111
|
1 /* Software floating-point emulation.
|
|
2 Basic one-word fraction declaration and manipulation.
|
145
|
3 Copyright (C) 1997-2019 Free Software Foundation, Inc.
|
111
|
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).
|
|
9
|
|
10 The GNU C Library is free software; you can redistribute it and/or
|
|
11 modify it under the terms of the GNU Lesser General Public
|
|
12 License as published by the Free Software Foundation; either
|
|
13 version 2.1 of the License, or (at your option) any later version.
|
|
14
|
|
15 In addition to the permissions in the GNU Lesser General Public
|
|
16 License, the Free Software Foundation gives you unlimited
|
|
17 permission to link the compiled version of this file into
|
|
18 combinations with other programs, and to distribute those
|
|
19 combinations without any restriction coming from the use of this
|
|
20 file. (The Lesser General Public License restrictions do apply in
|
|
21 other respects; for example, they cover modification of the file,
|
|
22 and distribution when not linked into a combine executable.)
|
|
23
|
|
24 The GNU C Library is distributed in the hope that it will be useful,
|
|
25 but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
26 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
27 Lesser General Public License for more details.
|
|
28
|
|
29 You should have received a copy of the GNU Lesser General Public
|
|
30 License along with the GNU C Library; if not, see
|
|
31 <http://www.gnu.org/licenses/>. */
|
|
32
|
|
33 #ifndef SOFT_FP_OP_1_H
|
|
34 #define SOFT_FP_OP_1_H 1
|
|
35
|
|
36 #define _FP_FRAC_DECL_1(X) _FP_W_TYPE X##_f _FP_ZERO_INIT
|
|
37 #define _FP_FRAC_COPY_1(D, S) (D##_f = S##_f)
|
|
38 #define _FP_FRAC_SET_1(X, I) (X##_f = I)
|
|
39 #define _FP_FRAC_HIGH_1(X) (X##_f)
|
|
40 #define _FP_FRAC_LOW_1(X) (X##_f)
|
|
41 #define _FP_FRAC_WORD_1(X, w) (X##_f)
|
|
42
|
|
43 #define _FP_FRAC_ADDI_1(X, I) (X##_f += I)
|
|
44 #define _FP_FRAC_SLL_1(X, N) \
|
|
45 do \
|
|
46 { \
|
|
47 if (__builtin_constant_p (N) && (N) == 1) \
|
|
48 X##_f += X##_f; \
|
|
49 else \
|
|
50 X##_f <<= (N); \
|
|
51 } \
|
|
52 while (0)
|
|
53 #define _FP_FRAC_SRL_1(X, N) (X##_f >>= N)
|
|
54
|
|
55 /* Right shift with sticky-lsb. */
|
|
56 #define _FP_FRAC_SRST_1(X, S, N, sz) __FP_FRAC_SRST_1 (X##_f, S, (N), (sz))
|
|
57 #define _FP_FRAC_SRS_1(X, N, sz) __FP_FRAC_SRS_1 (X##_f, (N), (sz))
|
|
58
|
|
59 #define __FP_FRAC_SRST_1(X, S, N, sz) \
|
|
60 do \
|
|
61 { \
|
|
62 S = (__builtin_constant_p (N) && (N) == 1 \
|
|
63 ? X & 1 \
|
|
64 : (X << (_FP_W_TYPE_SIZE - (N))) != 0); \
|
|
65 X = X >> (N); \
|
|
66 } \
|
|
67 while (0)
|
|
68
|
|
69 #define __FP_FRAC_SRS_1(X, N, sz) \
|
|
70 (X = (X >> (N) | (__builtin_constant_p (N) && (N) == 1 \
|
|
71 ? X & 1 \
|
|
72 : (X << (_FP_W_TYPE_SIZE - (N))) != 0)))
|
|
73
|
|
74 #define _FP_FRAC_ADD_1(R, X, Y) (R##_f = X##_f + Y##_f)
|
|
75 #define _FP_FRAC_SUB_1(R, X, Y) (R##_f = X##_f - Y##_f)
|
|
76 #define _FP_FRAC_DEC_1(X, Y) (X##_f -= Y##_f)
|
|
77 #define _FP_FRAC_CLZ_1(z, X) __FP_CLZ ((z), X##_f)
|
|
78
|
|
79 /* Predicates. */
|
|
80 #define _FP_FRAC_NEGP_1(X) ((_FP_WS_TYPE) X##_f < 0)
|
|
81 #define _FP_FRAC_ZEROP_1(X) (X##_f == 0)
|
|
82 #define _FP_FRAC_OVERP_1(fs, X) (X##_f & _FP_OVERFLOW_##fs)
|
|
83 #define _FP_FRAC_CLEAR_OVERP_1(fs, X) (X##_f &= ~_FP_OVERFLOW_##fs)
|
|
84 #define _FP_FRAC_HIGHBIT_DW_1(fs, X) (X##_f & _FP_HIGHBIT_DW_##fs)
|
|
85 #define _FP_FRAC_EQ_1(X, Y) (X##_f == Y##_f)
|
|
86 #define _FP_FRAC_GE_1(X, Y) (X##_f >= Y##_f)
|
|
87 #define _FP_FRAC_GT_1(X, Y) (X##_f > Y##_f)
|
|
88
|
|
89 #define _FP_ZEROFRAC_1 0
|
|
90 #define _FP_MINFRAC_1 1
|
|
91 #define _FP_MAXFRAC_1 (~(_FP_WS_TYPE) 0)
|
|
92
|
|
93 /* Unpack the raw bits of a native fp value. Do not classify or
|
|
94 normalize the data. */
|
|
95
|
|
96 #define _FP_UNPACK_RAW_1(fs, X, val) \
|
|
97 do \
|
|
98 { \
|
|
99 union _FP_UNION_##fs _FP_UNPACK_RAW_1_flo; \
|
|
100 _FP_UNPACK_RAW_1_flo.flt = (val); \
|
|
101 \
|
|
102 X##_f = _FP_UNPACK_RAW_1_flo.bits.frac; \
|
|
103 X##_e = _FP_UNPACK_RAW_1_flo.bits.exp; \
|
|
104 X##_s = _FP_UNPACK_RAW_1_flo.bits.sign; \
|
|
105 } \
|
|
106 while (0)
|
|
107
|
|
108 #define _FP_UNPACK_RAW_1_P(fs, X, val) \
|
|
109 do \
|
|
110 { \
|
|
111 union _FP_UNION_##fs *_FP_UNPACK_RAW_1_P_flo \
|
|
112 = (union _FP_UNION_##fs *) (val); \
|
|
113 \
|
|
114 X##_f = _FP_UNPACK_RAW_1_P_flo->bits.frac; \
|
|
115 X##_e = _FP_UNPACK_RAW_1_P_flo->bits.exp; \
|
|
116 X##_s = _FP_UNPACK_RAW_1_P_flo->bits.sign; \
|
|
117 } \
|
|
118 while (0)
|
|
119
|
|
120 /* Repack the raw bits of a native fp value. */
|
|
121
|
|
122 #define _FP_PACK_RAW_1(fs, val, X) \
|
|
123 do \
|
|
124 { \
|
|
125 union _FP_UNION_##fs _FP_PACK_RAW_1_flo; \
|
|
126 \
|
|
127 _FP_PACK_RAW_1_flo.bits.frac = X##_f; \
|
|
128 _FP_PACK_RAW_1_flo.bits.exp = X##_e; \
|
|
129 _FP_PACK_RAW_1_flo.bits.sign = X##_s; \
|
|
130 \
|
|
131 (val) = _FP_PACK_RAW_1_flo.flt; \
|
|
132 } \
|
|
133 while (0)
|
|
134
|
|
135 #define _FP_PACK_RAW_1_P(fs, val, X) \
|
|
136 do \
|
|
137 { \
|
|
138 union _FP_UNION_##fs *_FP_PACK_RAW_1_P_flo \
|
|
139 = (union _FP_UNION_##fs *) (val); \
|
|
140 \
|
|
141 _FP_PACK_RAW_1_P_flo->bits.frac = X##_f; \
|
|
142 _FP_PACK_RAW_1_P_flo->bits.exp = X##_e; \
|
|
143 _FP_PACK_RAW_1_P_flo->bits.sign = X##_s; \
|
|
144 } \
|
|
145 while (0)
|
|
146
|
|
147
|
|
148 /* Multiplication algorithms: */
|
|
149
|
|
150 /* Basic. Assuming the host word size is >= 2*FRACBITS, we can do the
|
|
151 multiplication immediately. */
|
|
152
|
|
153 #define _FP_MUL_MEAT_DW_1_imm(wfracbits, R, X, Y) \
|
|
154 do \
|
|
155 { \
|
|
156 R##_f = X##_f * Y##_f; \
|
|
157 } \
|
|
158 while (0)
|
|
159
|
|
160 #define _FP_MUL_MEAT_1_imm(wfracbits, R, X, Y) \
|
|
161 do \
|
|
162 { \
|
|
163 _FP_MUL_MEAT_DW_1_imm ((wfracbits), R, X, Y); \
|
|
164 /* Normalize since we know where the msb of the multiplicands \
|
|
165 were (bit B), we know that the msb of the of the product is \
|
|
166 at either 2B or 2B-1. */ \
|
|
167 _FP_FRAC_SRS_1 (R, (wfracbits)-1, 2*(wfracbits)); \
|
|
168 } \
|
|
169 while (0)
|
|
170
|
|
171 /* Given a 1W * 1W => 2W primitive, do the extended multiplication. */
|
|
172
|
|
173 #define _FP_MUL_MEAT_DW_1_wide(wfracbits, R, X, Y, doit) \
|
|
174 do \
|
|
175 { \
|
|
176 doit (R##_f1, R##_f0, X##_f, Y##_f); \
|
|
177 } \
|
|
178 while (0)
|
|
179
|
|
180 #define _FP_MUL_MEAT_1_wide(wfracbits, R, X, Y, doit) \
|
|
181 do \
|
|
182 { \
|
|
183 _FP_FRAC_DECL_2 (_FP_MUL_MEAT_1_wide_Z); \
|
|
184 _FP_MUL_MEAT_DW_1_wide ((wfracbits), _FP_MUL_MEAT_1_wide_Z, \
|
|
185 X, Y, doit); \
|
|
186 /* Normalize since we know where the msb of the multiplicands \
|
|
187 were (bit B), we know that the msb of the of the product is \
|
|
188 at either 2B or 2B-1. */ \
|
|
189 _FP_FRAC_SRS_2 (_FP_MUL_MEAT_1_wide_Z, (wfracbits)-1, \
|
|
190 2*(wfracbits)); \
|
|
191 R##_f = _FP_MUL_MEAT_1_wide_Z_f0; \
|
|
192 } \
|
|
193 while (0)
|
|
194
|
|
195 /* Finally, a simple widening multiply algorithm. What fun! */
|
|
196
|
|
197 #define _FP_MUL_MEAT_DW_1_hard(wfracbits, R, X, Y) \
|
|
198 do \
|
|
199 { \
|
|
200 _FP_W_TYPE _FP_MUL_MEAT_DW_1_hard_xh, _FP_MUL_MEAT_DW_1_hard_xl; \
|
|
201 _FP_W_TYPE _FP_MUL_MEAT_DW_1_hard_yh, _FP_MUL_MEAT_DW_1_hard_yl; \
|
|
202 _FP_FRAC_DECL_2 (_FP_MUL_MEAT_DW_1_hard_a); \
|
|
203 \
|
|
204 /* Split the words in half. */ \
|
|
205 _FP_MUL_MEAT_DW_1_hard_xh = X##_f >> (_FP_W_TYPE_SIZE/2); \
|
|
206 _FP_MUL_MEAT_DW_1_hard_xl \
|
|
207 = X##_f & (((_FP_W_TYPE) 1 << (_FP_W_TYPE_SIZE/2)) - 1); \
|
|
208 _FP_MUL_MEAT_DW_1_hard_yh = Y##_f >> (_FP_W_TYPE_SIZE/2); \
|
|
209 _FP_MUL_MEAT_DW_1_hard_yl \
|
|
210 = Y##_f & (((_FP_W_TYPE) 1 << (_FP_W_TYPE_SIZE/2)) - 1); \
|
|
211 \
|
|
212 /* Multiply the pieces. */ \
|
|
213 R##_f0 = _FP_MUL_MEAT_DW_1_hard_xl * _FP_MUL_MEAT_DW_1_hard_yl; \
|
|
214 _FP_MUL_MEAT_DW_1_hard_a_f0 \
|
|
215 = _FP_MUL_MEAT_DW_1_hard_xh * _FP_MUL_MEAT_DW_1_hard_yl; \
|
|
216 _FP_MUL_MEAT_DW_1_hard_a_f1 \
|
|
217 = _FP_MUL_MEAT_DW_1_hard_xl * _FP_MUL_MEAT_DW_1_hard_yh; \
|
|
218 R##_f1 = _FP_MUL_MEAT_DW_1_hard_xh * _FP_MUL_MEAT_DW_1_hard_yh; \
|
|
219 \
|
|
220 /* Reassemble into two full words. */ \
|
|
221 if ((_FP_MUL_MEAT_DW_1_hard_a_f0 += _FP_MUL_MEAT_DW_1_hard_a_f1) \
|
|
222 < _FP_MUL_MEAT_DW_1_hard_a_f1) \
|
|
223 R##_f1 += (_FP_W_TYPE) 1 << (_FP_W_TYPE_SIZE/2); \
|
|
224 _FP_MUL_MEAT_DW_1_hard_a_f1 \
|
|
225 = _FP_MUL_MEAT_DW_1_hard_a_f0 >> (_FP_W_TYPE_SIZE/2); \
|
|
226 _FP_MUL_MEAT_DW_1_hard_a_f0 \
|
|
227 = _FP_MUL_MEAT_DW_1_hard_a_f0 << (_FP_W_TYPE_SIZE/2); \
|
|
228 _FP_FRAC_ADD_2 (R, R, _FP_MUL_MEAT_DW_1_hard_a); \
|
|
229 } \
|
|
230 while (0)
|
|
231
|
|
232 #define _FP_MUL_MEAT_1_hard(wfracbits, R, X, Y) \
|
|
233 do \
|
|
234 { \
|
|
235 _FP_FRAC_DECL_2 (_FP_MUL_MEAT_1_hard_z); \
|
|
236 _FP_MUL_MEAT_DW_1_hard ((wfracbits), \
|
|
237 _FP_MUL_MEAT_1_hard_z, X, Y); \
|
|
238 \
|
|
239 /* Normalize. */ \
|
|
240 _FP_FRAC_SRS_2 (_FP_MUL_MEAT_1_hard_z, \
|
|
241 (wfracbits) - 1, 2*(wfracbits)); \
|
|
242 R##_f = _FP_MUL_MEAT_1_hard_z_f0; \
|
|
243 } \
|
|
244 while (0)
|
|
245
|
|
246
|
|
247 /* Division algorithms: */
|
|
248
|
|
249 /* Basic. Assuming the host word size is >= 2*FRACBITS, we can do the
|
|
250 division immediately. Give this macro either _FP_DIV_HELP_imm for
|
|
251 C primitives or _FP_DIV_HELP_ldiv for the ISO function. Which you
|
|
252 choose will depend on what the compiler does with divrem4. */
|
|
253
|
|
254 #define _FP_DIV_MEAT_1_imm(fs, R, X, Y, doit) \
|
|
255 do \
|
|
256 { \
|
|
257 _FP_W_TYPE _FP_DIV_MEAT_1_imm_q, _FP_DIV_MEAT_1_imm_r; \
|
|
258 X##_f <<= (X##_f < Y##_f \
|
|
259 ? R##_e--, _FP_WFRACBITS_##fs \
|
|
260 : _FP_WFRACBITS_##fs - 1); \
|
|
261 doit (_FP_DIV_MEAT_1_imm_q, _FP_DIV_MEAT_1_imm_r, X##_f, Y##_f); \
|
|
262 R##_f = _FP_DIV_MEAT_1_imm_q | (_FP_DIV_MEAT_1_imm_r != 0); \
|
|
263 } \
|
|
264 while (0)
|
|
265
|
|
266 /* GCC's longlong.h defines a 2W / 1W => (1W,1W) primitive udiv_qrnnd
|
|
267 that may be useful in this situation. This first is for a primitive
|
|
268 that requires normalization, the second for one that does not. Look
|
|
269 for UDIV_NEEDS_NORMALIZATION to tell which your machine needs. */
|
|
270
|
|
271 #define _FP_DIV_MEAT_1_udiv_norm(fs, R, X, Y) \
|
|
272 do \
|
|
273 { \
|
|
274 _FP_W_TYPE _FP_DIV_MEAT_1_udiv_norm_nh; \
|
|
275 _FP_W_TYPE _FP_DIV_MEAT_1_udiv_norm_nl; \
|
|
276 _FP_W_TYPE _FP_DIV_MEAT_1_udiv_norm_q; \
|
|
277 _FP_W_TYPE _FP_DIV_MEAT_1_udiv_norm_r; \
|
|
278 _FP_W_TYPE _FP_DIV_MEAT_1_udiv_norm_y; \
|
|
279 \
|
|
280 /* Normalize Y -- i.e. make the most significant bit set. */ \
|
|
281 _FP_DIV_MEAT_1_udiv_norm_y = Y##_f << _FP_WFRACXBITS_##fs; \
|
|
282 \
|
|
283 /* Shift X op correspondingly high, that is, up one full word. */ \
|
|
284 if (X##_f < Y##_f) \
|
|
285 { \
|
|
286 R##_e--; \
|
|
287 _FP_DIV_MEAT_1_udiv_norm_nl = 0; \
|
|
288 _FP_DIV_MEAT_1_udiv_norm_nh = X##_f; \
|
|
289 } \
|
|
290 else \
|
|
291 { \
|
|
292 _FP_DIV_MEAT_1_udiv_norm_nl = X##_f << (_FP_W_TYPE_SIZE - 1); \
|
|
293 _FP_DIV_MEAT_1_udiv_norm_nh = X##_f >> 1; \
|
|
294 } \
|
|
295 \
|
|
296 udiv_qrnnd (_FP_DIV_MEAT_1_udiv_norm_q, \
|
|
297 _FP_DIV_MEAT_1_udiv_norm_r, \
|
|
298 _FP_DIV_MEAT_1_udiv_norm_nh, \
|
|
299 _FP_DIV_MEAT_1_udiv_norm_nl, \
|
|
300 _FP_DIV_MEAT_1_udiv_norm_y); \
|
|
301 R##_f = (_FP_DIV_MEAT_1_udiv_norm_q \
|
|
302 | (_FP_DIV_MEAT_1_udiv_norm_r != 0)); \
|
|
303 } \
|
|
304 while (0)
|
|
305
|
|
306 #define _FP_DIV_MEAT_1_udiv(fs, R, X, Y) \
|
|
307 do \
|
|
308 { \
|
|
309 _FP_W_TYPE _FP_DIV_MEAT_1_udiv_nh, _FP_DIV_MEAT_1_udiv_nl; \
|
|
310 _FP_W_TYPE _FP_DIV_MEAT_1_udiv_q, _FP_DIV_MEAT_1_udiv_r; \
|
|
311 if (X##_f < Y##_f) \
|
|
312 { \
|
|
313 R##_e--; \
|
|
314 _FP_DIV_MEAT_1_udiv_nl = X##_f << _FP_WFRACBITS_##fs; \
|
|
315 _FP_DIV_MEAT_1_udiv_nh = X##_f >> _FP_WFRACXBITS_##fs; \
|
|
316 } \
|
|
317 else \
|
|
318 { \
|
|
319 _FP_DIV_MEAT_1_udiv_nl = X##_f << (_FP_WFRACBITS_##fs - 1); \
|
|
320 _FP_DIV_MEAT_1_udiv_nh = X##_f >> (_FP_WFRACXBITS_##fs + 1); \
|
|
321 } \
|
|
322 udiv_qrnnd (_FP_DIV_MEAT_1_udiv_q, _FP_DIV_MEAT_1_udiv_r, \
|
|
323 _FP_DIV_MEAT_1_udiv_nh, _FP_DIV_MEAT_1_udiv_nl, \
|
|
324 Y##_f); \
|
|
325 R##_f = _FP_DIV_MEAT_1_udiv_q | (_FP_DIV_MEAT_1_udiv_r != 0); \
|
|
326 } \
|
|
327 while (0)
|
|
328
|
|
329
|
|
330 /* Square root algorithms:
|
|
331 We have just one right now, maybe Newton approximation
|
|
332 should be added for those machines where division is fast. */
|
|
333
|
|
334 #define _FP_SQRT_MEAT_1(R, S, T, X, q) \
|
|
335 do \
|
|
336 { \
|
|
337 while ((q) != _FP_WORK_ROUND) \
|
|
338 { \
|
|
339 T##_f = S##_f + (q); \
|
|
340 if (T##_f <= X##_f) \
|
|
341 { \
|
|
342 S##_f = T##_f + (q); \
|
|
343 X##_f -= T##_f; \
|
|
344 R##_f += (q); \
|
|
345 } \
|
|
346 _FP_FRAC_SLL_1 (X, 1); \
|
|
347 (q) >>= 1; \
|
|
348 } \
|
|
349 if (X##_f) \
|
|
350 { \
|
|
351 if (S##_f < X##_f) \
|
|
352 R##_f |= _FP_WORK_ROUND; \
|
|
353 R##_f |= _FP_WORK_STICKY; \
|
|
354 } \
|
|
355 } \
|
|
356 while (0)
|
|
357
|
|
358 /* Assembly/disassembly for converting to/from integral types.
|
|
359 No shifting or overflow handled here. */
|
|
360
|
|
361 #define _FP_FRAC_ASSEMBLE_1(r, X, rsize) ((r) = X##_f)
|
|
362 #define _FP_FRAC_DISASSEMBLE_1(X, r, rsize) (X##_f = (r))
|
|
363
|
|
364
|
|
365 /* Convert FP values between word sizes. */
|
|
366
|
|
367 #define _FP_FRAC_COPY_1_1(D, S) (D##_f = S##_f)
|
|
368
|
|
369 #endif /* !SOFT_FP_OP_1_H */
|