1 /* $NetBSD: sfsub.c,v 1.3 2005/12/11 12:17:40 christos Exp $ */
3 /* $OpenBSD: sfsub.c,v 1.4 2001/03/29 03:58:19 mickey Exp $ */
6 * Copyright 1996 1995 by Open Software Foundation, Inc.
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.
15 * OSF DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE
16 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
17 * FOR A PARTICULAR PURPOSE.
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.
30 * (c) Copyright 1986 HEWLETT-PACKARD COMPANY
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.
44 #include <sys/cdefs.h>
45 __KERNEL_RCSID(0, "$NetBSD: sfsub.c,v 1.3 2005/12/11 12:17:40 christos Exp $");
47 #include "../spmath/float.h"
48 #include "../spmath/sgl_float.h"
51 * Single_subtract: subtract two single precision values.
54 sgl_fsub(leftptr
, rightptr
, dstptr
, status
)
55 sgl_floating_point
*leftptr
, *rightptr
, *dstptr
;
58 register unsigned int left
, right
, result
, extent
;
59 register unsigned int signless_upper_left
, signless_upper_right
, save
;
61 register int result_exponent
, right_exponent
, diff_exponent
;
62 register int sign_save
, jumpsize
;
63 register int inexact
= false, underflowtrap
;
65 /* Create local copies of the numbers */
69 /* A zero "save" helps discover equal operands (for later), *
70 * and is used in swapping operands (if needed). */
71 Sgl_xortointp1(left
,right
,/*to*/save
);
74 * check first operand for NaN's or infinity
76 if ((result_exponent
= Sgl_exponent(left
)) == SGL_INFINITY_EXPONENT
)
78 if (Sgl_iszero_mantissa(left
))
80 if (Sgl_isnotnan(right
))
82 if (Sgl_isinfinity(right
) && save
==0)
85 * invalid since operands are same signed infinity's
87 if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION
);
89 Sgl_makequietnan(result
);
103 * is NaN; signaling or quiet?
105 if (Sgl_isone_signaling(left
))
107 /* trap if INVALIDTRAP enabled */
108 if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION
);
114 * is second operand a signaling NaN?
116 else if (Sgl_is_signalingnan(right
))
118 /* trap if INVALIDTRAP enabled */
119 if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION
);
122 Sgl_set_quiet(right
);
132 } /* End left NaN or Infinity processing */
134 * check second operand for NaN's or infinity
136 if (Sgl_isinfinity_exponent(right
))
138 if (Sgl_iszero_mantissa(right
))
140 /* return infinity */
141 Sgl_invert_sign(right
);
146 * is NaN; signaling or quiet?
148 if (Sgl_isone_signaling(right
))
150 /* trap if INVALIDTRAP enabled */
151 if (Is_invalidtrap_enabled()) return(INVALIDEXCEPTION
);
154 Sgl_set_quiet(right
);
161 } /* End right NaN or Infinity processing */
163 /* Invariant: Must be dealing with finite numbers */
165 /* Compare operands by removing the sign */
166 Sgl_copytoint_exponentmantissa(left
,signless_upper_left
);
167 Sgl_copytoint_exponentmantissa(right
,signless_upper_right
);
169 /* sign difference selects sub or add operation. */
170 if(Sgl_ismagnitudeless(signless_upper_left
,signless_upper_right
))
172 /* Set the left operand to the larger one by XOR swap *
173 * First finish the first word using "save" */
174 Sgl_xorfromintp1(save
,right
,/*to*/right
);
175 Sgl_xorfromintp1(save
,left
,/*to*/left
);
176 result_exponent
= Sgl_exponent(left
);
177 Sgl_invert_sign(left
);
179 /* Invariant: left is not smaller than right. */
181 if((right_exponent
= Sgl_exponent(right
)) == 0)
183 /* Denormalized operands. First look for zeroes */
184 if(Sgl_iszero_mantissa(right
))
187 if(Sgl_iszero_exponentmantissa(left
))
189 /* Both operands are zeros */
190 Sgl_invert_sign(right
);
191 if(Is_rounding_mode(ROUNDMINUS
))
193 Sgl_or_signs(left
,/*with*/right
);
197 Sgl_and_signs(left
,/*with*/right
);
202 /* Left is not a zero and must be the result. Trapped
203 * underflows are signaled if left is denormalized. Result
204 * is always exact. */
205 if( (result_exponent
== 0) && Is_underflowtrap_enabled() )
207 /* need to normalize results mantissa */
208 sign_save
= Sgl_signextendedsign(left
);
209 Sgl_leftshiftby1(left
);
210 Sgl_normalize(left
,result_exponent
);
211 Sgl_set_sign(left
,/*using*/sign_save
);
212 Sgl_setwrapped_exponent(left
,result_exponent
,unfl
);
214 /* inexact = false */
215 return(UNDERFLOWEXCEPTION
);
222 /* Neither are zeroes */
223 Sgl_clear_sign(right
); /* Exponent is already cleared */
224 if(result_exponent
== 0 )
226 /* Both operands are denormalized. The result must be exact
227 * and is simply calculated. A sum could become normalized and a
228 * difference could cancel to a true zero. */
229 if( (/*signed*/int) save
>= 0 )
231 Sgl_subtract(left
,/*minus*/right
,/*into*/result
);
232 if(Sgl_iszero_mantissa(result
))
234 if(Is_rounding_mode(ROUNDMINUS
))
236 Sgl_setone_sign(result
);
240 Sgl_setzero_sign(result
);
248 Sgl_addition(left
,right
,/*into*/result
);
249 if(Sgl_isone_hidden(result
))
255 if(Is_underflowtrap_enabled())
257 /* need to normalize result */
258 sign_save
= Sgl_signextendedsign(result
);
259 Sgl_leftshiftby1(result
);
260 Sgl_normalize(result
,result_exponent
);
261 Sgl_set_sign(result
,/*using*/sign_save
);
262 Sgl_setwrapped_exponent(result
,result_exponent
,unfl
);
264 /* inexact = false */
265 return(UNDERFLOWEXCEPTION
);
270 right_exponent
= 1; /* Set exponent to reflect different bias
271 * with denomalized numbers. */
275 Sgl_clear_signexponent_set_hidden(right
);
277 Sgl_clear_exponent_set_hidden(left
);
278 diff_exponent
= result_exponent
- right_exponent
;
281 * Special case alignment of operands that would force alignment
282 * beyond the extent of the extension. A further optimization
283 * could special case this but only reduces the path length for this
286 if(diff_exponent
> SGL_THRESHOLD
)
288 diff_exponent
= SGL_THRESHOLD
;
291 /* Align right operand by shifting to right */
292 Sgl_right_align(/*operand*/right
,/*shifted by*/diff_exponent
,
293 /*and lower to*/extent
);
295 /* Treat sum and difference of the operands separately. */
296 if( (/*signed*/int) save
>= 0 )
299 * Difference of the two operands. Their can be no overflow. A
300 * borrow can occur out of the hidden bit and force a post
301 * normalization phase.
303 Sgl_subtract_withextension(left
,/*minus*/right
,/*with*/extent
,/*into*/result
);
304 if(Sgl_iszero_hidden(result
))
306 /* Handle normalization */
307 /* A straight foward algorithm would now shift the result
308 * and extension left until the hidden bit becomes one. Not
309 * all of the extension bits need participate in the shift.
310 * Only the two most significant bits (round and guard) are
311 * needed. If only a single shift is needed then the guard
312 * bit becomes a significant low order bit and the extension
313 * must participate in the rounding. If more than a single
314 * shift is needed, then all bits to the right of the guard
315 * bit are zeros, and the guard bit may or may not be zero. */
316 sign_save
= Sgl_signextendedsign(result
);
317 Sgl_leftshiftby1_withextent(result
,extent
,result
);
319 /* Need to check for a zero result. The sign and exponent
320 * fields have already been zeroed. The more efficient test
321 * of the full object can be used.
323 if(Sgl_iszero(result
))
324 /* Must have been "x-x" or "x+(-x)". */
326 if(Is_rounding_mode(ROUNDMINUS
)) Sgl_setone_sign(result
);
331 /* Look to see if normalization is finished. */
332 if(Sgl_isone_hidden(result
))
334 if(result_exponent
==0)
336 /* Denormalized, exponent should be zero. Left operand *
337 * was normalized, so extent (guard, round) was zero */
342 /* No further normalization is needed. */
343 Sgl_set_sign(result
,/*using*/sign_save
);
344 Ext_leftshiftby1(extent
);
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
;
354 /* Shift extension to complete one bit of normalization and
355 * update exponent. */
356 Ext_leftshiftby1(extent
);
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. */
363 while(Sgl_iszero_hiddenhigh7mantissa(result
))
365 Sgl_leftshiftby8(result
);
366 if((result_exponent
-= 8) <= 0 && !underflowtrap
)
369 /* Now narrow it down to the nibble */
370 if(Sgl_iszero_hiddenhigh3mantissa(result
))
372 /* The lower nibble contains the normalizing one */
373 Sgl_leftshiftby4(result
);
374 if((result_exponent
-= 4) <= 0 && !underflowtrap
)
377 /* Select case were first bit is set (already normalized)
378 * otherwise select the proper shift. */
379 if((jumpsize
= Sgl_hiddenhigh3mantissa(result
)) > 7)
381 /* Already normalized */
382 if(result_exponent
<= 0) goto underflow
;
383 Sgl_set_sign(result
,/*using*/sign_save
);
384 Sgl_set_exponent(result
,/*using*/result_exponent
);
388 Sgl_sethigh4bits(result
,/*using*/sign_save
);
393 Sgl_leftshiftby3(result
);
394 result_exponent
-= 3;
400 Sgl_leftshiftby2(result
);
401 result_exponent
-= 2;
409 Sgl_leftshiftby1(result
);
410 result_exponent
-= 1;
414 if(result_exponent
> 0)
416 Sgl_set_exponent(result
,/*using*/result_exponent
);
417 *dstptr
= result
; /* Sign bit is already set */
420 /* Fixup potential underflows */
422 if(Is_underflowtrap_enabled())
424 Sgl_set_sign(result
,sign_save
);
425 Sgl_setwrapped_exponent(result
,result_exponent
,unfl
);
427 /* inexact = false */
428 return(UNDERFLOWEXCEPTION
);
431 * Since we cannot get an inexact denormalized result,
434 Sgl_right_align(result
,/*by*/(1-result_exponent
),extent
);
435 Sgl_clear_signexponent(result
);
436 Sgl_set_sign(result
,sign_save
);
439 } /* end if(hidden...)... */
440 /* Fall through and round */
441 } /* end if(save >= 0)... */
445 Sgl_addition(left
,right
,/*to*/result
);
446 if(Sgl_isone_hiddenoverflow(result
))
448 /* Prenormalization required. */
449 Sgl_rightshiftby1_withextent(result
,extent
,extent
);
450 Sgl_arithrightshiftby1(result
);
452 } /* end if hiddenoverflow... */
453 } /* end else ...sub magnitudes... */
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. */
460 if(Ext_isnotzero(extent
))
463 switch(Rounding_mode())
465 case ROUNDNEAREST
: /* The default. */
466 if(Ext_isone_sign(extent
))
468 /* at least 1/2 ulp */
469 if(Ext_isnotzero_lower(extent
) ||
470 Sgl_isone_lowmantissa(result
))
472 /* either exactly half way and odd or more than 1/2ulp */
473 Sgl_increment(result
);
479 if(Sgl_iszero_sign(result
))
481 /* Round up positive results */
482 Sgl_increment(result
);
487 if(Sgl_isone_sign(result
))
489 /* Round down negative results */
490 Sgl_increment(result
);
494 /* truncate is simple */
495 } /* end switch... */
496 if(Sgl_isone_hiddenoverflow(result
)) result_exponent
++;
498 if(result_exponent
== SGL_INFINITY_EXPONENT
)
501 if(Is_overflowtrap_enabled())
503 Sgl_setwrapped_exponent(result
,result_exponent
,ovfl
);
506 if (Is_inexacttrap_enabled())
507 return(OVERFLOWEXCEPTION
| INEXACTEXCEPTION
);
508 else Set_inexactflag();
510 return(OVERFLOWEXCEPTION
);
516 Sgl_setoverflow(result
);
519 else Sgl_set_exponent(result
,result_exponent
);
522 if(Is_inexacttrap_enabled()) return(INEXACTEXCEPTION
);
523 else Set_inexactflag();