1 /*
2 * Copyright (c) 2006-2010 Tensilica Inc. ALL RIGHTS RESERVED.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a copy
5 * of this software and associated documentation files (the "Software"), to deal
6 * in the Software without restriction, including without limitation the rights
7 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8 * copies of the Software, and to permit persons to whom the Software is
9 * furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
20 * THE SOFTWARE.
21 */
22
23 #ifndef __C6X_COMPAT__H
24 #define __C6X_COMPAT__H
25
26 /* Unimplemented functions _gmpy, _gmpy4, _xormpy, _lssub, _cmpy, _cmpyr,
27 _cmpyr1, _ddotpl2r, _ddotph2r */
28
29
30 typedef long long C6X_COMPAT_LONG40;
31
32
33 #define _memd8(a) (*((double*)(a)))
34 #define _memd8_const(a) (*((const double*)(a)))
35
36 #define _amemd8(a) (*((double*)(a)))
37 #define _amemd8_const(a) (*((const double*)(a)))
38
39 #define _mem8(a) (*((unsigned long long*)(a)))
40 #define _mem8_const(a) (*((const unsigned long long*)(a)))
41
42 #define _mem4(a) (*((unsigned*)(a)))
43 #define _mem4_const(a) (*((const unsigned*)(a)))
44 #define _amem4_const(a) (*((const unsigned*)(a)))
45
46 /* NOTE: To emulate a C6X properly you should define global variables
47 for your Xtensa with these names. Some of the emulation routines
48 will set these values. */
49
50 extern int _carry;
51 extern int _overflow;
52
53 // Utility routines
54
55
56 #define TESTBIT(x,n) (((x) >> (n)) & 1)
57
58 #define NSA_BITS 32
59
norm_shift_amt_U_and_non_U(int is_signed,int inp)60 static inline unsigned int norm_shift_amt_U_and_non_U(int is_signed, int inp) {
61 int j=0, k=0;
62 int x=inp;
63 if (is_signed) {
64 /* Invert signed val if negative */
65 x= TESTBIT(x,(NSA_BITS-1))? ~x: x;
66 x= (x&1)|(x<<1); /* Shift up to return count-1 */
67 if (x ==0)
68 return NSA_BITS-1;
69 }
70 if (x ==0)
71 return NSA_BITS;
72 /* Now count leading zeros */
73 for (j=0, k=NSA_BITS-1; k>=0; j++, k--) {
74 if (TESTBIT(x,k))
75 return j;
76 }
77 return NSA_BITS;
78 }
79
80
81
82 static inline long long
orig_L40_set(long long L40_var1)83 orig_L40_set( long long L40_var1) {
84 long long L40_var_out;
85
86 L40_var_out = L40_var1 & 0x000000ffffffffffLL;
87
88 if( L40_var1 & 0x8000000000LL)
89 L40_var_out = L40_var_out | 0xffffff0000000000LL;
90
91 return( L40_var_out);
92 }
93
94
95
96 static inline signed long long
util_saturate_n_no_state(signed long long t,int n)97 util_saturate_n_no_state(signed long long t, int n)
98 {
99 signed long long maxv, minv;
100 maxv = (1LL << (n-1)) - 1;
101 minv = (-1LL << (n-1));
102 if (t > maxv) {
103 t = maxv;
104 } else if (t < minv) {
105 t = minv;
106 }
107 return t;
108 }
109
110
111 static inline signed long long
util_saturate_n_sgn(signed long long t,int n)112 util_saturate_n_sgn(signed long long t, int n)
113 {
114 signed long long result;
115 signed long long maxv, minv;
116 maxv = (1LL << (n-1)) - 1;
117 minv = (-1LL << (n-1));
118 if (t > 0) {
119 result = maxv;
120 _overflow = 1;
121 } else if (t < 0) {
122 result = minv;
123 _overflow = 1;
124 } else {
125 result = 0;
126 }
127 return result;
128 }
129
130
131
132
133 /* well-behaved signed shift right (left on negative) with
134 saturation */
135 static inline signed long long
util_shift_right_saturate_n(signed long long t,int shval,int n)136 util_shift_right_saturate_n(signed long long t, int shval, int n)
137 {
138 /* n should be <= 62 */
139 long long result;
140
141 signed long long mask;
142 int actual_shift = shval;
143 long long shft = actual_shift > 0 ? actual_shift : -actual_shift;
144
145 if (t == 0 || actual_shift == 0)
146 return t;
147
148 if (actual_shift >= n) {
149 return (t < 0) ? -1 : 0;
150 }
151 if (actual_shift <= -n) {
152 return util_saturate_n_sgn(t, n);
153 }
154 if (actual_shift > 0) {
155 return t >> actual_shift;
156 }
157 /* actual_shift < 0. Check for saturation after shift. */
158 mask = (-1LL << (n-shft-1));
159 if (t > 0 && ((mask & t) != 0)) {
160 return util_saturate_n_sgn(t, n);
161 }
162 if (t < 0 && ((mask & t) != mask)) {
163 return util_saturate_n_sgn(t, n);
164 }
165 result = t << shft;
166
167 return result;
168 }
169
170
171 /* Implemented c6x standard C compatibility functions (alphabetical
172 order) */
173
174
_abs(int src1)175 static inline int _abs(int src1) {
176 if ((unsigned) src1 == (unsigned) 0x80000000) {
177 return 0x7fffffff;
178 }
179 return abs(src1);
180 }
181
182
_abs2(int src1)183 static inline int _abs2(int src1) {
184 short s1[2],r[2];
185 int result;
186 *((int*)s1) = src1;
187 if ((unsigned short) s1[1] == (unsigned short) 0x8000) r[1] = 0x7fff;
188 else r[1] = abs(s1[1]);
189 if ((unsigned short) s1[0] == (unsigned short) 0x8000) r[0] = 0x7fff;
190 else r[0] = abs(s1[0]);
191 result = *(int*)r;
192 return result;
193 }
194
195
196
197
_add2(int src1,int src2)198 static inline int _add2(int src1, int src2) {
199 short s1[2], s2[2], r[2];
200 int result;
201 *((int*)s1) = src1;
202 *((int*)s2) = src2;
203 r[0] = s1[0] + s2[0];
204 r[1] = s1[1] + s2[1];
205 result = *(int*)r;
206 return result;
207 }
208
_add4(int src1,int src2)209 static inline int _add4(int src1, int src2) {
210 char c1[4], c2[4], r[4];
211 int result;
212 *((int*)c1) = src1;
213 *((int*)c2) = src2;
214 r[0] = c1[0] + c2[0];
215 r[1] = c1[1] + c2[1];
216 r[2] = c1[2] + c2[2];
217 r[3] = c1[3] + c2[3];
218 result = *(int*)r;
219 return result;
220 }
221
222
223
_addsub(unsigned int src1,unsigned int src2)224 static inline long long _addsub(unsigned int src1, unsigned int src2)
225 {
226
227 int res_lo;
228 int res_hi;
229
230 res_hi = src1+src2;
231 res_lo = src1-src2;
232 return (((unsigned long long) res_hi) << 32) | ((unsigned int) res_lo) ;
233 }
234
235
_addsub2(unsigned int src1,unsigned int src2)236 static inline long long _addsub2(unsigned int src1, unsigned int src2)
237 {
238 short s1[2], s2[2], ra[2], rs[2];
239 int res_lo;
240 int res_hi;
241
242 *((int*)s1) = src1;
243 *((int*)s2) = src2;
244 ra[0] = s1[0] + s2[0];
245 ra[1] = s1[1] + s2[1];
246 rs[0] = s1[0] - s2[0];
247 rs[1] = s1[1] - s2[1];
248
249 res_hi = *(int*)ra;
250 res_lo = *(int*)rs;
251 return (((unsigned long long) res_hi) << 32) | ((unsigned int) res_lo) ;
252 }
253
254
_avg2(int src1,int src2)255 static inline int _avg2(int src1, int src2) {
256 int low = (((int)1 + (short) src1 + (short) src2) >> 1) & 0XFFFF;
257 int high1 = src1 >> 16;
258 int high2 = src2 >> 16;
259 int high = ((high1 + high2 + 1) >> 1)<< 16;
260 return high | low;
261 }
262
263
264
_avgu4(unsigned int src1,unsigned int src2)265 static inline unsigned int _avgu4(unsigned int src1, unsigned int src2) {
266 unsigned int res0 = ((src1 & 0xFF) + (src2 & 0xFF) + 1) >> 1;
267 unsigned int res1 = (((src1 & 0xFF00) >> 8) + ((src2 & 0xFF00) >> 8) + 1) >> 1;
268 unsigned int res2 = (((src1 & 0xFF0000) >> 16) + ((src2 & 0xFF0000) >> 16) + 1) >> 1;
269 unsigned int res3 = (((src1 & 0xFF000000) >> 24) + ((src2 & 0xFF000000) >> 24) + 1) >> 1;
270 return (res3 << 24) | (res2 << 16) | (res1 << 8) | res0;
271 }
272
273
TEN_popc(unsigned char b)274 static inline int TEN_popc (unsigned char b)
275 {
276 int i, result = 0;
277 for (i = 0; i < 8; i++){
278 if (b & 0x1)
279 result++;
280 b >>= 1;
281 }
282 return result;
283 }
284
_bitc4(unsigned int src1)285 static inline unsigned int _bitc4(unsigned int src1)
286 {
287 unsigned int res0 = TEN_popc(src1 & 0xFF);
288 unsigned int res1 = TEN_popc((src1 & 0xFF00) >> 8);
289 unsigned int res2 = TEN_popc((src1 & 0xFF0000) >> 16);
290 unsigned int res3 = TEN_popc((src1 & 0xFF000000) >> 24);
291 return (res3 << 24) | (res2 << 16) | (res1 << 8) | res0;
292 }
293
_bitr(unsigned int src)294 static inline unsigned int _bitr(unsigned int src) {
295 int i;
296 unsigned r = 0;
297 for (i = 0; i< 32; ++i) {
298 r = r | (((src >> i) & 1)<<(31-i));
299 }
300 return r;
301 }
302
303
_clr(unsigned int src2,int csta,int cstb)304 static inline unsigned int _clr(unsigned int src2, int csta, int cstb)
305 {
306 csta &= 0x1f;
307 cstb &= 0x1f;
308 if (csta > cstb)
309 return src2;
310 else {
311 unsigned int mask = (((1 << (cstb - csta)) << 1) - 1) << csta;
312 return src2 & (~mask);
313 }
314 }
315
_clrr(unsigned int src2,int src1)316 static inline unsigned int _clrr(unsigned int src2, int src1)
317 {
318 unsigned int csta = (src1 >> 5) & 0x1f;
319 unsigned int cstb = src1 & 0x1f;
320 if (csta > cstb)
321 return src2;
322 else {
323 unsigned int mask = (((1 << (cstb - csta)) << 1) - 1) << csta;
324 return src2 & (~mask);
325 }
326 }
327
328
329
330
_cmpeq2(int src1,int src2)331 static inline int _cmpeq2(int src1, int src2) {
332 short s1[2], s2[2];
333 int r0, r1;
334 int result;
335 *((int*)s1) = src1;
336 *((int*)s2) = src2;
337 r0 = s1[0] == s2[0] ? 1 : 0;
338 r1 = s1[1] == s2[1] ? 1 : 0;
339 result = (r1 << 1) | r0;
340 return result;
341 }
342
_cmpeq4(int src1,int src2)343 static inline int _cmpeq4(int src1, int src2) {
344 char s1[4], s2[4];
345 int r0, r1, r2, r3;
346 int result;
347 *((int*)s1) = src1;
348 *((int*)s2) = src2;
349 r0 = s1[0] == s2[0] ? 1 : 0;
350 r1 = s1[1] == s2[1] ? 1 : 0;
351 r2 = s1[2] == s2[2] ? 1 : 0;
352 r3 = s1[3] == s2[3] ? 1 : 0;
353 result = (r3 << 3) | (r2 << 2) | (r1 << 1) | r0;
354 return result;
355 }
356
357
_cmpgt2(int src1,int src2)358 static inline int _cmpgt2(int src1, int src2) {
359 short s1[2], s2[2];
360 int r1, r0;
361 int result;
362 *((int*)s1) = src1;
363 *((int*)s2) = src2;
364 r0 = s1[0] > s2[0] ? 1 : 0;
365 r1 = s1[1] > s2[1] ? 1 : 0;
366 result = (r1<<1) | r0;
367 return result;
368 }
369
370
_cmpgtu4(unsigned int src1,unsigned int src2)371 static inline unsigned int _cmpgtu4(unsigned int src1, unsigned int src2) {
372 unsigned int s1_0 = (src1 & 0xFF);
373 unsigned int s1_1 = (src1 & 0xFF00) >> 8;
374 unsigned int s1_2 = (src1 & 0xFF0000) >> 16;
375 unsigned int s1_3 = (src1 & 0xFF000000) >> 24;
376
377 unsigned int s2_0 = (src2 & 0xFF);
378 unsigned int s2_1 = (src2 & 0xFF00) >> 8;
379 unsigned int s2_2 = (src2 & 0xFF0000) >> 16;
380 unsigned int s2_3 = (src2 & 0xFF000000) >> 24;
381
382 unsigned int result = 0;
383
384 if (s1_0 > s2_0)
385 result |= 0x1;
386
387 if (s1_1 > s2_1)
388 result |= 0x2;
389
390 if (s1_2 > s2_2)
391 result |= 0x4;
392
393 if (s1_3 > s2_3)
394 result |= 0x8;
395
396 return result;
397 }
398
399
400
401
_ddotp4(unsigned int src1,unsigned int src2)402 static inline long long _ddotp4(unsigned int src1, unsigned int src2) {
403 unsigned int res0, res1;
404 short s1_0 = (src1 & 0xffff);
405 short s1_1 = (src1 & 0xfff0000) >> 16;
406
407 unsigned short s2_0 = (src2 & 0xff);
408 unsigned short s2_1 = (src2 & 0xff00) >> 8;
409 unsigned short s2_2 = (src2 & 0xff0000) >> 16;
410 unsigned short s2_3 = (src2 & 0xff000000) >> 24;
411
412 res0 = ((int)s1_0) * s2_0 + ((int)s1_1) * s2_1;
413 res1 = ((int)s1_0) * s2_2 + ((int)s1_1) * s2_3;
414
415 return (res1 << 16) | res0;
416 }
417
418
_ddotph2(long long src1_o_src1_e,unsigned int src2)419 static inline long long _ddotph2(long long src1_o_src1_e, unsigned int src2)
420 {
421
422 unsigned int src1_o = src1_o_src1_e >> 32;
423 unsigned int src1_e = src1_o_src1_e & 0xFFFFFFFF;
424 short ls1_o = src1_o & 0XFFFF;
425 short hs1_o = src1_o >> 16;
426 // short ls1_e = src1_e & 0XFFFF;
427 short hs1_e = src1_e >> 16;
428 short ls2 = src2 & 0XFFFF;
429 short hs2 = src2 >> 16;
430
431 unsigned long long res_hi = ls2 * ls1_o + hs2 * hs1_o;
432 unsigned int res_lo = ls1_o * hs2 + hs1_e * ls2;
433 return (res_hi << 32) | res_lo;
434 }
435
436
_ddotpl2(long long src1_o_src1_e,unsigned int src2)437 static inline long long _ddotpl2(long long src1_o_src1_e, unsigned int src2)
438 {
439 unsigned int src1_o = src1_o_src1_e >> 32;
440 unsigned int src1_e = src1_o_src1_e & 0xFFFFFFFF;
441 short ls1_o = src1_o & 0XFFFF;
442 // short hs1_o = src1_o >> 16;
443 short ls1_e = src1_e & 0XFFFF;
444 short hs1_e = src1_e >> 16;
445 short ls2 = src2 & 0XFFFF;
446 short hs2 = src2 >> 16;
447
448 unsigned long long res_hi = ls2 * hs1_e + hs2 * ls1_o;
449 unsigned res_lo = hs1_e * hs2 + ls1_e * ls2;
450 return (res_hi << 32) | res_lo;
451 }
452
453
_deal(unsigned int src)454 static inline unsigned int _deal(unsigned int src)
455 {
456 int i;
457 unsigned short lo = 0, hi = 0;
458 for (i = 0; i < 32; i+= 2) {
459 lo >>= 1;
460 lo |= (src & 0x1) << 15;
461 src >>= 1;
462 hi >>= 1;
463 hi |= (src & 0x1) << 15;
464 src >>= 1;
465 }
466 return (hi << 16) | lo;
467 }
468
469
_dmv(unsigned int src1,unsigned int src2)470 static inline long long _dmv(unsigned int src1, unsigned int src2)
471 {
472 return (((long long) src1) << 32) | src2;
473 }
474
475
_dotpn2(int src1,int src2)476 static inline int _dotpn2(int src1, int src2) {
477 short int s1_h = src1>>16;
478 short int s1_l = src1;
479 short int s2_h = src2>>16;
480 short int s2_l = src2;
481 return s1_h * s2_h - s1_l * s2_l;
482 }
483
484
_dotp2(int src1,int src2)485 static inline int _dotp2(int src1, int src2) {
486 short int s1_h = src1>>16;
487 short int s1_l = src1;
488 short int s2_h = src2>>16;
489 short int s2_l = src2;
490 return s1_h * s2_h + s1_l * s2_l;
491 }
492
493
494
_dotpnrsu2(int src1,unsigned int src2)495 static inline int _dotpnrsu2(int src1, unsigned int src2)
496 {
497 short ls1 = src1 & 0XFFFF;
498 unsigned short ls2 = src2 & 0XFFFF;
499 short hs1 = src1 >> 16;
500 unsigned short hs2 = src2 >> 16;
501
502 int result = (((long long) (int)(hs1 * hs2)) - ((long long) (int)(ls1 * ls2)) + (1 << 15)) >> 16;
503 return result;
504 }
505
506
507
_dotprsu2(int src1,unsigned int src2)508 static inline int _dotprsu2(int src1, unsigned int src2) {
509 short ls1 = src1 & 0XFFFF;
510 unsigned short ls2 = (src2 & 0XFFFF);
511 short hs1 = src1 >> 16;
512 unsigned short hs2 = (src2 >> 16);
513
514 int result = (((long long) (int) (ls1 * ls2)) + ((long long) (int) (hs1 * hs2)) + (1LL << 15)) >> 16;
515 return result;
516 }
517
518
519
520
521
522
523
_dotpsu4(int src1,unsigned int src2)524 static inline int _dotpsu4(int src1, unsigned int src2) {
525 int result;
526 signed char s1_0 = (src1 & 0xff);
527 signed char s1_1 = (src1 & 0xff00) >> 8;
528 signed char s1_2 = (src1 & 0xff0000) >> 16;
529 signed char s1_3 = (src1 & 0xff000000) >> 24;
530
531 unsigned int s2_0 = (src2 & 0xff);
532 unsigned int s2_1 = (src2 & 0xff00) >> 8;
533 unsigned int s2_2 = (src2 & 0xff0000) >> 16;
534 unsigned int s2_3 = (src2 & 0xff000000) >> 24;
535
536 result = s1_0 * s2_0 + s1_1 * s2_1 + s1_2 * s2_2 + s1_3 * s2_3;
537 return result;
538 }
539
540
_dotpu4(unsigned int src1,unsigned int src2)541 static inline unsigned int _dotpu4(unsigned int src1, unsigned int src2) {
542 unsigned char v1_0 = src1 & 0xff;
543 unsigned char v1_1 = (src1>>8) & 0xff;
544 unsigned char v1_2 = (src1>>16) & 0xff;
545 unsigned char v1_3 = (src1>>24) & 0xff;
546
547 unsigned char v2_0 = src2 & 0xff;
548 unsigned char v2_1 = (src2>>8) & 0xff;
549 unsigned char v2_2 = (src2>>16) & 0xff;
550 unsigned char v2_3 = (src2>>24) & 0xff;
551
552 unsigned v = v1_0 * v2_0 + v1_1 * v2_1 + v1_2 * v2_2 + v1_3 * v2_3;
553 return v;
554 }
555
556
_dpack2(unsigned int src1,unsigned int src2)557 static inline long long _dpack2(unsigned int src1, unsigned int src2){
558 unsigned short s1[2], s2[2];
559 *((int*)s1) = src1;
560 *((int*)s2) = src2;
561 return ((unsigned long long) s1[1] << 48) | ((unsigned long long) s2[1] << 32) | ((unsigned long long) s1[0] << 16) | ((unsigned long long) s2[0]);
562 }
563
564
_dpackx2(unsigned int src1,unsigned int src2)565 static inline long long _dpackx2(unsigned int src1, unsigned int src2){
566 unsigned short s1[2], s2[2];
567 *((int*)s1) = src1;
568 *((int*)s2) = src2;
569 return ((unsigned long long) s2[0] << 48) | ((unsigned long long) s1[1] << 32) | ((unsigned long long) s1[0] << 16) | ((unsigned long long) s2[1]);
570 }
571
_ext(int src2,unsigned int csta,unsigned int cstb)572 static inline int _ext(int src2, unsigned int csta, unsigned int cstb)
573 {
574 return (src2 << csta) >> cstb;
575 }
576
_extr(int src2,int src1)577 static inline int _extr(int src2, int src1)
578 {
579 unsigned int csta = (src1 >> 5) & 0x1f;
580 unsigned int cstb = src1 & 0x1f;
581 return (src2 << csta) >> cstb;
582 }
583
_extu(unsigned int src2,unsigned int csta,unsigned int cstb)584 static inline unsigned int _extu(unsigned int src2, unsigned int csta, unsigned int cstb)
585 {
586 return (src2 << csta) >> cstb;
587 }
588
_extur(unsigned int src2,int src1)589 static inline unsigned int _extur(unsigned int src2, int src1)
590 {
591 unsigned int csta = (src1 >> 5) & 0x1f;
592 unsigned int cstb = src1 & 0x1f;
593 return (src2 << csta) >> cstb;
594 }
595
596
_hi(double src)597 static inline unsigned long long _hi(double src) {
598 unsigned long long v;
599 *(double*)&v = src;
600 return v>>32;
601 }
602
_hill(long long src)603 static inline unsigned int _hill (long long src)
604 {
605 return (unsigned int) (src >> 32);
606 }
607
608
609
_itod(unsigned hi,unsigned lo)610 static inline double _itod(unsigned hi, unsigned lo) {
611 double v;
612 unsigned long long ll = ((((unsigned long long)(hi))<<32) | (unsigned long long)((unsigned)lo));
613 *((unsigned long long *)&v) = ll;
614 return v;
615 }
616
617
_itoll(unsigned int src2,unsigned int src1)618 static inline long long _itoll(unsigned int src2, unsigned int src1)
619 {
620 return (((long long) src2) << 32) | src1;
621 }
622
623
_labs(C6X_COMPAT_LONG40 src2)624 static inline C6X_COMPAT_LONG40 _labs(C6X_COMPAT_LONG40 src2)
625 {
626 long long maxv = (1LL << (40 -1)) - 1;
627 long long minv = (-1LL << (40 - 1));
628 C6X_COMPAT_LONG40 lres = orig_L40_set(src2);
629
630 lres = lres < 0 ? -lres : lres;
631 if (lres > maxv) lres = maxv;
632 else if (lres < minv) lres = minv;
633
634 return lres;
635 }
636
637
_ldotp2(int src1,int src2)638 static inline C6X_COMPAT_LONG40 _ldotp2(int src1, int src2) {
639 return (C6X_COMPAT_LONG40) _dotp2(src1, src2);
640 }
641
642
_lmbd(unsigned int src1,unsigned int src2)643 static inline unsigned int _lmbd(unsigned int src1, unsigned int src2)
644 {
645 return norm_shift_amt_U_and_non_U(0,(((int) (src1 << 31)) >> 31) ^ (~src2));
646 }
647
648
_lnorm(C6X_COMPAT_LONG40 src2)649 static inline unsigned int _lnorm(C6X_COMPAT_LONG40 src2) {
650 if (src2 == 0)
651 return 39;
652 else {
653 int hi = (int)(src2 >> 32);
654 int lo = (int)src2;
655
656
657 long long temp = (unsigned long long)(unsigned)lo | (unsigned long long)hi << 32;
658 temp = orig_L40_set(temp);
659
660 if (temp == 0) return 0;
661 int cnt = 0;
662 while (((temp >> 39) & 1) == ((temp >> 38) & 1)) {
663 temp <<= 1;
664 cnt++;
665 }
666 return cnt;
667 }
668 }
669
670
_lo(double src)671 static inline unsigned long long _lo(double src) {
672 unsigned long long v;
673 *(double*)&v = src;
674 return v;
675 }
676
677
_loll(long long src)678 static inline unsigned int _loll (long long src)
679 {
680 return (unsigned int) src;
681 }
682
683
_lsadd(int src1,C6X_COMPAT_LONG40 src2)684 static inline C6X_COMPAT_LONG40 _lsadd(int src1, C6X_COMPAT_LONG40 src2)
685 {
686 long long maxv = (1LL << (40 -1)) - 1;
687 long long minv = (-1LL << (40 - 1));
688 int hi = (int)(src2 >> 32);
689 int lo = (int)src2;
690 long long src2_int = (unsigned long long)(unsigned)lo | (unsigned long long)hi << 32;
691
692
693 long long src2_int2 = orig_L40_set(src2_int);
694
695 long long res = src1 + src2_int2;
696
697 if (res > maxv) {
698 res = maxv;
699 _overflow = 1;
700 }
701 else if (res < minv) {
702 res = minv;
703 _overflow = 1;
704 }
705
706 long long res2 = orig_L40_set(res);
707
708 res2 = (signed char)(res2 >> 32);
709
710 C6X_COMPAT_LONG40 lres = (((C6X_COMPAT_LONG40) res2) << 32) | ((unsigned int)res);
711 return lres;
712 }
713
714
715
_max2(int src1,int src2)716 static inline int _max2 (int src1, int src2) {
717 short s1[2], s2[2], r[2];
718 int result;
719 *((int*)s1) = src1;
720 *((int*)s2) = src2;
721 r[0] = s1[0] > s2[0] ? s1[0] : s2[0];
722 r[1] = s1[1] > s2[1] ? s1[1] : s2[1];
723 result = *(int*)r;
724 return result;
725 }
726
727
728
729
730
731
_maxu4(unsigned int src1,unsigned int src2)732 static inline unsigned int _maxu4(unsigned int src1, unsigned int src2) {
733 unsigned int res0, res1, res2, res3;
734 unsigned int s1_0 = res0 = (src1 & 0xFF);
735 unsigned int s1_1 = res1 = (src1 & 0xFF00) >> 8;
736 unsigned int s1_2 = res2 = (src1 & 0xFF0000) >> 16;
737 unsigned int s1_3 = res3 = (src1 & 0xFF000000) >> 24;
738
739 unsigned int s2_0 = (src2 & 0xFF);
740 unsigned int s2_1 = (src2 & 0xFF00) >> 8;
741 unsigned int s2_2 = (src2 & 0xFF0000) >> 16;
742 unsigned int s2_3 = (src2 & 0xFF000000) >> 24;
743
744 // unsigned int res = 0;
745
746 if (s1_0 < s2_0)
747 res0 = s2_0;
748
749 if (s1_1 < s2_1)
750 res1 = s2_1;
751
752 if (s1_2 < s2_2)
753 res2 = s2_2;
754
755 if (s1_3 < s2_3)
756 res3 = s2_3;
757
758 return (res3 << 24) | (res2 << 16) | (res1 << 8) | res0;
759
760
761 }
762
_min2(int src1,int src2)763 static inline int _min2(int src1, int src2) {
764 short s1[2], s2[2], r[2];
765 int result;
766 *((int*)s1) = src1;
767 *((int*)s2) = src2;
768 r[0] = s1[0] < s2[0] ? s1[0] : s2[0];
769 r[1] = s1[1] < s2[1] ? s1[1] : s2[1];
770 result = *(int*)r;
771 return result;
772 }
773
774
_minu4(unsigned int src1,unsigned int src2)775 static inline unsigned int _minu4(unsigned int src1, unsigned int src2) {
776 unsigned int res0, res1, res2, res3;
777 unsigned int s1_0 = res0 = (src1 & 0xFF);
778 unsigned int s1_1 = res1 = (src1 & 0xFF00) >> 8;
779 unsigned int s1_2 = res2 = (src1 & 0xFF0000) >> 16;
780 unsigned int s1_3 = res3 = (src1 & 0xFF000000) >> 24;
781
782 unsigned int s2_0 = (src2 & 0xFF);
783 unsigned int s2_1 = (src2 & 0xFF00) >> 8;
784 unsigned int s2_2 = (src2 & 0xFF0000) >> 16;
785 unsigned int s2_3 = (src2 & 0xFF000000) >> 24;
786
787 // unsigned int res = 0;
788
789 if (s1_0 > s2_0)
790 res0 = s2_0;
791
792 if (s1_1 > s2_1)
793 res1 = s2_1;
794
795 if (s1_2 > s2_2)
796 res2 = s2_2;
797
798 if (s1_3 > s2_3)
799 res3 = s2_3;
800
801 return (res3 << 24) | (res2 << 16) | (res1 << 8) | res0;
802 }
803
804
_mpy(int src1,int src2)805 static inline int _mpy(int src1, int src2) {
806 return (short) src1 * (short) src2;
807 }
808
809
_mpyh(int src1,int src2)810 static inline int _mpyh(int src1, int src2) {
811 return (short) (src1 >> 16) * (short) (src2 >> 16);
812 }
813
814
_mpyhill(int src1,int src2)815 static inline long long _mpyhill (int src1, int src2)
816 {
817 short s1 = src1 >> 16;
818 return ((long long) src2) * s1;
819 }
820
_mpyhir(int src1,int src2)821 static inline int _mpyhir(int src1, int src2)
822 {
823 short s1 = src1 >> 16;
824 long long result = ((long long) src2) * s1 + (1 << 14);
825 result >>= 15;
826 return result;
827 }
828
829
_mpyhl(int src1,int src2)830 static inline int _mpyhl(int src1, int src2) {
831 return (short) (src1 >> 16) * (short) (src2);
832 }
833
_mpyhlu(unsigned int src1,unsigned int src2)834 static inline unsigned int _mpyhlu(unsigned int src1, unsigned int src2) {
835 return (unsigned short) (src1 >> 16) * (unsigned short) (src2);
836 }
837
_mpyhslu(int src1,unsigned int src2)838 static inline int _mpyhslu(int src1, unsigned int src2) {
839 return (short) (src1 >> 16) * (unsigned short) src2;
840 }
841
842
_mpyhsu(int src1,unsigned int src2)843 static inline int _mpyhsu(int src1, unsigned int src2) {
844 return (short) (src1 >>16) * (unsigned short) (src2 >>16);
845 }
846
847
_mpyhu(unsigned int src1,unsigned int src2)848 static inline unsigned int _mpyhu(unsigned int src1, unsigned int src2) {
849 return (unsigned short) (src1 >>16) * (unsigned short) (src2 >> 16);
850 }
851
852
_mpyhuls(unsigned int src1,int src2)853 static inline int _mpyhuls(unsigned int src1, int src2) {
854 return (unsigned short) (src1 >>16) * (signed short) (src2);
855 }
856
857
_mpyhus(unsigned int src1,int src2)858 static inline int _mpyhus(unsigned int src1, int src2) {
859 return (unsigned short) (src1 >> 16) * (short) (src2 >>16);
860 }
861
862
863
_mpyidll(int src1,int src2)864 static inline long long _mpyidll (int src1, int src2)
865 {
866 return (long long) src1 * src2;
867 }
868
869
_mpylh(int src1,int src2)870 static inline int _mpylh(int src1, int src2) {
871 return (signed short) (src1 & 0xffff) * (signed short) (src2 >> 16);
872 }
873
_mpylhu(unsigned int src1,unsigned int src2)874 static inline unsigned int _mpylhu(unsigned int src1, unsigned int src2) {
875 return (unsigned short) src1 * (unsigned short) (src2 >> 16);
876 }
877
878
_mpylill(int src1,int src2)879 static inline long long _mpylill (int src1, int src2)
880 {
881 return ((long long) src2) * ((short)src1);
882 }
883
884
885
_mpylir(int src1,int src2)886 static inline int _mpylir(int src1, int src2)
887 {
888 short s1 = src1;
889 long long result = ((long long) src2) * s1 + (1 << 14);
890 result >>= 15;
891 return result;
892 }
893
894
_mpylshu(int src1,unsigned int src2)895 static inline int _mpylshu(int src1, unsigned int src2) {
896 return (short) src1 * (unsigned short) (src2 >> 16);
897 }
898
899
_mpyluhs(unsigned int src1,int src2)900 static inline int _mpyluhs(unsigned int src1, int src2) {
901 return (unsigned short) src1 * (short) (src2 >> 16);
902 }
903
904
905
_mpysu(int src1,unsigned int src2)906 static inline int _mpysu(int src1, unsigned int src2) {
907 return (short) src1 * (unsigned short) src2;
908 }
909
910
911
_mpysu4ll(int src1,unsigned int src2)912 static inline long long _mpysu4ll (int src1, unsigned int src2) {
913 unsigned short res0, res1, res2, res3;
914 signed char s1_0 = (src1 & 0xff);
915 signed char s1_1 = (src1 & 0xff00) >> 8;
916 signed char s1_2 = (src1 & 0xff0000) >> 16;
917 signed char s1_3 = (src1 & 0xff000000) >> 24;
918
919 unsigned short s2_0 = (src2 & 0xff);
920 unsigned short s2_1 = (src2 & 0xff00) >> 8;
921 unsigned short s2_2 = (src2 & 0xff0000) >> 16;
922 unsigned short s2_3 = (src2 & 0xff000000) >> 24;
923
924 res0 = s1_0 * s2_0;
925 res1 = s1_1 * s2_1;
926 res2 = s1_2 * s2_2;
927 res3 = s1_3 * s2_3;
928
929 return (((unsigned long long) res3) << 48)
930 | (((unsigned long long) res2) << 32)
931 | (((unsigned long long) res1) << 16)
932 | res0;
933 }
934
_mpyu(unsigned int src1,unsigned int src2)935 static inline unsigned int _mpyu(unsigned int src1, unsigned int src2) {
936 unsigned v = (unsigned short)src1 * (unsigned short)src2;
937 return v;
938 }
939
_mpyus(unsigned int src1,int src2)940 static inline int _mpyus(unsigned int src1, int src2) {
941 return (unsigned short) src1 * (short) src2;
942 }
943
_mpyu4ll(unsigned int src1,unsigned int src2)944 static inline long long _mpyu4ll (unsigned int src1, unsigned int src2) {
945 unsigned short res0, res1, res2, res3;
946 unsigned char s1_0 = (src1 & 0xff);
947 unsigned char s1_1 = (src1 & 0xff00) >> 8;
948 unsigned char s1_2 = (src1 & 0xff0000) >> 16;
949 unsigned char s1_3 = (src1 & 0xff000000) >> 24;
950
951 unsigned short s2_0 = (src2 & 0xff);
952 unsigned short s2_1 = (src2 & 0xff00) >> 8;
953 unsigned short s2_2 = (src2 & 0xff0000) >> 16;
954 unsigned short s2_3 = (src2 & 0xff000000) >> 24;
955
956 res0 = s1_0 * s2_0;
957 res1 = s1_1 * s2_1;
958 res2 = s1_2 * s2_2;
959 res3 = s1_3 * s2_3;
960
961 return (((unsigned long long) res3) << 48)
962 | (((unsigned long long) res2) << 32)
963 | (((unsigned long long) res1) << 16)
964 | res0;
965 }
966
967
_mpy2ir(unsigned int src1,unsigned int src2)968 static inline long long _mpy2ir(unsigned int src1, unsigned int src2)
969 {
970 if ((src1 == 0x8000) && (src2 == 0x80000000)) {
971 _overflow = 1;
972 return 0;
973 }
974 else {
975 short ls1 = src1 & 0xffff;
976 short hs1 = src1 >> 16;
977 unsigned long long hi = (((long long) hs1) * (int) src2 + (1 << 14)) >> 15;
978 unsigned long long lo = ((((long long) ls1) * (int) src2 + (1 << 14)) >> 15) & 0xFFFFFFFF;
979 return (hi << 32) | lo;
980 }
981 }
982
983
_mpy2ll(int src1,int src2)984 static inline long long _mpy2ll (int src1, int src2) {
985 short ls1 = src1 & 0xffff;
986 short hs1 = src1 >> 16;
987 short ls2 = src2 & 0xffff;
988 short hs2 = src2 >> 16;
989
990 unsigned long long hi = hs1 * hs2;
991 unsigned long long lo = (ls1 * ls2) & 0xFFFFFFFF;
992
993 return (hi << 32) | lo;
994
995 }
996
997
_mpy32(int src1,int src2)998 static inline int _mpy32(int src1, int src2)
999 {
1000 return src1 * src2;
1001 }
1002
1003
_mpy32ll(int src1,int src2)1004 static inline long long _mpy32ll(int src1, int src2)
1005 {
1006 return ((long long) src1) * src2;
1007 }
1008
_mpy32su(int src1,unsigned int src2)1009 static inline long long _mpy32su(int src1, unsigned int src2)
1010 {
1011 return ((long long) src1) * ((int) src2);
1012 }
1013
_mpy32u(unsigned int src1,unsigned int src2)1014 static inline long long _mpy32u(unsigned int src1, unsigned int src2)
1015 {
1016 return ((long long) ((int) src1)) * ((long long) ((int) src2));
1017 }
1018
_mpy32us(unsigned int src1,int src2)1019 static inline long long _mpy32us(unsigned int src1, int src2)
1020 {
1021 return ((int) src1) * ((long long) src2);
1022 }
1023
_mvd(int src2)1024 static inline int _mvd (int src2)
1025 {
1026 return src2;
1027 }
1028
1029
_norm(int src2)1030 static inline unsigned int _norm(int src2)
1031 {
1032 return norm_shift_amt_U_and_non_U(1,src2);
1033 }
1034
1035
_pack2(unsigned int src1,unsigned int src2)1036 static inline unsigned int _pack2 (unsigned int src1, unsigned int src2) {
1037 short s1[2], s2[2], r[2];
1038 int result;
1039 *((int*)s1) = src1;
1040 *((int*)s2) = src2;
1041 r[0] = s2[0];
1042 r[1] = s1[0];
1043 result = *(int*)r;
1044 return result;
1045 }
1046
1047
_packh2(unsigned int src1,unsigned int src2)1048 static inline int _packh2 (unsigned int src1, unsigned int src2) {
1049 unsigned v0 = src1 & 0xffff0000;
1050 unsigned v1 = src2 >> 16;
1051 unsigned v = v0|v1;
1052 return v;
1053
1054 }
1055
_packh4(unsigned int src1,unsigned int src2)1056 static inline unsigned int _packh4 (unsigned int src1, unsigned int src2) {
1057 unsigned v3 = (src1 >> 24) & 0xff;
1058 unsigned v2 = (src1 >> 8) & 0xff;
1059 unsigned v1 = (src2 >> 24) & 0xff;
1060 unsigned v0 = (src2 >> 8) & 0xff;
1061 unsigned v = (v3<<24) | (v2<<16) | (v1 << 8) | v0;
1062 return v;
1063 }
1064
_packhl2(unsigned int src1,unsigned int src2)1065 static inline unsigned int _packhl2 (unsigned int src1, unsigned int src2) {
1066 unsigned v0 = src1 & 0xffff0000;
1067 unsigned v1 = src2 & 0x0000ffff;
1068 unsigned v = v0|v1;
1069 return v;
1070 }
1071
_packlh2(unsigned int src1,unsigned int src2)1072 static inline unsigned int _packlh2 (unsigned int src1, unsigned int src2) {
1073 unsigned v0 = src1 << 16;
1074 unsigned v1 = (src2 >> 16) & 0xffff;
1075 unsigned v = v0|v1;
1076 return v;
1077 }
1078
1079
1080
1081
_packl4(unsigned int src1,unsigned int src2)1082 static inline unsigned int _packl4 (unsigned int src1, unsigned int src2) {
1083 unsigned v3 = (src1 >> 16) & 0xff;
1084 unsigned v2 = (src1) & 0xff;
1085 unsigned v1 = (src2 >> 16) & 0xff;
1086 unsigned v0 = (src2) & 0xff;
1087 unsigned v = (v3<<24) | (v2<<16) | (v1 << 8) | v0;
1088 return v;
1089 }
1090
1091
1092
1093
_rpack2(unsigned int src1,unsigned int src2)1094 static inline unsigned int _rpack2 (unsigned int src1, unsigned int src2) {
1095 int s1 = (int) src1;
1096 int s2 = (int) src2;
1097 s1 = util_shift_right_saturate_n (s1, -1, 32);
1098 s2 = util_shift_right_saturate_n (s2, -1, 32);
1099 return (unsigned int) (s1 & 0xffff0000) | (unsigned int) ((s2 & 0xffff0000) >>16);
1100 }
1101
1102
_rotl(unsigned int src1,unsigned int src2)1103 static inline unsigned int _rotl (unsigned int src1, unsigned int src2)
1104 {
1105 src2 &= 0x1f;
1106 return (src1 << src2) | (src1 >> (32 - src2));
1107 }
1108
1109
_sadd(int src1,int src2)1110 static inline int _sadd(int src1, int src2) {
1111 signed long long res;
1112 signed long long maxv, minv;
1113 maxv = (1LL << (32-1)) - 1;
1114 minv = (-1LL << (32-1));
1115 res = (long long) src1 + (long long) src2;
1116 if (res > maxv) {
1117 res = maxv;
1118 _overflow = 1;
1119 }
1120 else if (res < minv ) {
1121 res = minv;
1122 _overflow = 1;
1123 }
1124 return (int) res;
1125 }
1126
_saddsub(unsigned int src1,unsigned int src2)1127 static inline long long _saddsub(unsigned int src1, unsigned int src2) {
1128 int radd;
1129 signed long long rsub;
1130
1131 signed long long maxv, minv;
1132 maxv = (1LL << (32-1)) - 1;
1133 minv = (-1LL << (32-1));
1134
1135 radd = (int) src1 + (int) src2;
1136
1137 // saturate on subtract, not add
1138
1139
1140 rsub = (long long) ((int) src1) - (long long) ((int) src2);
1141 if (rsub > maxv) {
1142 rsub = maxv;
1143 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1144 /* _overflow = 1; */
1145 }
1146 else if (rsub < minv ) {
1147 rsub = minv;
1148 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1149 /* _overflow = 1; */
1150 }
1151
1152 return (((unsigned long long) radd) << 32) | ( rsub & 0x00000000ffffffff ) ;
1153 }
1154
1155
1156
_saddsub2(unsigned int src1,unsigned int src2)1157 static inline long long _saddsub2(unsigned int src1, unsigned int src2) {
1158 signed int radd[2];
1159 signed int rsub[2];
1160 signed short s1[2], s2[2];
1161
1162 signed int maxv, minv;
1163 maxv = (1L << (16-1)) - 1;
1164 minv = (-1L << (16-1));
1165
1166 *((int*)s1) = src1;
1167 *((int*)s2) = src2;
1168
1169 radd[0] = (int) s1[0] + (int) s2[0];
1170 radd[1] = (int) s1[1] + (int) s2[1];
1171
1172 rsub[0] = (int) s1[0] - (int) s2[0];
1173 rsub[1] = (int) s1[1] - (int) s2[1];
1174
1175 if (radd[0] > maxv) {
1176 radd[0] = maxv;
1177 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1178 /* _overflow = 1; */
1179 }
1180 else if (radd[0] < minv ) {
1181 radd[0] = minv;
1182 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1183 /* _overflow = 1; */
1184 }
1185
1186 if (radd[1] > maxv) {
1187 radd[1] = maxv;
1188 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1189 /* _overflow = 1; */
1190 }
1191 else if (radd[1] < minv ) {
1192 radd[1] = minv;
1193 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1194 /* _overflow = 1; */
1195 }
1196
1197
1198 if (rsub[0] > maxv) {
1199 rsub[0] = maxv;
1200 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1201 /* _overflow = 1; */
1202 }
1203 else if (rsub[0] < minv ) {
1204 rsub[0] = minv;
1205 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1206 /* _overflow = 1; */
1207 }
1208
1209 if (rsub[1] > maxv) {
1210 rsub[1] = maxv;
1211 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1212 /* _overflow = 1; */
1213 }
1214 else if (rsub[1] < minv ) {
1215 rsub[1] = minv;
1216 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1217 /* _overflow = 1; */
1218 }
1219
1220
1221 return ((((unsigned long long) radd[1]) & 0x000000000000ffff) << 48) |
1222 ((((unsigned long long) radd[0]) & 0x000000000000ffff) << 32) |
1223 ((((unsigned long long) rsub[1]) & 0x000000000000ffff) << 16) |
1224 ((((unsigned long long) rsub[0]) & 0x000000000000ffff));
1225 }
1226
1227
1228
_sadd2(int src1,int src2)1229 static inline int _sadd2(int src1, int src2) {
1230 signed short s1[2], s2[2];
1231 signed int r[2], maxv, minv;
1232
1233 maxv = (1L << (16-1)) - 1;
1234 minv = (-1L << (16-1));
1235
1236
1237 *((int*)s1) = src1;
1238 *((int*)s2) = src2;
1239
1240 r[0] = (int) s1[0] + (int) s2[0];
1241 r[1] = (int) s1[1] + (int) s2[1];
1242
1243 if (r[0] > maxv) {
1244 r[0] = maxv;
1245 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1246 /* _overflow = 1; */
1247 }
1248 else if (r[0] < minv ) {
1249 r[0] = minv;
1250 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1251 /* _overflow = 1; */
1252 }
1253 if (r[1] > maxv) {
1254 r[1] = maxv;
1255 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1256 /* _overflow = 1; */
1257 }
1258 else if (r[1] < minv ) {
1259 r[1] = minv;
1260 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1261 /* _overflow = 1; */
1262 }
1263
1264 return ((r[1] & 0xffff) << 16 ) | (r[0] & 0xffff) ;
1265 }
1266
1267
_saddus2(unsigned int src1,int src2)1268 static inline int _saddus2(unsigned int src1, int src2) {
1269 int res0, res1;
1270 unsigned int s1_0 = (src1 & 0xffff);
1271 unsigned int s1_1 = (src1 & 0xffff0000) >> 16;
1272
1273 short s2_0 = (src2 & 0xffff);
1274 short s2_1 = (src2 & 0xffff0000) >> 16;
1275
1276 res0 = s1_0 + s2_0;
1277 res1 = s1_1 + s2_1;
1278
1279 if (res0 >= 0x10000)
1280 res0 = 0xffff;
1281 else if (res0 < 0)
1282 res0 = 0;
1283
1284 if (res1 >= 0x10000)
1285 res1 = 0xffff;
1286 else if (res1 < 0)
1287 res1 = 0;
1288
1289 return (res1 << 16) | res0;
1290 }
1291
1292
_saddu4(unsigned int src1,unsigned int src2)1293 static inline unsigned int _saddu4(unsigned int src1, unsigned int src2) {
1294 unsigned int res0, res1, res2, res3;
1295 unsigned int s1_0 = (src1 & 0xff);
1296 unsigned int s1_1 = (src1 & 0xff00) >> 8;
1297 unsigned int s1_2 = (src1 & 0xff0000) >> 16;
1298 unsigned int s1_3 = (src1 & 0xff000000) >> 24;
1299
1300 unsigned int s2_0 = (src2 & 0xff);
1301 unsigned int s2_1 = (src2 & 0xff00) >> 8;
1302 unsigned int s2_2 = (src2 & 0xff0000) >> 16;
1303 unsigned int s2_3 = (src2 & 0xff000000) >> 24;
1304
1305 res0 = s1_0 + s2_0;
1306 res1 = s1_1 + s2_1;
1307 res2 = s1_2 + s2_2;
1308 res3 = s1_3 + s2_3;
1309
1310 if (res0 >= 0x100)
1311 res0 = 0xff;
1312
1313 if (res1 >= 0x100)
1314 res1 = 0xff;
1315
1316 if (res2 >= 0x100)
1317 res2 = 0xff;
1318
1319 if (res3 >= 0x100)
1320 res3 = 0xff;
1321
1322 return (res3 << 24) | (res2 << 16) | (res1 << 8) | res0;
1323
1324 }
1325
1326
1327
_sat(C6X_COMPAT_LONG40 src2)1328 static inline int _sat(C6X_COMPAT_LONG40 src2)
1329 {
1330 long long maxv = (1LL << (32-1)) - 1;
1331 long long minv = (-1LL << (32-1));
1332
1333 int hi = (int)(src2 >> 32);
1334 int lo = (int)src2;
1335 long long temp = (unsigned long long)(unsigned)lo | (unsigned long long)hi << 32;
1336 temp = orig_L40_set(temp);
1337
1338 if (temp > maxv) {
1339 temp = maxv;
1340 _overflow = 1;
1341 }
1342 else if (temp < minv) {
1343 temp = minv;
1344 _overflow = 1;
1345 }
1346 return (int) temp;
1347 }
1348
_set(unsigned int src2,unsigned int csta,unsigned int cstb)1349 static inline unsigned int _set(unsigned int src2, unsigned int csta, unsigned int cstb)
1350 {
1351 csta &= 0x1f;
1352 cstb &= 0x1f;
1353 if (csta > cstb)
1354 return src2;
1355 else {
1356 unsigned int mask = (((1 << (cstb - csta)) << 1) - 1) << csta;
1357 return src2 | mask;
1358 }
1359 }
1360
_setr(unsigned int src2,int src1)1361 static inline unsigned int _setr(unsigned int src2, int src1)
1362 {
1363 unsigned int csta = (src1 >> 5) & 0x1f;
1364 unsigned int cstb = src1 & 0x1f;
1365 if (csta > cstb)
1366 return src2;
1367 else {
1368 unsigned int mask = (((1 << (cstb - csta)) << 1) - 1) << csta;
1369 return src2 | mask;
1370 }
1371 }
1372
1373
_shfl(unsigned int src2)1374 static inline unsigned int _shfl (unsigned int src2)
1375 {
1376 unsigned short lo = src2;
1377 unsigned short hi = src2 >> 16;
1378 unsigned int result = 0;
1379 int i;
1380 for (i = 0; i < 32; i+= 2) {
1381 result >>= 1;
1382 result |= (lo & 0x1) << 31;
1383 lo >>= 1;
1384 result >>= 1;
1385 result |= (hi & 0x1) << 31;
1386 hi >>= 1;
1387 }
1388 return result;
1389 }
1390
_shfl3(unsigned int src1,unsigned int src2)1391 static inline long long _shfl3 (unsigned int src1, unsigned int src2)
1392 {
1393 unsigned short lo = src2;
1394 unsigned short hi = src1 >> 16;
1395 unsigned short mid = src1;
1396 unsigned long long result = 0;
1397 int i;
1398 for (i = 0; i < 32; i+= 2) {
1399 result >>= 1;
1400 result |= ((unsigned long long) (lo & 0x1)) << 47;
1401 lo >>= 1;
1402 result >>= 1;
1403 result |= ((unsigned long long) (mid & 0x1)) << 47;
1404 mid >>= 1;
1405 result >>= 1;
1406 result |= ((unsigned long long) (hi & 0x1)) << 47;
1407 hi >>= 1;
1408 }
1409 return result;
1410 }
1411
1412
1413
_shlmb(unsigned int src1,unsigned int src2)1414 static inline unsigned int _shlmb (unsigned int src1, unsigned int src2)
1415 {
1416 return (src2 << 8) | (src1 >> 24);
1417 }
1418
_shrmb(unsigned int src1,unsigned int src2)1419 static inline unsigned int _shrmb (unsigned int src1, unsigned int src2)
1420 {
1421 return (src2 >> 8) | (src1 << 24);
1422 }
1423
1424
_shru2(unsigned int src1,unsigned int src2)1425 static inline unsigned int _shru2 (unsigned int src1, unsigned int src2) {
1426 unsigned short hs1 = src1 >> 16;
1427 unsigned short ls1 = src1 & 0xFFFF;
1428 hs1 >>= src2;
1429 ls1 >>= src2;
1430 return (hs1 << 16) | ls1;
1431 }
1432
1433
_shr2(int src1,unsigned int src2)1434 static inline int _shr2 (int src1, unsigned int src2) {
1435 short s1[2], result[2];
1436 *((int*)s1) = src1;
1437 src2 = src2 & 31;
1438 result[0] = (int)s1[0] >> src2;
1439 result[1] = (int)s1[1] >> src2;
1440
1441 return *(int*)result;
1442 }
1443
1444
_smpy(int src1,int src2)1445 static inline int _smpy (int src1, int src2) {
1446 unsigned long long result;
1447 result = (((short) src1 * (short) src2) << 1);
1448
1449 if ((result & 0xffffffff) == 0x80000000){
1450 result = 0x7fffffff;
1451 _overflow = 1;
1452 }
1453 return (int) (result);
1454 }
1455
_smpyh(int src1,int src2)1456 static inline int _smpyh (int src1, int src2) {
1457 unsigned long long result;
1458 result = ((short) (src1 >> 16) * (short) (src2 >> 16)) << 1;
1459 if ((result & 0xffffffff) == 0x80000000){
1460 result = 0x7fffffff;
1461 _overflow = 1;
1462 }
1463 return (int) (result);
1464 }
1465
_smpyhl(int src1,int src2)1466 static inline int _smpyhl (int src1, int src2) {
1467 unsigned long long result;
1468 result = ((short) (src1 >> 16) * (short) (src2)) << 1;
1469 if ((result & 0xffffffff) == 0x80000000){
1470 result = 0x7fffffff;
1471 _overflow = 1;
1472 }
1473 return (int) (result);
1474 }
1475
_smpylh(int src1,int src2)1476 static inline int _smpylh (int src1, int src2) {
1477 unsigned long long result;
1478 result = ((short) (src1) * (short) (src2 >> 16)) << 1;
1479 if ((result & 0xffffffff) == 0x80000000){
1480 result = 0x7fffffff;
1481 _overflow = 1;
1482 }
1483 return (int) (result);
1484 }
1485
_smpy2ll(int src1,int src2)1486 static inline long long _smpy2ll (int src1, int src2) {
1487 short ls1 = src1 & 0XFFFF;
1488 short hs1 = src1 >> 16;
1489 short ls2 = src2 & 0XFFFF;
1490 short hs2 = src2 >> 16;
1491
1492 unsigned long long hi = (hs1 * hs2) << 1;
1493 unsigned long long lo = ((ls1 * ls2) << 1) & 0xFFFFFFFF;
1494 if ((hi & 0xffffffff) == 0x80000000){
1495 hi = 0x7fffffff;
1496 _overflow = 1;
1497 }
1498
1499 if ((lo & 0xffffffff) == 0x80000000){
1500 lo = 0x7fffffff;
1501 _overflow = 1;
1502 }
1503
1504 return (hi << 32) | lo;
1505 }
1506
1507
1508
1509
_smpy32(int src1,int src2)1510 static inline int _smpy32(int src1, int src2)
1511 {
1512 long long res = (long long) src1 * src2;
1513 res <<= 1;
1514 res >>= 32;
1515 return res;
1516 }
1517
TEN_satu8(short src)1518 static inline unsigned char TEN_satu8 (short src)
1519 {
1520 if (src > 0xff)
1521 return 0xff;
1522 else if (src < 0)
1523 return 0;
1524 else
1525 return src;
1526 }
1527
_spack2(int src1,int src2)1528 static inline int _spack2 (int src1, int src2) {
1529 short s1 = (short) util_saturate_n_no_state(src1,16);
1530 short s2 = (short) util_saturate_n_no_state(src2,16);
1531 return ( (unsigned int) s1 << 16) | (((int) s2) & 0xFFFF);
1532 }
1533
1534
_spacku4(int src1,int src2)1535 static inline unsigned int _spacku4 (int src1, int src2) {
1536 short lolo = src2;
1537 short lohi = src2 >> 16;
1538 short hilo = src1;
1539 short hihi = src1 >> 16;
1540
1541 lolo = TEN_satu8(lolo);
1542 lohi = TEN_satu8(lohi);
1543 hilo = TEN_satu8(hilo);
1544 hihi = TEN_satu8(hihi);
1545
1546 return (((unsigned int) hihi) << 24) | (((unsigned int) hilo) << 16) | (lohi << 8) | lolo;
1547 }
1548
1549
1550
_sshl(int src1,unsigned int src2)1551 static inline int _sshl (int src1, unsigned int src2) {
1552 short local2 = (short)(src2 & 0x7FFF);
1553 return (int) util_shift_right_saturate_n(src1, -local2, 32);
1554 }
1555
1556
1557
1558
_sshvl(int src2,int src1)1559 static inline int _sshvl (int src2, int src1) {
1560 short s1;
1561 if (src1 > 31)
1562 s1 = 31;
1563 else if (src1 < -31)
1564 s1 = -31;
1565 else
1566 s1 = src1;
1567
1568 return (int) util_shift_right_saturate_n(src2, -s1, 32);
1569 }
1570
1571
1572
1573
1574
_sshvr(int src2,int src1)1575 static inline int _sshvr (int src2, int src1) {
1576 short s1;
1577 if (src1 > 31)
1578 s1 = 31;
1579 else if (src1 < -31)
1580 s1 = -31;
1581 else
1582 s1 = src1;
1583 return (int) util_shift_right_saturate_n(src2, s1, 32);
1584 }
1585
1586
1587
1588
_ssub(int src1,int src2)1589 static inline int _ssub(int src1, int src2) {
1590 signed long long res;
1591 signed long long maxv, minv;
1592 maxv = (1LL << (32-1)) - 1;
1593 minv = (-1LL << (32-1));
1594 res = (long long) src1 - (long long) src2;
1595 if (res > maxv) {
1596 res = maxv;
1597 _overflow = 1;
1598 }
1599 else if (res < minv ) {
1600 res = minv;
1601 _overflow = 1;
1602 }
1603 return (int) res;
1604 }
1605
_ssub2(int src1,int src2)1606 static inline int _ssub2(int src1, int src2) {
1607 signed short s1[2], s2[2];
1608 signed int r[2], maxv, minv;
1609
1610 maxv = (1L << (16-1)) - 1;
1611 minv = (-1L << (16-1));
1612
1613
1614 *((int*)s1) = src1;
1615 *((int*)s2) = src2;
1616
1617 r[0] = (int) s1[0] - (int) s2[0];
1618 r[1] = (int) s1[1] - (int) s2[1];
1619
1620 if (r[0] > maxv) {
1621 r[0] = maxv;
1622 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1623 /* _overflow = 1; */
1624 }
1625 else if (r[0] < minv ) {
1626 r[0] = minv;
1627 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1628 /* _overflow = 1; */
1629 }
1630 if (r[1] > maxv) {
1631 r[1] = maxv;
1632 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1633 /* _overflow = 1; */
1634 }
1635 else if (r[1] < minv ) {
1636 r[1] = minv;
1637 /* NOTE: TI c6x does NOT set the overflow register even if results saturate */
1638 /* _overflow = 1; */
1639 }
1640
1641 return ((r[1] & 0xffff) << 16 ) | (r[0] & 0xffff) ;
1642 }
1643
1644
_subabs4(int src1,int src2)1645 static inline int _subabs4 (int src1, int src2) {
1646 int res0, res1, res2, res3;
1647 unsigned int s1_0 = (src1 & 0xff);
1648 unsigned int s1_1 = (src1 & 0xff00) >> 8;
1649 unsigned int s1_2 = (src1 & 0xff0000) >> 16;
1650 unsigned int s1_3 = (src1 & 0xff000000) >> 24;
1651
1652 unsigned int s2_0 = (src2 & 0xff);
1653 unsigned int s2_1 = (src2 & 0xff00) >> 8;
1654 unsigned int s2_2 = (src2 & 0xff0000) >> 16;
1655 unsigned int s2_3 = (src2 & 0xff000000) >> 24;
1656
1657 res0 = s1_0 - s2_0;
1658 res1 = s1_1 - s2_1;
1659 res2 = s1_2 - s2_2;
1660 res3 = s1_3 - s2_3;
1661
1662 if (res0 < 0)
1663 res0 = -res0;
1664
1665 if (res1 < 0)
1666 res1 = -res1;
1667
1668 if (res2 < 0)
1669 res2 = -res2;
1670
1671 if (res3 < 0)
1672 res3 = -res3;
1673
1674 return (res3 << 24) | (res2 << 16) | (res1 << 8) | res0;
1675 }
1676
1677
_subc(unsigned int src1,unsigned int src2)1678 static inline unsigned int _subc (unsigned int src1, unsigned int src2)
1679 {
1680 if ( src1 >= src2)
1681 return ((src1 - src2) << 1) + 1;
1682 else
1683 return src1 << 1;
1684 }
1685
1686
1687
_sub2(int src1,int src2)1688 static inline int _sub2(int src1, int src2) {
1689 short s1[2], s2[2], r[2];
1690 int result;
1691 *((int*)s1) = src1;
1692 *((int*)s2) = src2;
1693 r[0] = s1[0] - s2[0];
1694 r[1] = s1[1] - s2[1];
1695 result = *(int*)r;
1696 return result;
1697 }
1698
1699
_sub4(int src1,int src2)1700 static inline int _sub4(int src1, int src2) {
1701 char c1[4], c2[4], r[4];
1702 int result;
1703 *((int*)c1) = src1;
1704 *((int*)c2) = src2;
1705 r[0] = c1[0] - c2[0];
1706 r[1] = c1[1] - c2[1];
1707 r[2] = c1[2] - c2[2];
1708 r[3] = c1[3] - c2[3];
1709 result = *(int*)r;
1710 return result;
1711 }
1712
1713
_swap4(unsigned int src1)1714 static inline int _swap4 (unsigned int src1) {
1715 unsigned char v0 = src1;
1716 unsigned char v1 = src1 >> 8;
1717 unsigned char v2 = src1 >> 16;
1718 unsigned char v3 = src1 >> 24;
1719 unsigned v = v0<<8 | v1 | v2<<24 | v3<<16;
1720 return v;
1721 }
1722
_unpkhu4(unsigned int src1)1723 static inline unsigned int _unpkhu4 (unsigned int src1) {
1724 unsigned v0 = src1>>24;
1725 unsigned v1 = (src1>>16) & 0xff;
1726 return (v0<<16) | v1;
1727 }
1728
_unpklu4(unsigned int src1)1729 static inline unsigned int _unpklu4 (unsigned int src1) {
1730 unsigned v1 = (src1>>8) & 0xff;
1731 unsigned v0 = (src1) & 0xff;
1732 return (v1<<16) | v0;
1733 }
1734
1735
1736
1737
_xpnd2(unsigned int src1)1738 static inline unsigned int _xpnd2 (unsigned int src1) {
1739 int v0 = (src1 & 0x1) ? 0x0000ffff : 0x00000000;
1740 int v1 = (src1 & 0x2) ? 0xffff0000 : 0x00000000;
1741 return v0|v1;
1742 }
1743
_xpnd4(unsigned int src1)1744 static inline unsigned int _xpnd4 (unsigned int src1) {
1745 int v0 = (src1 & 0x1) ? 0x000000ff : 0x00000000;
1746 int v1 = (src1 & 0x2) ? 0x0000ff00 : 0x00000000;
1747 int v2 = (src1 & 0x4) ? 0x00ff0000 : 0x00000000;
1748 int v3 = (src1 & 0x8) ? 0xff000000 : 0x00000000;
1749 int r = v0|v1|v2|v3;
1750 return r;
1751 }
1752
1753
1754
1755 // end of Implemented in alphabetical order
1756
1757
1758 #endif /* __C6X_COMPAT__H */
1759
