 		  Dbl_isone_lowmantissap2(resultp2))
460 		    {
461 		    /* either exactly half way and odd or more than 1/2ulp */
462 		    Dbl_increment(resultp1,resultp2);
463 		    }
464 		}
465 	    break;
466 
467 	    case ROUNDPLUS:
468 	    if(Dbl_iszero_sign(resultp1))
469 		{
470 		/* Round up positive results */
471 		Dbl_increment(resultp1,resultp2);
472 		}
473 	    break;
474 
475 	    case ROUNDMINUS:
476 	    if(Dbl_isone_sign(resultp1))
477 		{
478 		/* Round down negative results */
479 		Dbl_increment(resultp1,resultp2);
480 		}
481 
482 	    case ROUNDZERO:;
483 	    /* truncate is simple */
484 	    } /* end switch... */
485 	if(Dbl_isone_hiddenoverflow(resultp1)) result_exponent++;
486 	}
487     if(result_exponent == DBL_INFINITY_EXPONENT)
488         {
489         /* Overflow */
490         if(Is_overflowtrap_enabled())
491 	    {
492 	    Dbl_setwrapped_exponent(resultp1,result_exponent,ovfl);
493 	    Dbl_copytoptr(resultp1,resultp2,dstptr);
494 	    if (inexact)
495 	    if (Is_inexacttrap_enabled())
496 		return(OVERFLOWEXCEPTION | INEXACTEXCEPTION);
497 		else Set_inexactflag();
498 	    return(OVERFLOWEXCEPTION);
499 	    }
500         else
501 	    {
502 	    inexact = TRUE;
503 	    Set_overflowflag();
504 	    Dbl_setoverflow(resultp1,resultp2);
505 	    }
506 	}
507     else Dbl_set_exponent(resultp1,result_exponent);
508     Dbl_copytoptr(resultp1,resultp2,dstptr);
509     if(inexact)
510 	if(Is_inexacttrap_enabled()) return(INEXACTEXCEPTION);
511 	else Set_inexactflag();
512     return(NOEXCEPTION);
513     }
514