1 | /* Copyright (C) 2007-2024 Free Software Foundation, Inc. |
2 | |
3 | This file is part of GCC. |
4 | |
5 | GCC is free software; you can redistribute it and/or modify it under |
6 | the terms of the GNU General Public License as published by the Free |
7 | Software Foundation; either version 3, or (at your option) any later |
8 | version. |
9 | |
10 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
11 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
12 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
13 | for more details. |
14 | |
15 | Under Section 7 of GPL version 3, you are granted additional |
16 | permissions described in the GCC Runtime Library Exception, version |
17 | 3.1, as published by the Free Software Foundation. |
18 | |
19 | You should have received a copy of the GNU General Public License and |
20 | a copy of the GCC Runtime Library Exception along with this program; |
21 | see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
22 | <http://www.gnu.org/licenses/>. */ |
23 | |
24 | #undef IN_LIBGCC2 |
25 | #include "bid-dpd.h" |
26 | |
27 | /* get full 64x64bit product */ |
28 | #define __mul_64x64_to_128(P, CX, CY) \ |
29 | { \ |
30 | UINT64 CXH, CXL, CYH,CYL,PL,PH,PM,PM2; \ |
31 | CXH = (CX) >> 32; \ |
32 | CXL = (UINT32)(CX); \ |
33 | CYH = (CY) >> 32; \ |
34 | CYL = (UINT32)(CY); \ |
35 | \ |
36 | PM = CXH*CYL; \ |
37 | PH = CXH*CYH; \ |
38 | PL = CXL*CYL; \ |
39 | PM2 = CXL*CYH; \ |
40 | PH += (PM>>32); \ |
41 | PM = (UINT64)((UINT32)PM)+PM2+(PL>>32); \ |
42 | \ |
43 | (P).w[1] = PH + (PM>>32); \ |
44 | (P).w[0] = (PM<<32)+(UINT32)PL; \ |
45 | } |
46 | |
47 | /* add 64-bit value to 128-bit */ |
48 | #define __add_128_64(R128, A128, B64) \ |
49 | { \ |
50 | UINT64 R64H; \ |
51 | R64H = (A128).w[1]; \ |
52 | (R128).w[0] = (B64) + (A128).w[0]; \ |
53 | if((R128).w[0] < (B64)) R64H ++; \ |
54 | (R128).w[1] = R64H; \ |
55 | } |
56 | |
57 | /* add 128-bit value to 128-bit (assume no carry-out) */ |
58 | #define __add_128_128(R128, A128, B128) \ |
59 | { \ |
60 | UINT128 Q128; \ |
61 | Q128.w[1] = (A128).w[1]+(B128).w[1]; \ |
62 | Q128.w[0] = (B128).w[0] + (A128).w[0]; \ |
63 | if(Q128.w[0] < (B128).w[0]) Q128.w[1] ++; \ |
64 | (R128).w[1] = Q128.w[1]; \ |
65 | (R128).w[0] = Q128.w[0]; \ |
66 | } |
67 | |
68 | #define __mul_128x128_high(Q, A, B) \ |
69 | { \ |
70 | UINT128 ALBL, ALBH, AHBL, AHBH, QM, QM2; \ |
71 | \ |
72 | __mul_64x64_to_128(ALBH, (A).w[0], (B).w[1]); \ |
73 | __mul_64x64_to_128(AHBL, (B).w[0], (A).w[1]); \ |
74 | __mul_64x64_to_128(ALBL, (A).w[0], (B).w[0]); \ |
75 | __mul_64x64_to_128(AHBH, (A).w[1],(B).w[1]); \ |
76 | \ |
77 | __add_128_128(QM, ALBH, AHBL); \ |
78 | __add_128_64(QM2, QM, ALBL.w[1]); \ |
79 | __add_128_64((Q), AHBH, QM2.w[1]); \ |
80 | } |
81 | |
82 | #include "bid2dpd_dpd2bid.h" |
83 | |
84 | static const unsigned int dm103[] = |
85 | { 0, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10000, 11000 }; |
86 | |
87 | void _bid_to_dpd32 (_Decimal32 *, _Decimal32 *); |
88 | |
89 | void |
90 | _bid_to_dpd32 (_Decimal32 *pres, _Decimal32 *px) { |
91 | unsigned int sign, coefficient_x, exp, dcoeff; |
92 | unsigned int b2, b1, b0, b01, res; |
93 | _Decimal32 x = *px; |
94 | |
95 | sign = (x & 0x80000000); |
96 | if ((x & 0x60000000ul) == 0x60000000ul) { |
97 | /* special encodings */ |
98 | if ((x & 0x78000000ul) == 0x78000000ul) { |
99 | *pres = x; /* NaN or Infinity */ |
100 | return; |
101 | } |
102 | /* coefficient */ |
103 | coefficient_x = (x & 0x001ffffful) | 0x00800000ul; |
104 | if (coefficient_x >= 10000000) coefficient_x = 0; |
105 | /* get exponent */ |
106 | exp = (x >> 21) & 0xff; |
107 | } else { |
108 | exp = (x >> 23) & 0xff; |
109 | coefficient_x = (x & 0x007ffffful); |
110 | } |
111 | b01 = coefficient_x / 1000; |
112 | b2 = coefficient_x - 1000 * b01; |
113 | b0 = b01 / 1000; |
114 | b1 = b01 - 1000 * b0; |
115 | dcoeff = b2d[b2] | b2d2[b1]; |
116 | if (b0 >= 8) { /* is b0 8 or 9? */ |
117 | res = sign | ((0x600 | ((exp >> 6) << 7) | |
118 | ((b0 & 1) << 6) | (exp & 0x3f)) << 20) | dcoeff; |
119 | } else { /* else b0 is 0..7 */ |
120 | res = sign | ((((exp >> 6) << 9) | (b0 << 6) | |
121 | (exp & 0x3f)) << 20) | dcoeff; |
122 | } |
123 | *pres = res; |
124 | } |
125 | |
126 | void _dpd_to_bid32 (_Decimal32 *, _Decimal32 *); |
127 | |
128 | void |
129 | _dpd_to_bid32 (_Decimal32 *pres, _Decimal32 *px) { |
130 | unsigned int r; |
131 | unsigned int sign, exp, bcoeff; |
132 | UINT64 trailing; |
133 | unsigned int d0, d1, d2; |
134 | _Decimal32 x = *px; |
135 | |
136 | sign = (x & 0x80000000); |
137 | trailing = (x & 0x000fffff); |
138 | if ((x & 0x78000000) == 0x78000000) { |
139 | *pres = x; |
140 | return; |
141 | } |
142 | /* normal number */ |
143 | if ((x & 0x60000000) == 0x60000000) { /* G0..G1 = 11 -> d0 = 8 + G4 */ |
144 | d0 = d2b3[((x >> 26) & 1) | 8]; /* d0 = (comb & 0x0100 ? 9 : 8); */ |
145 | exp = (x >> 27) & 3; /* exp leading bits are G2..G3 */ |
146 | } else { |
147 | d0 = d2b3[(x >> 26) & 0x7]; |
148 | exp = (x >> 29) & 3; /* exp loading bits are G0..G1 */ |
149 | } |
150 | d1 = d2b2[(trailing >> 10) & 0x3ff]; |
151 | d2 = d2b[(trailing) & 0x3ff]; |
152 | bcoeff = d2 + d1 + d0; |
153 | exp = (exp << 6) + ((x >> 20) & 0x3f); |
154 | if (bcoeff < (1 << 23)) { |
155 | r = exp; |
156 | r <<= 23; |
157 | r |= (bcoeff | sign); |
158 | } else { |
159 | r = exp; |
160 | r <<= 21; |
161 | r |= (sign | 0x60000000ul); |
162 | /* add coeff, without leading bits */ |
163 | r |= (((unsigned int) bcoeff) & 0x1fffff); |
164 | } |
165 | *pres = r; |
166 | } |
167 | |
168 | void _bid_to_dpd64 (_Decimal64 *, _Decimal64 *); |
169 | |
170 | void |
171 | _bid_to_dpd64 (_Decimal64 *pres, _Decimal64 *px) { |
172 | UINT64 res; |
173 | UINT64 sign, comb, exp, B34, B01; |
174 | UINT64 d103, D61; |
175 | UINT64 b0, b2, b3, b5; |
176 | unsigned int b1, b4; |
177 | UINT64 bcoeff; |
178 | UINT64 dcoeff; |
179 | unsigned int yhi, ylo; |
180 | _Decimal64 x = *px; |
181 | |
182 | sign = (x & 0x8000000000000000ull); |
183 | comb = (x & 0x7ffc000000000000ull) >> 51; |
184 | if ((comb & 0xf00) == 0xf00) { |
185 | *pres = x; |
186 | return; |
187 | } |
188 | /* Normal number */ |
189 | if ((comb & 0xc00) == 0xc00) { /* G0..G1 = 11 -> exp is G2..G11 */ |
190 | exp = (comb) & 0x3ff; |
191 | bcoeff = (x & 0x0007ffffffffffffull) | 0x0020000000000000ull; |
192 | if (bcoeff >= 10000000000000000ull) |
193 | bcoeff = 0; |
194 | } else { |
195 | exp = (comb >> 2) & 0x3ff; |
196 | bcoeff = (x & 0x001fffffffffffffull); |
197 | } |
198 | D61 = 2305843009ull; /* Floor(2^61 / 10^9) */ |
199 | /* Multiply the binary coefficient by ceil(2^64 / 1000), and take the upper |
200 | 64-bits in order to compute a division by 1000. */ |
201 | yhi = (D61 * (UINT64)(bcoeff >> (UINT64)27)) >> (UINT64)34; |
202 | ylo = bcoeff - 1000000000ull * yhi; |
203 | if (ylo >= 1000000000) { |
204 | ylo = ylo - 1000000000; |
205 | yhi = yhi + 1; |
206 | } |
207 | d103 = 0x4189374c; |
208 | B34 = ((UINT64) ylo * d103) >> (32 + 8); |
209 | B01 = ((UINT64) yhi * d103) >> (32 + 8); |
210 | b5 = ylo - B34 * 1000; |
211 | b2 = yhi - B01 * 1000; |
212 | b3 = ((UINT64) B34 * d103) >> (32 + 8); |
213 | b0 = ((UINT64) B01 * d103) >> (32 + 8); |
214 | b4 = (unsigned int) B34 - (unsigned int) b3 *1000; |
215 | b1 = (unsigned int) B01 - (unsigned int) dm103[b0]; |
216 | dcoeff = b2d[b5] | b2d2[b4] | b2d3[b3] | b2d4[b2] | b2d5[b1]; |
217 | if (b0 >= 8) /* is b0 8 or 9? */ |
218 | res = sign | ((0x1800 | ((exp >> 8) << 9) | ((b0 & 1) << 8) | |
219 | (exp & 0xff)) << 50) | dcoeff; |
220 | else /* else b0 is 0..7 */ |
221 | res = sign | ((((exp >> 8) << 11) | (b0 << 8) | |
222 | (exp & 0xff)) << 50) | dcoeff; |
223 | *pres = res; |
224 | } |
225 | |
226 | void _dpd_to_bid64 (_Decimal64 *, _Decimal64 *); |
227 | |
228 | void |
229 | _dpd_to_bid64 (_Decimal64 *pres, _Decimal64 *px) { |
230 | UINT64 res; |
231 | UINT64 sign, comb, exp; |
232 | UINT64 trailing; |
233 | UINT64 d0, d1, d2; |
234 | unsigned int d3, d4, d5; |
235 | UINT64 bcoeff, mask; |
236 | _Decimal64 x = *px; |
237 | |
238 | sign = (x & 0x8000000000000000ull); |
239 | comb = (x & 0x7ffc000000000000ull) >> 50; |
240 | trailing = (x & 0x0003ffffffffffffull); |
241 | if ((comb & 0x1e00) == 0x1e00) { |
242 | *pres = x; |
243 | return; |
244 | } |
245 | /* normal number */ |
246 | if ((comb & 0x1800) == 0x1800) { /* G0..G1 = 11 -> d0 = 8 + G4 */ |
247 | d0 = d2b6[((comb >> 8) & 1) | 8]; /* d0 = (comb & 0x0100 ? 9 : 8); */ |
248 | exp = (comb & 0x600) >> 1; /* exp = (comb & 0x0400 ? 1 : 0) * 0x200 + |
249 | (comb & 0x0200 ? 1 : 0) * 0x100; exp leading bits are G2..G3 */ |
250 | } else { |
251 | d0 = d2b6[(comb >> 8) & 0x7]; |
252 | exp = (comb & 0x1800) >> 3; /* exp = (comb & 0x1000 ? 1 : 0) * 0x200 + |
253 | (comb & 0x0800 ? 1 : 0) * 0x100; exp loading bits are G0..G1 */ |
254 | } |
255 | d1 = d2b5[(trailing >> 40) & 0x3ff]; |
256 | d2 = d2b4[(trailing >> 30) & 0x3ff]; |
257 | d3 = d2b3[(trailing >> 20) & 0x3ff]; |
258 | d4 = d2b2[(trailing >> 10) & 0x3ff]; |
259 | d5 = d2b[(trailing) & 0x3ff]; |
260 | bcoeff = (d5 + d4 + d3) + d2 + d1 + d0; |
261 | exp += (comb & 0xff); |
262 | mask = 1; |
263 | mask <<= 53; |
264 | if (bcoeff < mask) { /* check whether coefficient fits in 10*5+3 bits */ |
265 | res = exp; |
266 | res <<= 53; |
267 | res |= (bcoeff | sign); |
268 | *pres = res; |
269 | return; |
270 | } |
271 | /* special format */ |
272 | res = (exp << 51) | (sign | 0x6000000000000000ull); |
273 | /* add coeff, without leading bits */ |
274 | mask = (mask >> 2) - 1; |
275 | bcoeff &= mask; |
276 | res |= bcoeff; |
277 | *pres = res; |
278 | } |
279 | |
280 | void _bid_to_dpd128 (_Decimal128 *, _Decimal128 *); |
281 | |
282 | void |
283 | _bid_to_dpd128 (_Decimal128 *pres, _Decimal128 *px) { |
284 | UINT128 res; |
285 | UINT128 sign; |
286 | unsigned int comb; |
287 | UINT128 bcoeff; |
288 | UINT128 dcoeff; |
289 | UINT128 BH, d1018, BT2, BT1; |
290 | UINT64 exp, BL, d109; |
291 | UINT64 d106, d103; |
292 | UINT64 k1, k2, k4, k5, k7, k8, k10, k11; |
293 | unsigned int BHH32, BLL32, BHL32, BLH32, k0, k3, k6, k9, amount; |
294 | _Decimal128 x = *px; |
295 | |
296 | sign.w[1] = (x.w[1] & 0x8000000000000000ull); |
297 | sign.w[0] = 0; |
298 | comb = (x.w[1] /*& 0x7fffc00000000000ull */ ) >> 46; |
299 | exp = 0; |
300 | if ((comb & 0x1e000) == 0x1e000) { |
301 | res = x; |
302 | } else { /* normal number */ |
303 | if ((comb & 0x18000) == 0x18000) { |
304 | /* Noncanonical significand (prepending 8 or 9 to any 110-bit |
305 | trailing significand field produces a value above 10^34). */ |
306 | exp = (comb & 0x7fff) >> 1; |
307 | bcoeff.w[1] = 0; |
308 | bcoeff.w[0] = 0; |
309 | } else { |
310 | exp = ((x.w[1] & 0x7fff000000000000ull) >> 49) & 0x3fff; |
311 | bcoeff.w[1] = (x.w[1] & 0x0001ffffffffffffull); |
312 | bcoeff.w[0] = x.w[0]; |
313 | if (bcoeff.w[1] > 0x1ed09bead87c0ull |
314 | || (bcoeff.w[1] == 0x1ed09bead87c0ull |
315 | && bcoeff.w[0] >= 0x378d8e6400000000ull)) { |
316 | bcoeff.w[1] = 0; |
317 | bcoeff.w[0] = 0; |
318 | } |
319 | } |
320 | d1018 = reciprocals10_128[18]; |
321 | __mul_128x128_high (BH, bcoeff, d1018); |
322 | amount = recip_scale[18]; |
323 | BH.w[0] = (BH.w[0] >> amount) | (BH.w[1] << (64 - amount)); |
324 | BL = bcoeff.w[0] - BH.w[0] * 1000000000000000000ull; |
325 | d109 = reciprocals10_64[9]; |
326 | __mul_64x64_to_128 (BT1, BH.w[0], d109); |
327 | BHH32 = (unsigned int) (BT1.w[1] >> short_recip_scale[9]); |
328 | BHL32 = (unsigned int) BH.w[0] - BHH32 * 1000000000; |
329 | __mul_64x64_to_128 (BT2, BL, d109); |
330 | BLH32 = (unsigned int) (BT2.w[1] >> short_recip_scale[9]); |
331 | BLL32 = (unsigned int) BL - BLH32 * 1000000000; |
332 | d106 = 0x431BDE83; |
333 | d103 = 0x4189374c; |
334 | k0 = ((UINT64) BHH32 * d106) >> (32 + 18); |
335 | BHH32 -= (unsigned int) k0 *1000000; |
336 | k1 = ((UINT64) BHH32 * d103) >> (32 + 8); |
337 | k2 = BHH32 - (unsigned int) k1 *1000; |
338 | k3 = ((UINT64) BHL32 * d106) >> (32 + 18); |
339 | BHL32 -= (unsigned int) k3 *1000000; |
340 | k4 = ((UINT64) BHL32 * d103) >> (32 + 8); |
341 | k5 = BHL32 - (unsigned int) k4 *1000; |
342 | k6 = ((UINT64) BLH32 * d106) >> (32 + 18); |
343 | BLH32 -= (unsigned int) k6 *1000000; |
344 | k7 = ((UINT64) BLH32 * d103) >> (32 + 8); |
345 | k8 = BLH32 - (unsigned int) k7 *1000; |
346 | k9 = ((UINT64) BLL32 * d106) >> (32 + 18); |
347 | BLL32 -= (unsigned int) k9 *1000000; |
348 | k10 = ((UINT64) BLL32 * d103) >> (32 + 8); |
349 | k11 = BLL32 - (unsigned int) k10 *1000; |
350 | dcoeff.w[1] = (b2d[k5] >> 4) | (b2d[k4] << 6) | (b2d[k3] << 16) | |
351 | (b2d[k2] << 26) | (b2d[k1] << 36); |
352 | dcoeff.w[0] = b2d[k11] | (b2d[k10] << 10) | (b2d[k9] << 20) | |
353 | (b2d[k8] << 30) | (b2d[k7] << 40) | (b2d[k6] << 50) | (b2d[k5] << 60); |
354 | res.w[0] = dcoeff.w[0]; |
355 | if (k0 >= 8) { |
356 | res.w[1] = sign.w[1] | ((0x18000 | ((exp >> 12) << 13) | |
357 | ((k0 & 1) << 12) | (exp & 0xfff)) << 46) | dcoeff.w[1]; |
358 | } else { |
359 | res.w[1] = sign.w[1] | ((((exp >> 12) << 15) | (k0 << 12) | |
360 | (exp & 0xfff)) << 46) | dcoeff.w[1]; |
361 | } |
362 | } |
363 | *pres = res; |
364 | } |
365 | |
366 | void _dpd_to_bid128 (_Decimal128 *, _Decimal128 *); |
367 | |
368 | void |
369 | _dpd_to_bid128 (_Decimal128 *pres, _Decimal128 *px) { |
370 | UINT128 res; |
371 | UINT128 sign; |
372 | UINT64 exp, comb; |
373 | UINT128 trailing; |
374 | UINT64 d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10, d11; |
375 | UINT128 bcoeff; |
376 | UINT64 tl, th; |
377 | _Decimal128 x = *px; |
378 | |
379 | sign.w[1] = (x.w[1] & 0x8000000000000000ull); |
380 | sign.w[0] = 0; |
381 | comb = (x.w[1] & 0x7fffc00000000000ull) >> 46; |
382 | trailing.w[1] = x.w[1]; |
383 | trailing.w[0] = x.w[0]; |
384 | if ((comb & 0x1e000) == 0x1e000) { |
385 | *pres = x; |
386 | return; |
387 | } |
388 | if ((comb & 0x18000) == 0x18000) { /* G0..G1 = 11 -> d0 = 8 + G4 */ |
389 | d0 = d2b6[8 + ((comb & 0x01000) >> 12)]; |
390 | exp = (comb & 0x06000) >> 1; /* exp leading bits are G2..G3 */ |
391 | } else { |
392 | d0 = d2b6[((comb & 0x07000) >> 12)]; |
393 | exp = (comb & 0x18000) >> 3; /* exp loading bits are G0..G1 */ |
394 | } |
395 | d11 = d2b[(trailing.w[0]) & 0x3ff]; |
396 | d10 = d2b2[(trailing.w[0] >> 10) & 0x3ff]; |
397 | d9 = d2b3[(trailing.w[0] >> 20) & 0x3ff]; |
398 | d8 = d2b4[(trailing.w[0] >> 30) & 0x3ff]; |
399 | d7 = d2b5[(trailing.w[0] >> 40) & 0x3ff]; |
400 | d6 = d2b6[(trailing.w[0] >> 50) & 0x3ff]; |
401 | d5 = d2b[(trailing.w[0] >> 60) | ((trailing.w[1] & 0x3f) << 4)]; |
402 | d4 = d2b2[(trailing.w[1] >> 6) & 0x3ff]; |
403 | d3 = d2b3[(trailing.w[1] >> 16) & 0x3ff]; |
404 | d2 = d2b4[(trailing.w[1] >> 26) & 0x3ff]; |
405 | d1 = d2b5[(trailing.w[1] >> 36) & 0x3ff]; |
406 | tl = d11 + d10 + d9 + d8 + d7 + d6; |
407 | th = d5 + d4 + d3 + d2 + d1 + d0; |
408 | __mul_64x64_to_128 (bcoeff, th, 1000000000000000000ull); |
409 | __add_128_64 (bcoeff, bcoeff, tl); |
410 | exp += (comb & 0xfff); |
411 | res.w[0] = bcoeff.w[0]; |
412 | res.w[1] = (exp << 49) | sign.w[1] | bcoeff.w[1]; |
413 | *pres = res; |
414 | } |
415 | |