| blob | f5235f5d9726cc3099e87c2d8a970c98d7e6f060 |
1 /* Software floating-point emulation.
2 Basic four-word fraction declaration and manipulation.
3 Copyright (C) 1997,1998,1999 Free Software Foundation, Inc.
4 This file is part of the GNU C Library.
5 Contributed by Richard Henderson (rth@cygnus.com),
6 Jakub Jelinek (jj@ultra.linux.cz),
7 David S. Miller (davem@redhat.com) and
8 Peter Maydell (pmaydell@chiark.greenend.org.uk).
10 The GNU C Library is free software; you can redistribute it and/or
11 modify it under the terms of the GNU Library General Public License as
12 published by the Free Software Foundation; either version 2 of the
13 License, or (at your option) any later version.
15 The GNU C Library is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 Library General Public License for more details.
20 You should have received a copy of the GNU Library General Public
21 License along with the GNU C Library; see the file COPYING.LIB. If
22 not, write to the Free Software Foundation, Inc.,
23 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. */
25 #define _FP_FRAC_DECL_4(X) _FP_W_TYPE X##_f[4]
26 #define _FP_FRAC_COPY_4(D,S) \
27 (D##_f[0] = S##_f[0], D##_f[1] = S##_f[1], \
28 D##_f[2] = S##_f[2], D##_f[3] = S##_f[3])
29 #define _FP_FRAC_SET_4(X,I) __FP_FRAC_SET_4(X, I)
30 #define _FP_FRAC_HIGH_4(X) (X##_f[3])
31 #define _FP_FRAC_LOW_4(X) (X##_f[0])
32 #define _FP_FRAC_WORD_4(X,w) (X##_f[w])
34 #define _FP_FRAC_SLL_4(X,N) \
35 do { \
36 _FP_I_TYPE _up, _down, _skip, _i; \
37 _skip = (N) / _FP_W_TYPE_SIZE; \
38 _up = (N) % _FP_W_TYPE_SIZE; \
39 _down = _FP_W_TYPE_SIZE - _up; \
40 if (!_up) \
41 for (_i = 3; _i >= _skip; --_i) \
42 X##_f[_i] = X##_f[_i-_skip]; \
43 else \
44 { \
45 for (_i = 3; _i > _skip; --_i) \
46 X##_f[_i] = X##_f[_i-_skip] << _up \
47 | X##_f[_i-_skip-1] >> _down; \
48 X##_f[_i--] = X##_f[0] << _up; \
49 } \
50 for (; _i >= 0; --_i) \
51 X##_f[_i] = 0; \
52 } while (0)
54 /* This one was broken too */
55 #define _FP_FRAC_SRL_4(X,N) \
56 do { \
57 _FP_I_TYPE _up, _down, _skip, _i; \
58 _skip = (N) / _FP_W_TYPE_SIZE; \
59 _down = (N) % _FP_W_TYPE_SIZE; \
60 _up = _FP_W_TYPE_SIZE - _down; \
61 if (!_down) \
62 for (_i = 0; _i <= 3-_skip; ++_i) \
63 X##_f[_i] = X##_f[_i+_skip]; \
64 else \
65 { \
66 for (_i = 0; _i < 3-_skip; ++_i) \
67 X##_f[_i] = X##_f[_i+_skip] >> _down \
68 | X##_f[_i+_skip+1] << _up; \
69 X##_f[_i++] = X##_f[3] >> _down; \
70 } \
71 for (; _i < 4; ++_i) \
72 X##_f[_i] = 0; \
73 } while (0)
76 /* Right shift with sticky-lsb.
77 * What this actually means is that we do a standard right-shift,
78 * but that if any of the bits that fall off the right hand side
79 * were one then we always set the LSbit.
80 */
81 #define _FP_FRAC_SRS_4(X,N,size) \
82 do { \
83 _FP_I_TYPE _up, _down, _skip, _i; \
84 _FP_W_TYPE _s; \
85 _skip = (N) / _FP_W_TYPE_SIZE; \
86 _down = (N) % _FP_W_TYPE_SIZE; \
87 _up = _FP_W_TYPE_SIZE - _down; \
88 for (_s = _i = 0; _i < _skip; ++_i) \
89 _s |= X##_f[_i]; \
90 _s |= X##_f[_i] << _up; \
92 if (!_down) \
93 for (_i = 0; _i <= 3-_skip; ++_i) \
94 X##_f[_i] = X##_f[_i+_skip]; \
95 else \
96 { \
97 for (_i = 0; _i < 3-_skip; ++_i) \
98 X##_f[_i] = X##_f[_i+_skip] >> _down \
99 | X##_f[_i+_skip+1] << _up; \
100 X##_f[_i++] = X##_f[3] >> _down; \
101 } \
102 for (; _i < 4; ++_i) \
103 X##_f[_i] = 0; \
105 X##_f[0] |= (_s != 0); \
106 } while (0)
108 #define _FP_FRAC_ADD_4(R,X,Y) \
109 __FP_FRAC_ADD_4(R##_f[3], R##_f[2], R##_f[1], R##_f[0], \
110 X##_f[3], X##_f[2], X##_f[1], X##_f[0], \
111 Y##_f[3], Y##_f[2], Y##_f[1], Y##_f[0])
113 #define _FP_FRAC_SUB_4(R,X,Y) \
114 __FP_FRAC_SUB_4(R##_f[3], R##_f[2], R##_f[1], R##_f[0], \
115 X##_f[3], X##_f[2], X##_f[1], X##_f[0], \
116 Y##_f[3], Y##_f[2], Y##_f[1], Y##_f[0])
118 #define _FP_FRAC_DEC_4(X,Y) \
119 __FP_FRAC_DEC_4(X##_f[3], X##_f[2], X##_f[1], X##_f[0], \
120 Y##_f[3], Y##_f[2], Y##_f[1], Y##_f[0])
122 #define _FP_FRAC_ADDI_4(X,I) \
123 __FP_FRAC_ADDI_4(X##_f[3], X##_f[2], X##_f[1], X##_f[0], I)
125 #define _FP_ZEROFRAC_4 0,0,0,0
126 #define _FP_MINFRAC_4 0,0,0,1
127 #define _FP_MAXFRAC_4 (~(_FP_WS_TYPE)0), (~(_FP_WS_TYPE)0), (~(_FP_WS_TYPE)0), (~(_FP_WS_TYPE)0)
129 #define _FP_FRAC_ZEROP_4(X) ((X##_f[0] | X##_f[1] | X##_f[2] | X##_f[3]) == 0)
130 #define _FP_FRAC_NEGP_4(X) ((_FP_WS_TYPE)X##_f[3] < 0)
131 #define _FP_FRAC_OVERP_4(fs,X) (_FP_FRAC_HIGH_##fs(X) & _FP_OVERFLOW_##fs)
133 #define _FP_FRAC_EQ_4(X,Y) \
134 (X##_f[0] == Y##_f[0] && X##_f[1] == Y##_f[1] \
135 && X##_f[2] == Y##_f[2] && X##_f[3] == Y##_f[3])
137 #define _FP_FRAC_GT_4(X,Y) \
138 (X##_f[3] > Y##_f[3] || \
139 (X##_f[3] == Y##_f[3] && (X##_f[2] > Y##_f[2] || \
140 (X##_f[2] == Y##_f[2] && (X##_f[1] > Y##_f[1] || \
141 (X##_f[1] == Y##_f[1] && X##_f[0] > Y##_f[0]) \
142 )) \
143 )) \
144 )
146 #define _FP_FRAC_GE_4(X,Y) \
147 (X##_f[3] > Y##_f[3] || \
148 (X##_f[3] == Y##_f[3] && (X##_f[2] > Y##_f[2] || \
149 (X##_f[2] == Y##_f[2] && (X##_f[1] > Y##_f[1] || \
150 (X##_f[1] == Y##_f[1] && X##_f[0] >= Y##_f[0]) \
151 )) \
152 )) \
153 )
156 #define _FP_FRAC_CLZ_4(R,X) \
157 do { \
158 if (X##_f[3]) \
159 { \
160 __FP_CLZ(R,X##_f[3]); \
161 } \
162 else if (X##_f[2]) \
163 { \
164 __FP_CLZ(R,X##_f[2]); \
165 R += _FP_W_TYPE_SIZE; \
166 } \
167 else if (X##_f[1]) \
168 { \
169 __FP_CLZ(R,X##_f[2]); \
170 R += _FP_W_TYPE_SIZE*2; \
171 } \
172 else \
173 { \
174 __FP_CLZ(R,X##_f[0]); \
175 R += _FP_W_TYPE_SIZE*3; \
176 } \
177 } while(0)
180 #define _FP_UNPACK_RAW_4(fs, X, val) \
181 do { \
182 union _FP_UNION_##fs _flo; _flo.flt = (val); \
183 X##_f[0] = _flo.bits.frac0; \
184 X##_f[1] = _flo.bits.frac1; \
185 X##_f[2] = _flo.bits.frac2; \
186 X##_f[3] = _flo.bits.frac3; \
187 X##_e = _flo.bits.exp; \
188 X##_s = _flo.bits.sign; \
189 } while (0)
191 #define _FP_UNPACK_RAW_4_P(fs, X, val) \
192 do { \
193 union _FP_UNION_##fs *_flo = \
194 (union _FP_UNION_##fs *)(val); \
195 \
196 X##_f[0] = _flo->bits.frac0; \
197 X##_f[1] = _flo->bits.frac1; \
198 X##_f[2] = _flo->bits.frac2; \
199 X##_f[3] = _flo->bits.frac3; \
200 X##_e = _flo->bits.exp; \
201 X##_s = _flo->bits.sign; \
202 } while (0)
204 #define _FP_PACK_RAW_4(fs, val, X) \
205 do { \
206 union _FP_UNION_##fs _flo; \
207 _flo.bits.frac0 = X##_f[0]; \
208 _flo.bits.frac1 = X##_f[1]; \
209 _flo.bits.frac2 = X##_f[2]; \
210 _flo.bits.frac3 = X##_f[3]; \
211 _flo.bits.exp = X##_e; \
212 _flo.bits.sign = X##_s; \
213 (val) = _flo.flt; \
214 } while (0)
216 #define _FP_PACK_RAW_4_P(fs, val, X) \
217 do { \
218 union _FP_UNION_##fs *_flo = \
219 (union _FP_UNION_##fs *)(val); \
220 \
221 _flo->bits.frac0 = X##_f[0]; \
222 _flo->bits.frac1 = X##_f[1]; \
223 _flo->bits.frac2 = X##_f[2]; \
224 _flo->bits.frac3 = X##_f[3]; \
225 _flo->bits.exp = X##_e; \
226 _flo->bits.sign = X##_s; \
227 } while (0)
229 /*
230 * Multiplication algorithms:
231 */
233 /* Given a 1W * 1W => 2W primitive, do the extended multiplication. */
235 #define _FP_MUL_MEAT_4_wide(wfracbits, R, X, Y, doit) \
236 do { \
237 _FP_FRAC_DECL_8(_z); _FP_FRAC_DECL_2(_b); _FP_FRAC_DECL_2(_c); \
238 _FP_FRAC_DECL_2(_d); _FP_FRAC_DECL_2(_e); _FP_FRAC_DECL_2(_f); \
239 \
240 doit(_FP_FRAC_WORD_8(_z,1), _FP_FRAC_WORD_8(_z,0), X##_f[0], Y##_f[0]); \
241 doit(_b_f1, _b_f0, X##_f[0], Y##_f[1]); \
242 doit(_c_f1, _c_f0, X##_f[1], Y##_f[0]); \
243 doit(_d_f1, _d_f0, X##_f[1], Y##_f[1]); \
244 doit(_e_f1, _e_f0, X##_f[0], Y##_f[2]); \
245 doit(_f_f1, _f_f0, X##_f[2], Y##_f[0]); \
246 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,3),_FP_FRAC_WORD_8(_z,2), \
247 _FP_FRAC_WORD_8(_z,1), 0,_b_f1,_b_f0, \
248 0,0,_FP_FRAC_WORD_8(_z,1)); \
249 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,3),_FP_FRAC_WORD_8(_z,2), \
250 _FP_FRAC_WORD_8(_z,1), 0,_c_f1,_c_f0, \
251 _FP_FRAC_WORD_8(_z,3),_FP_FRAC_WORD_8(_z,2), \
252 _FP_FRAC_WORD_8(_z,1)); \
253 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,4),_FP_FRAC_WORD_8(_z,3), \
254 _FP_FRAC_WORD_8(_z,2), 0,_d_f1,_d_f0, \
255 0,_FP_FRAC_WORD_8(_z,3),_FP_FRAC_WORD_8(_z,2)); \
256 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,4),_FP_FRAC_WORD_8(_z,3), \
257 _FP_FRAC_WORD_8(_z,2), 0,_e_f1,_e_f0, \
258 _FP_FRAC_WORD_8(_z,4),_FP_FRAC_WORD_8(_z,3), \
259 _FP_FRAC_WORD_8(_z,2)); \
260 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,4),_FP_FRAC_WORD_8(_z,3), \
261 _FP_FRAC_WORD_8(_z,2), 0,_f_f1,_f_f0, \
262 _FP_FRAC_WORD_8(_z,4),_FP_FRAC_WORD_8(_z,3), \
263 _FP_FRAC_WORD_8(_z,2)); \
264 doit(_b_f1, _b_f0, X##_f[0], Y##_f[3]); \
265 doit(_c_f1, _c_f0, X##_f[3], Y##_f[0]); \
266 doit(_d_f1, _d_f0, X##_f[1], Y##_f[2]); \
267 doit(_e_f1, _e_f0, X##_f[2], Y##_f[1]); \
268 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4), \
269 _FP_FRAC_WORD_8(_z,3), 0,_b_f1,_b_f0, \
270 0,_FP_FRAC_WORD_8(_z,4),_FP_FRAC_WORD_8(_z,3)); \
271 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4), \
272 _FP_FRAC_WORD_8(_z,3), 0,_c_f1,_c_f0, \
273 _FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4), \
274 _FP_FRAC_WORD_8(_z,3)); \
275 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4), \
276 _FP_FRAC_WORD_8(_z,3), 0,_d_f1,_d_f0, \
277 _FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4), \
278 _FP_FRAC_WORD_8(_z,3)); \
279 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4), \
280 _FP_FRAC_WORD_8(_z,3), 0,_e_f1,_e_f0, \
281 _FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4), \
282 _FP_FRAC_WORD_8(_z,3)); \
283 doit(_b_f1, _b_f0, X##_f[2], Y##_f[2]); \
284 doit(_c_f1, _c_f0, X##_f[1], Y##_f[3]); \
285 doit(_d_f1, _d_f0, X##_f[3], Y##_f[1]); \
286 doit(_e_f1, _e_f0, X##_f[2], Y##_f[3]); \
287 doit(_f_f1, _f_f0, X##_f[3], Y##_f[2]); \
288 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,6),_FP_FRAC_WORD_8(_z,5), \
289 _FP_FRAC_WORD_8(_z,4), 0,_b_f1,_b_f0, \
290 0,_FP_FRAC_WORD_8(_z,5),_FP_FRAC_WORD_8(_z,4)); \
291 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,6),_FP_FRAC_WORD_8(_z,5), \
292 _FP_FRAC_WORD_8(_z,4), 0,_c_f1,_c_f0, \
293 _FP_FRAC_WORD_8(_z,6),_FP_FRAC_WORD_8(_z,5), \
294 _FP_FRAC_WORD_8(_z,4)); \
295 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,6),_FP_FRAC_WORD_8(_z,5), \
296 _FP_FRAC_WORD_8(_z,4), 0,_d_f1,_d_f0, \
297 _FP_FRAC_WORD_8(_z,6),_FP_FRAC_WORD_8(_z,5), \
298 _FP_FRAC_WORD_8(_z,4)); \
299 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,7),_FP_FRAC_WORD_8(_z,6), \
300 _FP_FRAC_WORD_8(_z,5), 0,_e_f1,_e_f0, \
301 0,_FP_FRAC_WORD_8(_z,6),_FP_FRAC_WORD_8(_z,5)); \
302 __FP_FRAC_ADD_3(_FP_FRAC_WORD_8(_z,7),_FP_FRAC_WORD_8(_z,6), \
303 _FP_FRAC_WORD_8(_z,5), 0,_f_f1,_f_f0, \
304 _FP_FRAC_WORD_8(_z,7),_FP_FRAC_WORD_8(_z,6), \
305 _FP_FRAC_WORD_8(_z,5)); \
306 doit(_b_f1, _b_f0, X##_f[3], Y##_f[3]); \
307 __FP_FRAC_ADD_2(_FP_FRAC_WORD_8(_z,7),_FP_FRAC_WORD_8(_z,6), \
308 _b_f1,_b_f0, \
309 _FP_FRAC_WORD_8(_z,7),_FP_FRAC_WORD_8(_z,6)); \
310 \
311 /* Normalize since we know where the msb of the multiplicands \
312 were (bit B), we know that the msb of the of the product is \
314 _FP_FRAC_SRS_8(_z, wfracbits-1, 2*wfracbits); \
315 __FP_FRAC_SET_4(R, _FP_FRAC_WORD_8(_z,3), _FP_FRAC_WORD_8(_z,2), \
316 _FP_FRAC_WORD_8(_z,1), _FP_FRAC_WORD_8(_z,0)); \
317 } while (0)
319 #define _FP_MUL_MEAT_4_gmp(wfracbits, R, X, Y) \
320 do { \
321 _FP_FRAC_DECL_8(_z); \
322 \
323 mpn_mul_n(_z_f, _x_f, _y_f, 4); \
324 \
325 /* Normalize since we know where the msb of the multiplicands \
326 were (bit B), we know that the msb of the of the product is \
328 _FP_FRAC_SRS_8(_z, wfracbits-1, 2*wfracbits); \
329 __FP_FRAC_SET_4(R, _FP_FRAC_WORD_8(_z,3), _FP_FRAC_WORD_8(_z,2), \
330 _FP_FRAC_WORD_8(_z,1), _FP_FRAC_WORD_8(_z,0)); \
331 } while (0)
333 /*
334 * Helper utility for _FP_DIV_MEAT_4_udiv:
335 * pppp = m * nnn
336 */
337 #define umul_ppppmnnn(p3,p2,p1,p0,m,n2,n1,n0) \
338 do { \
339 UWtype _t; \
340 umul_ppmm(p1,p0,m,n0); \
341 umul_ppmm(p2,_t,m,n1); \
342 __FP_FRAC_ADDI_2(p2,p1,_t); \
343 umul_ppmm(p3,_t,m,n2); \
344 __FP_FRAC_ADDI_2(p3,p2,_t); \
345 } while (0)
347 /*
348 * Division algorithms:
349 */
351 #define _FP_DIV_MEAT_4_udiv(fs, R, X, Y) \
352 do { \
353 int _i; \
354 _FP_FRAC_DECL_4(_n); _FP_FRAC_DECL_4(_m); \
355 _FP_FRAC_SET_4(_n, _FP_ZEROFRAC_4); \
356 if (_FP_FRAC_GT_4(X, Y)) \
357 { \
358 _n_f[3] = X##_f[0] << (_FP_W_TYPE_SIZE - 1); \
359 _FP_FRAC_SRL_4(X, 1); \
360 } \
361 else \
362 R##_e--; \
363 \
364 /* Normalize, i.e. make the most significant bit of the \
366 _FP_FRAC_SLL_4(Y, _FP_WFRACXBITS_##fs); \
367 \
368 for (_i = 3; ; _i--) \
369 { \
370 if (X##_f[3] == Y##_f[3]) \
371 { \
372 /* This is a special case, not an optimization \
373 (X##_f[3]/Y##_f[3] would not fit into UWtype). \
374 As X## is guaranteed to be < Y, R##_f[_i] can be either \
376 R##_f[_i] = -1; \
377 if (!_i) \
378 break; \
379 __FP_FRAC_SUB_4(X##_f[3], X##_f[2], X##_f[1], X##_f[0], \
380 Y##_f[2], Y##_f[1], Y##_f[0], 0, \
381 X##_f[2], X##_f[1], X##_f[0], _n_f[_i]); \
382 _FP_FRAC_SUB_4(X, Y, X); \
383 if (X##_f[3] > Y##_f[3]) \
384 { \
385 R##_f[_i] = -2; \
386 _FP_FRAC_ADD_4(X, Y, X); \
387 } \
388 } \
389 else \
390 { \
391 udiv_qrnnd(R##_f[_i], X##_f[3], X##_f[3], X##_f[2], Y##_f[3]); \
392 umul_ppppmnnn(_m_f[3], _m_f[2], _m_f[1], _m_f[0], \
393 R##_f[_i], Y##_f[2], Y##_f[1], Y##_f[0]); \
394 X##_f[2] = X##_f[1]; \
395 X##_f[1] = X##_f[0]; \
396 X##_f[0] = _n_f[_i]; \
397 if (_FP_FRAC_GT_4(_m, X)) \
398 { \
399 R##_f[_i]--; \
400 _FP_FRAC_ADD_4(X, Y, X); \
401 if (_FP_FRAC_GE_4(X, Y) && _FP_FRAC_GT_4(_m, X)) \
402 { \
403 R##_f[_i]--; \
404 _FP_FRAC_ADD_4(X, Y, X); \
405 } \
406 } \
407 _FP_FRAC_DEC_4(X, _m); \
408 if (!_i) \
409 { \
410 if (!_FP_FRAC_EQ_4(X, _m)) \
411 R##_f[0] |= _FP_WORK_STICKY; \
412 break; \
413 } \
414 } \
415 } \
416 } while (0)
419 /*
420 * Square root algorithms:
421 * We have just one right now, maybe Newton approximation
422 * should be added for those machines where division is fast.
423 */
425 #define _FP_SQRT_MEAT_4(R, S, T, X, q) \
426 do { \
427 while (q) \
428 { \
429 T##_f[3] = S##_f[3] + q; \
430 if (T##_f[3] <= X##_f[3]) \
431 { \
432 S##_f[3] = T##_f[3] + q; \
433 X##_f[3] -= T##_f[3]; \
434 R##_f[3] += q; \
435 } \
436 _FP_FRAC_SLL_4(X, 1); \
437 q >>= 1; \
438 } \
439 q = (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE - 1); \
440 while (q) \
441 { \
442 T##_f[2] = S##_f[2] + q; \
443 T##_f[3] = S##_f[3]; \
444 if (T##_f[3] < X##_f[3] || \
445 (T##_f[3] == X##_f[3] && T##_f[2] <= X##_f[2])) \
446 { \
447 S##_f[2] = T##_f[2] + q; \
448 S##_f[3] += (T##_f[2] > S##_f[2]); \
449 __FP_FRAC_DEC_2(X##_f[3], X##_f[2], \
450 T##_f[3], T##_f[2]); \
451 R##_f[2] += q; \
452 } \
453 _FP_FRAC_SLL_4(X, 1); \
454 q >>= 1; \
455 } \
456 q = (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE - 1); \
457 while (q) \
458 { \
459 T##_f[1] = S##_f[1] + q; \
460 T##_f[2] = S##_f[2]; \
461 T##_f[3] = S##_f[3]; \
462 if (T##_f[3] < X##_f[3] || \
463 (T##_f[3] == X##_f[3] && (T##_f[2] < X##_f[2] || \
464 (T##_f[2] == X##_f[2] && T##_f[1] <= X##_f[1])))) \
465 { \
466 S##_f[1] = T##_f[1] + q; \
467 S##_f[2] += (T##_f[1] > S##_f[1]); \
468 S##_f[3] += (T##_f[2] > S##_f[2]); \
469 __FP_FRAC_DEC_3(X##_f[3], X##_f[2], X##_f[1], \
470 T##_f[3], T##_f[2], T##_f[1]); \
471 R##_f[1] += q; \
472 } \
473 _FP_FRAC_SLL_4(X, 1); \
474 q >>= 1; \
475 } \
476 q = (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE - 1); \
477 while (q != _FP_WORK_ROUND) \
478 { \
479 T##_f[0] = S##_f[0] + q; \
480 T##_f[1] = S##_f[1]; \
481 T##_f[2] = S##_f[2]; \
482 T##_f[3] = S##_f[3]; \
483 if (_FP_FRAC_GE_4(X,T)) \
484 { \
485 S##_f[0] = T##_f[0] + q; \
486 S##_f[1] += (T##_f[0] > S##_f[0]); \
487 S##_f[2] += (T##_f[1] > S##_f[1]); \
488 S##_f[3] += (T##_f[2] > S##_f[2]); \
489 _FP_FRAC_DEC_4(X, T); \
490 R##_f[0] += q; \
491 } \
492 _FP_FRAC_SLL_4(X, 1); \
493 q >>= 1; \
494 } \
495 if (!_FP_FRAC_ZEROP_4(X)) \
496 { \
497 if (_FP_FRAC_GT_4(X,S)) \
498 R##_f[0] |= _FP_WORK_ROUND; \
499 R##_f[0] |= _FP_WORK_STICKY; \
500 } \
501 } while (0)
504 /*
505 * Internals
506 */
508 #define __FP_FRAC_SET_4(X,I3,I2,I1,I0) \
509 (X##_f[3] = I3, X##_f[2] = I2, X##_f[1] = I1, X##_f[0] = I0)
511 #ifndef __FP_FRAC_ADD_3
512 #define __FP_FRAC_ADD_3(r2,r1,r0,x2,x1,x0,y2,y1,y0) \
513 (r0 = x0 + y0, \
514 r1 = x1 + y1 + (r0 < x0), \
515 r2 = x2 + y2 + (r1 < x1))
516 #endif
518 #ifndef __FP_FRAC_ADD_4
519 #define __FP_FRAC_ADD_4(r3,r2,r1,r0,x3,x2,x1,x0,y3,y2,y1,y0) \
520 (r0 = x0 + y0, \
521 r1 = x1 + y1 + (r0 < x0), \
522 r2 = x2 + y2 + (r1 < x1), \
523 r3 = x3 + y3 + (r2 < x2))
524 #endif
526 #ifndef __FP_FRAC_SUB_3
527 #define __FP_FRAC_SUB_3(r2,r1,r0,x2,x1,x0,y2,y1,y0) \
528 (r0 = x0 - y0, \
529 r1 = x1 - y1 - (r0 > x0), \
530 r2 = x2 - y2 - (r1 > x1))
531 #endif
533 #ifndef __FP_FRAC_SUB_4
534 #define __FP_FRAC_SUB_4(r3,r2,r1,r0,x3,x2,x1,x0,y3,y2,y1,y0) \
535 (r0 = x0 - y0, \
536 r1 = x1 - y1 - (r0 > x0), \
537 r2 = x2 - y2 - (r1 > x1), \
538 r3 = x3 - y3 - (r2 > x2))
539 #endif
541 #ifndef __FP_FRAC_DEC_3
542 #define __FP_FRAC_DEC_3(x2,x1,x0,y2,y1,y0) \
543 do { \
544 UWtype _t0, _t1; \
545 _t0 = x0; \
546 x0 -= y0; \
547 _t1 = x1; \
548 x1 -= y1 + (x0 > _t0); \
549 x2 -= y2 + (x1 > _t1); \
550 } while (0)
551 #endif
553 #ifndef __FP_FRAC_DEC_4
554 #define __FP_FRAC_DEC_4(x3,x2,x1,x0,y3,y2,y1,y0) \
555 do { \
556 UWtype _t0, _t1; \
557 _t0 = x0; \
558 x0 -= y0; \
559 _t1 = x1; \
560 x1 -= y1 + (x0 > _t0); \
561 _t0 = x2; \
562 x2 -= y2 + (x1 > _t1); \
563 x3 -= y3 + (x2 > _t0); \
564 } while (0)
565 #endif
567 #ifndef __FP_FRAC_ADDI_4
568 #define __FP_FRAC_ADDI_4(x3,x2,x1,x0,i) \
569 do { \
570 UWtype _t; \
571 _t = ((x0 += i) < i); \
572 x1 += _t; _t = (x1 < _t); \
573 x2 += _t; _t = (x2 < _t); \
574 x3 += _t; \
575 } while (0)
576 #endif
578 /* Convert FP values between word sizes. This appears to be more
579 * complicated than I'd have expected it to be, so these might be
580 * wrong... These macros are in any case somewhat bogus because they
581 * use information about what various FRAC_n variables look like
582 * internally [eg, that 2 word vars are X_f0 and x_f1]. But so do
583 * the ones in op-2.h and op-1.h.
584 */
585 #define _FP_FRAC_CONV_1_4(dfs, sfs, D, S) \
586 do { \
587 if (S##_c != FP_CLS_NAN) \
588 _FP_FRAC_SRS_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs), \
589 _FP_WFRACBITS_##sfs); \
590 else \
591 _FP_FRAC_SRL_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs)); \
592 D##_f = S##_f[0]; \
593 } while (0)
595 #define _FP_FRAC_CONV_2_4(dfs, sfs, D, S) \
596 do { \
597 if (S##_c != FP_CLS_NAN) \
598 _FP_FRAC_SRS_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs), \
599 _FP_WFRACBITS_##sfs); \
600 else \
601 _FP_FRAC_SRL_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs)); \
602 D##_f0 = S##_f[0]; \
603 D##_f1 = S##_f[1]; \
604 } while (0)
606 /* Assembly/disassembly for converting to/from integral types.
607 * No shifting or overflow handled here.
608 */
609 /* Put the FP value X into r, which is an integer of size rsize. */
610 #define _FP_FRAC_ASSEMBLE_4(r, X, rsize) \
611 do { \
612 if (rsize <= _FP_W_TYPE_SIZE) \
613 r = X##_f[0]; \
614 else if (rsize <= 2*_FP_W_TYPE_SIZE) \
615 { \
616 r = X##_f[1]; \
617 r <<= _FP_W_TYPE_SIZE; \
618 r += X##_f[0]; \
619 } \
620 else \
621 { \
624 r = X##_f[3]; \
625 r <<= _FP_W_TYPE_SIZE; \
626 r += X##_f[2]; \
627 r <<= _FP_W_TYPE_SIZE; \
628 r += X##_f[1]; \
629 r <<= _FP_W_TYPE_SIZE; \
630 r += X##_f[0]; \
631 } \
632 } while (0)
634 /* "No disassemble Number Five!" */
635 /* move an integer of size rsize into X's fractional part. We rely on
636 * the _f[] array consisting of words of size _FP_W_TYPE_SIZE to avoid
637 * having to mask the values we store into it.
638 */
639 #define _FP_FRAC_DISASSEMBLE_4(X, r, rsize) \
640 do { \
641 X##_f[0] = r; \
642 X##_f[1] = (rsize <= _FP_W_TYPE_SIZE ? 0 : r >> _FP_W_TYPE_SIZE); \
643 X##_f[2] = (rsize <= 2*_FP_W_TYPE_SIZE ? 0 : r >> 2*_FP_W_TYPE_SIZE); \
644 X##_f[3] = (rsize <= 3*_FP_W_TYPE_SIZE ? 0 : r >> 3*_FP_W_TYPE_SIZE); \
645 } while (0);
647 #define _FP_FRAC_CONV_4_1(dfs, sfs, D, S) \
648 do { \
649 D##_f[0] = S##_f; \
650 D##_f[1] = D##_f[2] = D##_f[3] = 0; \
651 _FP_FRAC_SLL_4(D, (_FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs)); \
652 } while (0)
654 #define _FP_FRAC_CONV_4_2(dfs, sfs, D, S) \
655 do { \
656 D##_f[0] = S##_f0; \
657 D##_f[1] = S##_f1; \
658 D##_f[2] = D##_f[3] = 0; \
659 _FP_FRAC_SLL_4(D, (_FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs)); \
660 } while (0)