]> git.proxmox.com Git - qemu.git/blame - target-i386/ops_sse.h
projects / qemu.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
sse fix
[qemu.git] / target-i386 / ops_sse.h
CommitLineData
664e0f19
FB		 1/*
2 * MMX/SSE/SSE2/PNI support
3 *
4 * Copyright (c) 2005 Fabrice Bellard
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 */
20#if SHIFT == 0
21#define Reg MMXReg
22#define XMM_ONLY(x...)
23#define B(n) MMX_B(n)
24#define W(n) MMX_W(n)
25#define L(n) MMX_L(n)
26#define Q(n) q
27#define SUFFIX _mmx
28#else
29#define Reg XMMReg
30#define XMM_ONLY(x...) x
31#define B(n) XMM_B(n)
32#define W(n) XMM_W(n)
33#define L(n) XMM_L(n)
34#define Q(n) XMM_Q(n)
35#define SUFFIX _xmm
36#endif
37
38void OPPROTO glue(op_psrlw, SUFFIX)(void)
39{
40 Reg *d, *s;
41 int shift;
42
43 d = (Reg *)((char *)env + PARAM1);
44 s = (Reg *)((char *)env + PARAM2);
45
46 if (s->Q(0) > 15) {
47 d->Q(0) = 0;
48#if SHIFT == 1
49 d->Q(1) = 0;
50#endif
51 } else {
52 shift = s->B(0);
53 d->W(0) >>= shift;
54 d->W(1) >>= shift;
55 d->W(2) >>= shift;
56 d->W(3) >>= shift;
57#if SHIFT == 1
58 d->W(4) >>= shift;
59 d->W(5) >>= shift;
60 d->W(6) >>= shift;
61 d->W(7) >>= shift;
62#endif
63 }
64}
65
66void OPPROTO glue(op_psraw, SUFFIX)(void)
67{
68 Reg *d, *s;
69 int shift;
70
71 d = (Reg *)((char *)env + PARAM1);
72 s = (Reg *)((char *)env + PARAM2);
73
74 if (s->Q(0) > 15) {
75 shift = 15;
76 } else {
77 shift = s->B(0);
78 }
79 d->W(0) = (int16_t)d->W(0) >> shift;
80 d->W(1) = (int16_t)d->W(1) >> shift;
81 d->W(2) = (int16_t)d->W(2) >> shift;
82 d->W(3) = (int16_t)d->W(3) >> shift;
83#if SHIFT == 1
84 d->W(4) = (int16_t)d->W(4) >> shift;
85 d->W(5) = (int16_t)d->W(5) >> shift;
86 d->W(6) = (int16_t)d->W(6) >> shift;
87 d->W(7) = (int16_t)d->W(7) >> shift;
88#endif
89}
90
91void OPPROTO glue(op_psllw, SUFFIX)(void)
92{
93 Reg *d, *s;
94 int shift;
95
96 d = (Reg *)((char *)env + PARAM1);
97 s = (Reg *)((char *)env + PARAM2);
98
99 if (s->Q(0) > 15) {
100 d->Q(0) = 0;
101#if SHIFT == 1
102 d->Q(1) = 0;
103#endif
104 } else {
105 shift = s->B(0);
106 d->W(0) <<= shift;
107 d->W(1) <<= shift;
108 d->W(2) <<= shift;
109 d->W(3) <<= shift;
110#if SHIFT == 1
111 d->W(4) <<= shift;
112 d->W(5) <<= shift;
113 d->W(6) <<= shift;
114 d->W(7) <<= shift;
115#endif
116 }
117}
118
119void OPPROTO glue(op_psrld, SUFFIX)(void)
120{
121 Reg *d, *s;
122 int shift;
123
124 d = (Reg *)((char *)env + PARAM1);
125 s = (Reg *)((char *)env + PARAM2);
126
127 if (s->Q(0) > 31) {
128 d->Q(0) = 0;
129#if SHIFT == 1
130 d->Q(1) = 0;
131#endif
132 } else {
133 shift = s->B(0);
134 d->L(0) >>= shift;
135 d->L(1) >>= shift;
136#if SHIFT == 1
137 d->L(2) >>= shift;
138 d->L(3) >>= shift;
139#endif
140 }
141}
142
143void OPPROTO glue(op_psrad, SUFFIX)(void)
144{
145 Reg *d, *s;
146 int shift;
147
148 d = (Reg *)((char *)env + PARAM1);
149 s = (Reg *)((char *)env + PARAM2);
150
151 if (s->Q(0) > 31) {
152 shift = 31;
153 } else {
154 shift = s->B(0);
155 }
156 d->L(0) = (int32_t)d->L(0) >> shift;
157 d->L(1) = (int32_t)d->L(1) >> shift;
158#if SHIFT == 1
159 d->L(2) = (int32_t)d->L(2) >> shift;
160 d->L(3) = (int32_t)d->L(3) >> shift;
161#endif
162}
163
164void OPPROTO glue(op_pslld, SUFFIX)(void)
165{
166 Reg *d, *s;
167 int shift;
168
169 d = (Reg *)((char *)env + PARAM1);
170 s = (Reg *)((char *)env + PARAM2);
171
172 if (s->Q(0) > 31) {
173 d->Q(0) = 0;
174#if SHIFT == 1
175 d->Q(1) = 0;
176#endif
177 } else {
178 shift = s->B(0);
179 d->L(0) <<= shift;
180 d->L(1) <<= shift;
181#if SHIFT == 1
182 d->L(2) <<= shift;
183 d->L(3) <<= shift;
184#endif
185 }
186}
187
188void OPPROTO glue(op_psrlq, SUFFIX)(void)
189{
190 Reg *d, *s;
191 int shift;
192
193 d = (Reg *)((char *)env + PARAM1);
194 s = (Reg *)((char *)env + PARAM2);
195
196 if (s->Q(0) > 63) {
197 d->Q(0) = 0;
198#if SHIFT == 1
199 d->Q(1) = 0;
200#endif
201 } else {
202 shift = s->B(0);
203 d->Q(0) >>= shift;
204#if SHIFT == 1
205 d->Q(1) >>= shift;
206#endif
207 }
208}
209
210void OPPROTO glue(op_psllq, SUFFIX)(void)
211{
212 Reg *d, *s;
213 int shift;
214
215 d = (Reg *)((char *)env + PARAM1);
216 s = (Reg *)((char *)env + PARAM2);
217
218 if (s->Q(0) > 63) {
219 d->Q(0) = 0;
220#if SHIFT == 1
221 d->Q(1) = 0;
222#endif
223 } else {
224 shift = s->B(0);
225 d->Q(0) <<= shift;
226#if SHIFT == 1
227 d->Q(1) <<= shift;
228#endif
229 }
230}
231
232#if SHIFT == 1
233void OPPROTO glue(op_psrldq, SUFFIX)(void)
234{
235 Reg *d, *s;
236 int shift, i;
237
238 d = (Reg *)((char *)env + PARAM1);
239 s = (Reg *)((char *)env + PARAM2);
240 shift = s->L(0);
241 if (shift > 16)
242 shift = 16;
243 for(i = 0; i < 16 - shift; i++)
244 d->B(i) = d->B(i + shift);
245 for(i = 16 - shift; i < 16; i++)
246 d->B(i) = 0;
247 FORCE_RET();
248}
249
250void OPPROTO glue(op_pslldq, SUFFIX)(void)
251{
252 Reg *d, *s;
253 int shift, i;
254
255 d = (Reg *)((char *)env + PARAM1);
256 s = (Reg *)((char *)env + PARAM2);
257 shift = s->L(0);
258 if (shift > 16)
259 shift = 16;
260 for(i = 15; i >= shift; i--)
261 d->B(i) = d->B(i - shift);
262 for(i = 0; i < shift; i++)
263 d->B(i) = 0;
264 FORCE_RET();
265}
266#endif
267
268#define SSE_OP_B(name, F)\
269void OPPROTO glue(name, SUFFIX) (void)\
270{\
271 Reg *d, *s;\
272 d = (Reg *)((char *)env + PARAM1);\
273 s = (Reg *)((char *)env + PARAM2);\
274 d->B(0) = F(d->B(0), s->B(0));\
275 d->B(1) = F(d->B(1), s->B(1));\
276 d->B(2) = F(d->B(2), s->B(2));\
277 d->B(3) = F(d->B(3), s->B(3));\
278 d->B(4) = F(d->B(4), s->B(4));\
279 d->B(5) = F(d->B(5), s->B(5));\
280 d->B(6) = F(d->B(6), s->B(6));\
281 d->B(7) = F(d->B(7), s->B(7));\
282 XMM_ONLY(\
283 d->B(8) = F(d->B(8), s->B(8));\
284 d->B(9) = F(d->B(9), s->B(9));\
285 d->B(10) = F(d->B(10), s->B(10));\
286 d->B(11) = F(d->B(11), s->B(11));\
287 d->B(12) = F(d->B(12), s->B(12));\
288 d->B(13) = F(d->B(13), s->B(13));\
289 d->B(14) = F(d->B(14), s->B(14));\
290 d->B(15) = F(d->B(15), s->B(15));\
291 )\
292}
293
294#define SSE_OP_W(name, F)\
295void OPPROTO glue(name, SUFFIX) (void)\
296{\
297 Reg *d, *s;\
298 d = (Reg *)((char *)env + PARAM1);\
299 s = (Reg *)((char *)env + PARAM2);\
300 d->W(0) = F(d->W(0), s->W(0));\
301 d->W(1) = F(d->W(1), s->W(1));\
302 d->W(2) = F(d->W(2), s->W(2));\
303 d->W(3) = F(d->W(3), s->W(3));\
304 XMM_ONLY(\
305 d->W(4) = F(d->W(4), s->W(4));\
306 d->W(5) = F(d->W(5), s->W(5));\
307 d->W(6) = F(d->W(6), s->W(6));\
308 d->W(7) = F(d->W(7), s->W(7));\
309 )\
310}
311
312#define SSE_OP_L(name, F)\
313void OPPROTO glue(name, SUFFIX) (void)\
314{\
315 Reg *d, *s;\
316 d = (Reg *)((char *)env + PARAM1);\
317 s = (Reg *)((char *)env + PARAM2);\
318 d->L(0) = F(d->L(0), s->L(0));\
319 d->L(1) = F(d->L(1), s->L(1));\
320 XMM_ONLY(\
321 d->L(2) = F(d->L(2), s->L(2));\
322 d->L(3) = F(d->L(3), s->L(3));\
323 )\
324}
325
326#define SSE_OP_Q(name, F)\
327void OPPROTO glue(name, SUFFIX) (void)\
328{\
329 Reg *d, *s;\
330 d = (Reg *)((char *)env + PARAM1);\
331 s = (Reg *)((char *)env + PARAM2);\
332 d->Q(0) = F(d->Q(0), s->Q(0));\
333 XMM_ONLY(\
334 d->Q(1) = F(d->Q(1), s->Q(1));\
335 )\
336}
337
338#if SHIFT == 0
339static inline int satub(int x)
340{
341 if (x < 0)
342 return 0;
343 else if (x > 255)
344 return 255;
345 else
346 return x;
347}
348
349static inline int satuw(int x)
350{
351 if (x < 0)
352 return 0;
353 else if (x > 65535)
354 return 65535;
355 else
356 return x;
357}
358
359static inline int satsb(int x)
360{
361 if (x < -128)
362 return -128;
363 else if (x > 127)
364 return 127;
365 else
366 return x;
367}
368
369static inline int satsw(int x)
370{
371 if (x < -32768)
372 return -32768;
373 else if (x > 32767)
374 return 32767;
375 else
376 return x;
377}
378
379#define FADD(a, b) ((a) + (b))
380#define FADDUB(a, b) satub((a) + (b))
381#define FADDUW(a, b) satuw((a) + (b))
382#define FADDSB(a, b) satsb((int8_t)(a) + (int8_t)(b))
383#define FADDSW(a, b) satsw((int16_t)(a) + (int16_t)(b))
384
385#define FSUB(a, b) ((a) - (b))
386#define FSUBUB(a, b) satub((a) - (b))
387#define FSUBUW(a, b) satuw((a) - (b))
388#define FSUBSB(a, b) satsb((int8_t)(a) - (int8_t)(b))
389#define FSUBSW(a, b) satsw((int16_t)(a) - (int16_t)(b))
390#define FMINUB(a, b) ((a) < (b)) ? (a) : (b)
391#define FMINSW(a, b) ((int16_t)(a) < (int16_t)(b)) ? (a) : (b)
392#define FMAXUB(a, b) ((a) > (b)) ? (a) : (b)
393#define FMAXSW(a, b) ((int16_t)(a) > (int16_t)(b)) ? (a) : (b)
394
395#define FAND(a, b) (a) & (b)
396#define FANDN(a, b) ((~(a)) & (b))
397#define FOR(a, b) (a) | (b)
398#define FXOR(a, b) (a) ^ (b)
399
400#define FCMPGTB(a, b) (int8_t)(a) > (int8_t)(b) ? -1 : 0
401#define FCMPGTW(a, b) (int16_t)(a) > (int16_t)(b) ? -1 : 0
402#define FCMPGTL(a, b) (int32_t)(a) > (int32_t)(b) ? -1 : 0
403#define FCMPEQ(a, b) (a) == (b) ? -1 : 0
404
405#define FMULLW(a, b) (a) * (b)
406#define FMULHUW(a, b) (a) * (b) >> 16
407#define FMULHW(a, b) (int16_t)(a) * (int16_t)(b) >> 16
408
409#define FAVG(a, b) ((a) + (b) + 1) >> 1
410#endif
411
412SSE_OP_B(op_paddb, FADD)
413SSE_OP_W(op_paddw, FADD)
414SSE_OP_L(op_paddl, FADD)
415SSE_OP_Q(op_paddq, FADD)
416
417SSE_OP_B(op_psubb, FSUB)
418SSE_OP_W(op_psubw, FSUB)
419SSE_OP_L(op_psubl, FSUB)
420SSE_OP_Q(op_psubq, FSUB)
421
422SSE_OP_B(op_paddusb, FADDUB)
423SSE_OP_B(op_paddsb, FADDSB)
424SSE_OP_B(op_psubusb, FSUBUB)
425SSE_OP_B(op_psubsb, FSUBSB)
426
427SSE_OP_W(op_paddusw, FADDUW)
428SSE_OP_W(op_paddsw, FADDSW)
429SSE_OP_W(op_psubusw, FSUBUW)
430SSE_OP_W(op_psubsw, FSUBSW)
431
432SSE_OP_B(op_pminub, FMINUB)
433SSE_OP_B(op_pmaxub, FMAXUB)
434
435SSE_OP_W(op_pminsw, FMINSW)
436SSE_OP_W(op_pmaxsw, FMAXSW)
437
438SSE_OP_Q(op_pand, FAND)
439SSE_OP_Q(op_pandn, FANDN)
440SSE_OP_Q(op_por, FOR)
441SSE_OP_Q(op_pxor, FXOR)
442
443SSE_OP_B(op_pcmpgtb, FCMPGTB)
444SSE_OP_W(op_pcmpgtw, FCMPGTW)
445SSE_OP_L(op_pcmpgtl, FCMPGTL)
446
447SSE_OP_B(op_pcmpeqb, FCMPEQ)
448SSE_OP_W(op_pcmpeqw, FCMPEQ)
449SSE_OP_L(op_pcmpeql, FCMPEQ)
450
451SSE_OP_W(op_pmullw, FMULLW)
452SSE_OP_W(op_pmulhuw, FMULHUW)
453SSE_OP_W(op_pmulhw, FMULHW)
454
455SSE_OP_B(op_pavgb, FAVG)
456SSE_OP_W(op_pavgw, FAVG)
457
458void OPPROTO glue(op_pmuludq, SUFFIX) (void)
459{
460 Reg *d, *s;
461 d = (Reg *)((char *)env + PARAM1);
462 s = (Reg *)((char *)env + PARAM2);
463
464 d->Q(0) = (uint64_t)s->L(0) * (uint64_t)d->L(0);
465#if SHIFT == 1
466 d->Q(1) = (uint64_t)s->L(2) * (uint64_t)d->L(2);
467#endif
468}
469
470void OPPROTO glue(op_pmaddwd, SUFFIX) (void)
471{
472 int i;
473 Reg *d, *s;
474 d = (Reg *)((char *)env + PARAM1);
475 s = (Reg *)((char *)env + PARAM2);
476
477 for(i = 0; i < (2 << SHIFT); i++) {
478 d->L(i) = (int16_t)s->W(2*i) * (int16_t)d->W(2*i) +
479 (int16_t)s->W(2*i+1) * (int16_t)d->W(2*i+1);
480 }
481}
482
483#if SHIFT == 0
484static inline int abs1(int a)
485{
486 if (a < 0)
487 return -a;
488 else
489 return a;
490}
491#endif
492void OPPROTO glue(op_psadbw, SUFFIX) (void)
493{
494 unsigned int val;
495 Reg *d, *s;
496 d = (Reg *)((char *)env + PARAM1);
497 s = (Reg *)((char *)env + PARAM2);
498
499 val = 0;
500 val += abs1(d->B(0) - s->B(0));
501 val += abs1(d->B(1) - s->B(1));
502 val += abs1(d->B(2) - s->B(2));
503 val += abs1(d->B(3) - s->B(3));
504 val += abs1(d->B(4) - s->B(4));
505 val += abs1(d->B(5) - s->B(5));
506 val += abs1(d->B(6) - s->B(6));
507 val += abs1(d->B(7) - s->B(7));
508 d->Q(0) = val;
509#if SHIFT == 1
510 val = 0;
511 val += abs1(d->B(8) - s->B(8));
512 val += abs1(d->B(9) - s->B(9));
513 val += abs1(d->B(10) - s->B(10));
514 val += abs1(d->B(11) - s->B(11));
515 val += abs1(d->B(12) - s->B(12));
516 val += abs1(d->B(13) - s->B(13));
517 val += abs1(d->B(14) - s->B(14));
518 val += abs1(d->B(15) - s->B(15));
519 d->Q(1) = val;
520#endif
521}
522
523void OPPROTO glue(op_maskmov, SUFFIX) (void)
524{
525 int i;
526 Reg *d, *s;
527 d = (Reg *)((char *)env + PARAM1);
528 s = (Reg *)((char *)env + PARAM2);
529 for(i = 0; i < (8 << SHIFT); i++) {
530 if (s->B(i) & 0x80)
d52cf7a6 531 stb(A0 + i, d->B(i));
664e0f19
FB		 532 }
533}
534
535void OPPROTO glue(op_movl_mm_T0, SUFFIX) (void)
536{
537 Reg *d;
538 d = (Reg *)((char *)env + PARAM1);
539 d->L(0) = T0;
540 d->L(1) = 0;
541#if SHIFT == 1
542 d->Q(1) = 0;
543#endif
544}
545
546void OPPROTO glue(op_movl_T0_mm, SUFFIX) (void)
547{
548 Reg *s;
549 s = (Reg *)((char *)env + PARAM1);
550 T0 = s->L(0);
551}
552
553#if SHIFT == 0
554void OPPROTO glue(op_pshufw, SUFFIX) (void)
555{
556 Reg r, *d, *s;
557 int order;
558 d = (Reg *)((char *)env + PARAM1);
559 s = (Reg *)((char *)env + PARAM2);
560 order = PARAM3;
561 r.W(0) = s->W(order & 3);
562 r.W(1) = s->W((order >> 2) & 3);
563 r.W(2) = s->W((order >> 4) & 3);
564 r.W(3) = s->W((order >> 6) & 3);
565 *d = r;
566}
567#else
d52cf7a6
FB		 568void OPPROTO op_shufps(void)
569{
570 Reg r, *d, *s;
571 int order;
572 d = (Reg *)((char *)env + PARAM1);
573 s = (Reg *)((char *)env + PARAM2);
574 order = PARAM3;
575 r.L(0) = d->L(order & 3);
576 r.L(1) = d->L((order >> 2) & 3);
577 r.L(2) = s->L((order >> 4) & 3);
578 r.L(3) = s->L((order >> 6) & 3);
579 *d = r;
580}
581
664e0f19
FB		 582void OPPROTO op_shufpd(void)
583{
584 Reg r, *d, *s;
585 int order;
586 d = (Reg *)((char *)env + PARAM1);
587 s = (Reg *)((char *)env + PARAM2);
588 order = PARAM3;
d52cf7a6 589 r.Q(0) = d->Q(order & 1);
664e0f19
FB		 590 r.Q(1) = s->Q((order >> 1) & 1);
591 *d = r;
592}
593
594void OPPROTO glue(op_pshufd, SUFFIX) (void)
595{
596 Reg r, *d, *s;
597 int order;
598 d = (Reg *)((char *)env + PARAM1);
599 s = (Reg *)((char *)env + PARAM2);
600 order = PARAM3;
601 r.L(0) = s->L(order & 3);
602 r.L(1) = s->L((order >> 2) & 3);
603 r.L(2) = s->L((order >> 4) & 3);
604 r.L(3) = s->L((order >> 6) & 3);
605 *d = r;
606}
607
608void OPPROTO glue(op_pshuflw, SUFFIX) (void)
609{
610 Reg r, *d, *s;
611 int order;
612 d = (Reg *)((char *)env + PARAM1);
613 s = (Reg *)((char *)env + PARAM2);
614 order = PARAM3;
615 r.W(0) = s->W(order & 3);
616 r.W(1) = s->W((order >> 2) & 3);
617 r.W(2) = s->W((order >> 4) & 3);
618 r.W(3) = s->W((order >> 6) & 3);
619 r.Q(1) = s->Q(1);
620 *d = r;
621}
622
623void OPPROTO glue(op_pshufhw, SUFFIX) (void)
624{
625 Reg r, *d, *s;
626 int order;
627 d = (Reg *)((char *)env + PARAM1);
628 s = (Reg *)((char *)env + PARAM2);
629 order = PARAM3;
630 r.Q(0) = s->Q(0);
631 r.W(4) = s->W(4 + (order & 3));
632 r.W(5) = s->W(4 + ((order >> 2) & 3));
633 r.W(6) = s->W(4 + ((order >> 4) & 3));
634 r.W(7) = s->W(4 + ((order >> 6) & 3));
635 *d = r;
636}
637#endif
638
639#if SHIFT == 1
640/* FPU ops */
641/* XXX: not accurate */
642
643#define SSE_OP_S(name, F)\
644void OPPROTO op_ ## name ## ps (void)\
645{\
646 Reg *d, *s;\
647 d = (Reg *)((char *)env + PARAM1);\
648 s = (Reg *)((char *)env + PARAM2);\
649 d->XMM_S(0) = F(d->XMM_S(0), s->XMM_S(0));\
650 d->XMM_S(1) = F(d->XMM_S(1), s->XMM_S(1));\
651 d->XMM_S(2) = F(d->XMM_S(2), s->XMM_S(2));\
652 d->XMM_S(3) = F(d->XMM_S(3), s->XMM_S(3));\
653}\
654\
655void OPPROTO op_ ## name ## ss (void)\
656{\
657 Reg *d, *s;\
658 d = (Reg *)((char *)env + PARAM1);\
659 s = (Reg *)((char *)env + PARAM2);\
660 d->XMM_S(0) = F(d->XMM_S(0), s->XMM_S(0));\
661}\
662void OPPROTO op_ ## name ## pd (void)\
663{\
664 Reg *d, *s;\
665 d = (Reg *)((char *)env + PARAM1);\
666 s = (Reg *)((char *)env + PARAM2);\
667 d->XMM_D(0) = F(d->XMM_D(0), s->XMM_D(0));\
668 d->XMM_D(1) = F(d->XMM_D(1), s->XMM_D(1));\
669}\
670\
671void OPPROTO op_ ## name ## sd (void)\
672{\
673 Reg *d, *s;\
674 d = (Reg *)((char *)env + PARAM1);\
675 s = (Reg *)((char *)env + PARAM2);\
676 d->XMM_D(0) = F(d->XMM_D(0), s->XMM_D(0));\
677}
678
679#define FPU_ADD(a, b) (a) + (b)
680#define FPU_SUB(a, b) (a) - (b)
681#define FPU_MUL(a, b) (a) * (b)
682#define FPU_DIV(a, b) (a) / (b)
683#define FPU_MIN(a, b) (a) < (b) ? (a) : (b)
684#define FPU_MAX(a, b) (a) > (b) ? (a) : (b)
685#define FPU_SQRT(a, b) sqrt(b)
686
687SSE_OP_S(add, FPU_ADD)
688SSE_OP_S(sub, FPU_SUB)
689SSE_OP_S(mul, FPU_MUL)
690SSE_OP_S(div, FPU_DIV)
691SSE_OP_S(min, FPU_MIN)
692SSE_OP_S(max, FPU_MAX)
693SSE_OP_S(sqrt, FPU_SQRT)
694
695
696/* float to float conversions */
697void OPPROTO op_cvtps2pd(void)
698{
699 float s0, s1;
700 Reg *d, *s;
701 d = (Reg *)((char *)env + PARAM1);
702 s = (Reg *)((char *)env + PARAM2);
703 s0 = s->XMM_S(0);
704 s1 = s->XMM_S(1);
705 d->XMM_D(0) = s0;
706 d->XMM_D(1) = s1;
707}
708
709void OPPROTO op_cvtpd2ps(void)
710{
711 Reg *d, *s;
712 d = (Reg *)((char *)env + PARAM1);
713 s = (Reg *)((char *)env + PARAM2);
714 d->XMM_S(0) = s->XMM_D(0);
715 d->XMM_S(1) = s->XMM_D(1);
716 d->Q(1) = 0;
717}
718
719void OPPROTO op_cvtss2sd(void)
720{
721 Reg *d, *s;
722 d = (Reg *)((char *)env + PARAM1);
723 s = (Reg *)((char *)env + PARAM2);
724 d->XMM_D(0) = s->XMM_S(0);
725}
726
727void OPPROTO op_cvtsd2ss(void)
728{
729 Reg *d, *s;
730 d = (Reg *)((char *)env + PARAM1);
731 s = (Reg *)((char *)env + PARAM2);
732 d->XMM_S(0) = s->XMM_D(0);
733}
734
735/* integer to float */
736void OPPROTO op_cvtdq2ps(void)
737{
738 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
739 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
740 d->XMM_S(0) = (int32_t)s->XMM_L(0);
741 d->XMM_S(1) = (int32_t)s->XMM_L(1);
742 d->XMM_S(2) = (int32_t)s->XMM_L(2);
743 d->XMM_S(3) = (int32_t)s->XMM_L(3);
744}
745
746void OPPROTO op_cvtdq2pd(void)
747{
748 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
749 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
750 int32_t l0, l1;
751 l0 = (int32_t)s->XMM_L(0);
752 l1 = (int32_t)s->XMM_L(1);
753 d->XMM_D(0) = l0;
754 d->XMM_D(1) = l1;
755}
756
757void OPPROTO op_cvtpi2ps(void)
758{
759 XMMReg *d = (Reg *)((char *)env + PARAM1);
760 MMXReg *s = (MMXReg *)((char *)env + PARAM2);
761 d->XMM_S(0) = (int32_t)s->MMX_L(0);
762 d->XMM_S(1) = (int32_t)s->MMX_L(1);
763}
764
765void OPPROTO op_cvtpi2pd(void)
766{
767 XMMReg *d = (Reg *)((char *)env + PARAM1);
768 MMXReg *s = (MMXReg *)((char *)env + PARAM2);
769 d->XMM_D(0) = (int32_t)s->MMX_L(0);
770 d->XMM_D(1) = (int32_t)s->MMX_L(1);
771}
772
773void OPPROTO op_cvtsi2ss(void)
774{
775 XMMReg *d = (Reg *)((char *)env + PARAM1);
776 d->XMM_S(0) = (int32_t)T0;
777}
778
779void OPPROTO op_cvtsi2sd(void)
780{
781 XMMReg *d = (Reg *)((char *)env + PARAM1);
782 d->XMM_D(0) = (int32_t)T0;
783}
784
785#ifdef TARGET_X86_64
786void OPPROTO op_cvtsq2ss(void)
787{
788 XMMReg *d = (Reg *)((char *)env + PARAM1);
789 d->XMM_S(0) = (int64_t)T0;
790}
791
792void OPPROTO op_cvtsq2sd(void)
793{
794 XMMReg *d = (Reg *)((char *)env + PARAM1);
795 d->XMM_D(0) = (int64_t)T0;
796}
797#endif
798
799/* float to integer */
800void OPPROTO op_cvtps2dq(void)
801{
802 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
803 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
804 d->XMM_L(0) = lrint(s->XMM_S(0));
805 d->XMM_L(1) = lrint(s->XMM_S(1));
806 d->XMM_L(2) = lrint(s->XMM_S(2));
807 d->XMM_L(3) = lrint(s->XMM_S(3));
808}
809
810void OPPROTO op_cvtpd2dq(void)
811{
812 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
813 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
814 d->XMM_L(0) = lrint(s->XMM_D(0));
815 d->XMM_L(1) = lrint(s->XMM_D(1));
816 d->XMM_Q(1) = 0;
817}
818
819void OPPROTO op_cvtps2pi(void)
820{
821 MMXReg *d = (MMXReg *)((char *)env + PARAM1);
822 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
823 d->MMX_L(0) = lrint(s->XMM_S(0));
824 d->MMX_L(1) = lrint(s->XMM_S(1));
825}
826
827void OPPROTO op_cvtpd2pi(void)
828{
829 MMXReg *d = (MMXReg *)((char *)env + PARAM1);
830 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
831 d->MMX_L(0) = lrint(s->XMM_D(0));
832 d->MMX_L(1) = lrint(s->XMM_D(1));
833}
834
835void OPPROTO op_cvtss2si(void)
836{
837 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
838 T0 = (int32_t)lrint(s->XMM_S(0));
839}
840
841void OPPROTO op_cvtsd2si(void)
842{
843 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
844 T0 = (int32_t)lrint(s->XMM_D(0));
845}
846
847#ifdef TARGET_X86_64
848void OPPROTO op_cvtss2sq(void)
849{
850 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
851 T0 = llrint(s->XMM_S(0));
852}
853
854void OPPROTO op_cvtsd2sq(void)
855{
856 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
857 T0 = llrint(s->XMM_D(0));
858}
859#endif
860
861/* float to integer truncated */
862void OPPROTO op_cvttps2dq(void)
863{
864 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
865 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
866 d->XMM_L(0) = (int32_t)s->XMM_S(0);
867 d->XMM_L(1) = (int32_t)s->XMM_S(1);
868 d->XMM_L(2) = (int32_t)s->XMM_S(2);
869 d->XMM_L(3) = (int32_t)s->XMM_S(3);
870}
871
872void OPPROTO op_cvttpd2dq(void)
873{
874 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
875 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
876 d->XMM_L(0) = (int32_t)s->XMM_D(0);
877 d->XMM_L(1) = (int32_t)s->XMM_D(1);
878 d->XMM_Q(1) = 0;
879}
880
881void OPPROTO op_cvttps2pi(void)
882{
883 MMXReg *d = (MMXReg *)((char *)env + PARAM1);
884 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
885 d->MMX_L(0) = (int32_t)(s->XMM_S(0));
886 d->MMX_L(1) = (int32_t)(s->XMM_S(1));
887}
888
889void OPPROTO op_cvttpd2pi(void)
890{
891 MMXReg *d = (MMXReg *)((char *)env + PARAM1);
892 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
893 d->MMX_L(0) = (int32_t)(s->XMM_D(0));
894 d->MMX_L(1) = (int32_t)(s->XMM_D(1));
895}
896
897void OPPROTO op_cvttss2si(void)
898{
899 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
900 T0 = (int32_t)(s->XMM_S(0));
901}
902
903void OPPROTO op_cvttsd2si(void)
904{
905 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
906 T0 = (int32_t)(s->XMM_D(0));
907}
908
909#ifdef TARGET_X86_64
910void OPPROTO op_cvttss2sq(void)
911{
912 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
913 T0 = (int64_t)(s->XMM_S(0));
914}
915
916void OPPROTO op_cvttsd2sq(void)
917{
918 XMMReg *s = (XMMReg *)((char *)env + PARAM1);
919 T0 = (int64_t)(s->XMM_D(0));
920}
921#endif
922
923void OPPROTO op_rsqrtps(void)
924{
925 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
926 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
927 d->XMM_S(0) = approx_rsqrt(s->XMM_S(0));
928 d->XMM_S(1) = approx_rsqrt(s->XMM_S(1));
929 d->XMM_S(2) = approx_rsqrt(s->XMM_S(2));
930 d->XMM_S(3) = approx_rsqrt(s->XMM_S(3));
931}
932
933void OPPROTO op_rsqrtss(void)
934{
935 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
936 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
937 d->XMM_S(0) = approx_rsqrt(s->XMM_S(0));
938}
939
940void OPPROTO op_rcpps(void)
941{
942 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
943 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
944 d->XMM_S(0) = approx_rcp(s->XMM_S(0));
945 d->XMM_S(1) = approx_rcp(s->XMM_S(1));
946 d->XMM_S(2) = approx_rcp(s->XMM_S(2));
947 d->XMM_S(3) = approx_rcp(s->XMM_S(3));
948}
949
950void OPPROTO op_rcpss(void)
951{
952 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
953 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
954 d->XMM_S(0) = approx_rcp(s->XMM_S(0));
955}
956
957void OPPROTO op_haddps(void)
958{
959 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
960 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
961 XMMReg r;
962 r.XMM_S(0) = d->XMM_S(0) + d->XMM_S(1);
963 r.XMM_S(1) = d->XMM_S(2) + d->XMM_S(3);
964 r.XMM_S(2) = s->XMM_S(0) + s->XMM_S(1);
965 r.XMM_S(3) = s->XMM_S(2) + s->XMM_S(3);
966 *d = r;
967}
968
969void OPPROTO op_haddpd(void)
970{
971 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
972 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
973 XMMReg r;
974 r.XMM_D(0) = d->XMM_D(0) + d->XMM_D(1);
975 r.XMM_D(1) = s->XMM_D(0) + s->XMM_D(1);
976 *d = r;
977}
978
979void OPPROTO op_hsubps(void)
980{
981 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
982 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
983 XMMReg r;
984 r.XMM_S(0) = d->XMM_S(0) - d->XMM_S(1);
985 r.XMM_S(1) = d->XMM_S(2) - d->XMM_S(3);
986 r.XMM_S(2) = s->XMM_S(0) - s->XMM_S(1);
987 r.XMM_S(3) = s->XMM_S(2) - s->XMM_S(3);
988 *d = r;
989}
990
991void OPPROTO op_hsubpd(void)
992{
993 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
994 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
995 XMMReg r;
996 r.XMM_D(0) = d->XMM_D(0) - d->XMM_D(1);
997 r.XMM_D(1) = s->XMM_D(0) - s->XMM_D(1);
998 *d = r;
999}
1000
1001void OPPROTO op_addsubps(void)
1002{
1003 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
1004 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
1005 d->XMM_S(0) = d->XMM_S(0) - s->XMM_S(0);
1006 d->XMM_S(1) = d->XMM_S(1) + s->XMM_S(1);
1007 d->XMM_S(2) = d->XMM_S(2) - s->XMM_S(2);
1008 d->XMM_S(3) = d->XMM_S(3) + s->XMM_S(3);
1009}
1010
1011void OPPROTO op_addsubpd(void)
1012{
1013 XMMReg *d = (XMMReg *)((char *)env + PARAM1);
1014 XMMReg *s = (XMMReg *)((char *)env + PARAM2);
1015 d->XMM_D(0) = d->XMM_D(0) - s->XMM_D(0);
1016 d->XMM_D(1) = d->XMM_D(1) + s->XMM_D(1);
1017}
1018
1019/* XXX: unordered */
1020#define SSE_OP_CMP(name, F)\
1021void OPPROTO op_ ## name ## ps (void)\
1022{\
1023 Reg *d, *s;\
1024 d = (Reg *)((char *)env + PARAM1);\
1025 s = (Reg *)((char *)env + PARAM2);\
1026 d->XMM_L(0) = F(d->XMM_S(0), s->XMM_S(0));\
1027 d->XMM_L(1) = F(d->XMM_S(1), s->XMM_S(1));\
1028 d->XMM_L(2) = F(d->XMM_S(2), s->XMM_S(2));\
1029 d->XMM_L(3) = F(d->XMM_S(3), s->XMM_S(3));\
1030}\
1031\
1032void OPPROTO op_ ## name ## ss (void)\
1033{\
1034 Reg *d, *s;\
1035 d = (Reg *)((char *)env + PARAM1);\
1036 s = (Reg *)((char *)env + PARAM2);\
1037 d->XMM_L(0) = F(d->XMM_S(0), s->XMM_S(0));\
1038}\
1039void OPPROTO op_ ## name ## pd (void)\
1040{\
1041 Reg *d, *s;\
1042 d = (Reg *)((char *)env + PARAM1);\
1043 s = (Reg *)((char *)env + PARAM2);\
1044 d->XMM_Q(0) = F(d->XMM_D(0), s->XMM_D(0));\
1045 d->XMM_Q(1) = F(d->XMM_D(1), s->XMM_D(1));\
1046}\
1047\
1048void OPPROTO op_ ## name ## sd (void)\
1049{\
1050 Reg *d, *s;\
1051 d = (Reg *)((char *)env + PARAM1);\
1052 s = (Reg *)((char *)env + PARAM2);\
1053 d->XMM_Q(0) = F(d->XMM_D(0), s->XMM_D(0));\
1054}
1055
1056#define FPU_CMPEQ(a, b) (a) == (b) ? -1 : 0
1057#define FPU_CMPLT(a, b) (a) < (b) ? -1 : 0
1058#define FPU_CMPLE(a, b) (a) <= (b) ? -1 : 0
1059#define FPU_CMPUNORD(a, b) (fpu_isnan(a) || fpu_isnan(b)) ? - 1 : 0
1060#define FPU_CMPNEQ(a, b) (a) == (b) ? 0 : -1
1061#define FPU_CMPNLT(a, b) (a) < (b) ? 0 : -1
1062#define FPU_CMPNLE(a, b) (a) <= (b) ? 0 : -1
1063#define FPU_CMPORD(a, b) (!fpu_isnan(a) && !fpu_isnan(b)) ? - 1 : 0
1064
1065SSE_OP_CMP(cmpeq, FPU_CMPEQ)
1066SSE_OP_CMP(cmplt, FPU_CMPLT)
1067SSE_OP_CMP(cmple, FPU_CMPLE)
1068SSE_OP_CMP(cmpunord, FPU_CMPUNORD)
1069SSE_OP_CMP(cmpneq, FPU_CMPNEQ)
1070SSE_OP_CMP(cmpnlt, FPU_CMPNLT)
1071SSE_OP_CMP(cmpnle, FPU_CMPNLE)
1072SSE_OP_CMP(cmpord, FPU_CMPORD)
1073
1074void OPPROTO op_ucomiss(void)
1075{
1076 int eflags;
1077 float s0, s1;
1078 Reg *d, *s;
1079 d = (Reg *)((char *)env + PARAM1);
1080 s = (Reg *)((char *)env + PARAM2);
1081
1082 s0 = d->XMM_S(0);
1083 s1 = s->XMM_S(0);
1084 if (s0 < s1)
1085 eflags = CC_C;
1086 else if (s0 == s1)
1087 eflags = CC_Z;
1088 else
1089 eflags = 0;
1090 CC_SRC = eflags;
1091 FORCE_RET();
1092}
1093
1094void OPPROTO op_comiss(void)
1095{
1096 int eflags;
1097 float s0, s1;
1098 Reg *d, *s;
1099 d = (Reg *)((char *)env + PARAM1);
1100 s = (Reg *)((char *)env + PARAM2);
1101
1102 s0 = d->XMM_S(0);
1103 s1 = s->XMM_S(0);
1104 if (s0 < s1)
1105 eflags = CC_C;
1106 else if (s0 == s1)
1107 eflags = CC_Z;
1108 else
1109 eflags = 0;
1110 CC_SRC = eflags;
1111 FORCE_RET();
1112}
1113
1114void OPPROTO op_ucomisd(void)
1115{
1116 int eflags;
1117 double d0, d1;
1118 Reg *d, *s;
1119 d = (Reg *)((char *)env + PARAM1);
1120 s = (Reg *)((char *)env + PARAM2);
1121
1122 d0 = d->XMM_D(0);
1123 d1 = s->XMM_D(0);
1124 if (d0 < d1)
1125 eflags = CC_C;
1126 else if (d0 == d1)
1127 eflags = CC_Z;
1128 else
1129 eflags = 0;
1130 CC_SRC = eflags;
1131 FORCE_RET();
1132}
1133
1134void OPPROTO op_comisd(void)
1135{
1136 int eflags;
1137 double d0, d1;
1138 Reg *d, *s;
1139 d = (Reg *)((char *)env + PARAM1);
1140 s = (Reg *)((char *)env + PARAM2);
1141
1142 d0 = d->XMM_D(0);
1143 d1 = s->XMM_D(0);
1144 if (d0 < d1)
1145 eflags = CC_C;
1146 else if (d0 == d1)
1147 eflags = CC_Z;
1148 else
1149 eflags = 0;
1150 CC_SRC = eflags;
1151 FORCE_RET();
1152}
1153
1154void OPPROTO op_movmskps(void)
1155{
1156 int b0, b1, b2, b3;
1157 Reg *s;
1158 s = (Reg *)((char *)env + PARAM1);
1159 b0 = s->XMM_L(0) >> 31;
1160 b1 = s->XMM_L(1) >> 31;
1161 b2 = s->XMM_L(2) >> 31;
1162 b3 = s->XMM_L(3) >> 31;
1163 T0 = b0 | (b1 << 1) | (b2 << 2) | (b3 << 3);
1164}
1165
1166void OPPROTO op_movmskpd(void)
1167{
1168 int b0, b1;
1169 Reg *s;
1170 s = (Reg *)((char *)env + PARAM1);
1171 b0 = s->XMM_L(1) >> 31;
1172 b1 = s->XMM_L(3) >> 31;
1173 T0 = b0 | (b1 << 1);
1174}
1175
1176#endif
1177
1178void OPPROTO glue(op_pmovmskb, SUFFIX)(void)
1179{
1180 Reg *s;
1181 s = (Reg *)((char *)env + PARAM1);
1182 T0 = 0;
1183 T0 |= (s->XMM_B(0) >> 7);
1184 T0 |= (s->XMM_B(1) >> 6) & 0x02;
1185 T0 |= (s->XMM_B(2) >> 5) & 0x04;
1186 T0 |= (s->XMM_B(3) >> 4) & 0x08;
1187 T0 |= (s->XMM_B(4) >> 3) & 0x10;
1188 T0 |= (s->XMM_B(5) >> 2) & 0x20;
1189 T0 |= (s->XMM_B(6) >> 1) & 0x40;
1190 T0 |= (s->XMM_B(7)) & 0x80;
1191#if SHIFT == 1
1192 T0 |= (s->XMM_B(8) << 1) & 0x0100;
1193 T0 |= (s->XMM_B(9) << 2) & 0x0200;
1194 T0 |= (s->XMM_B(10) << 3) & 0x0400;
1195 T0 |= (s->XMM_B(11) << 4) & 0x0800;
1196 T0 |= (s->XMM_B(12) << 5) & 0x1000;
1197 T0 |= (s->XMM_B(13) << 6) & 0x2000;
1198 T0 |= (s->XMM_B(14) << 7) & 0x4000;
1199 T0 |= (s->XMM_B(15) << 8) & 0x8000;
1200#endif
1201}
1202
1203void OPPROTO glue(op_pinsrw, SUFFIX) (void)
1204{
1205 Reg *d = (Reg *)((char *)env + PARAM1);
1206 int pos = PARAM2;
1207
1208 d->W(pos) = T0;
1209}
1210
1211void OPPROTO glue(op_pextrw, SUFFIX) (void)
1212{
1213 Reg *s = (Reg *)((char *)env + PARAM1);
1214 int pos = PARAM2;
1215
1216 T0 = s->W(pos);
1217}
1218
1219void OPPROTO glue(op_packsswb, SUFFIX) (void)
1220{
1221 Reg r, *d, *s;
1222 d = (Reg *)((char *)env + PARAM1);
1223 s = (Reg *)((char *)env + PARAM2);
1224
1225 r.B(0) = satsb((int16_t)d->W(0));
1226 r.B(1) = satsb((int16_t)d->W(1));
1227 r.B(2) = satsb((int16_t)d->W(2));
1228 r.B(3) = satsb((int16_t)d->W(3));
1229#if SHIFT == 1
1230 r.B(4) = satsb((int16_t)d->W(4));
1231 r.B(5) = satsb((int16_t)d->W(5));
1232 r.B(6) = satsb((int16_t)d->W(6));
1233 r.B(7) = satsb((int16_t)d->W(7));
1234#endif
1235 r.B((4 << SHIFT) + 0) = satsb((int16_t)s->W(0));
1236 r.B((4 << SHIFT) + 1) = satsb((int16_t)s->W(1));
1237 r.B((4 << SHIFT) + 2) = satsb((int16_t)s->W(2));
1238 r.B((4 << SHIFT) + 3) = satsb((int16_t)s->W(3));
1239#if SHIFT == 1
1240 r.B(12) = satsb((int16_t)s->W(4));
1241 r.B(13) = satsb((int16_t)s->W(5));
1242 r.B(14) = satsb((int16_t)s->W(6));
1243 r.B(15) = satsb((int16_t)s->W(7));
1244#endif
1245 *d = r;
1246}
1247
1248void OPPROTO glue(op_packuswb, SUFFIX) (void)
1249{
1250 Reg r, *d, *s;
1251 d = (Reg *)((char *)env + PARAM1);
1252 s = (Reg *)((char *)env + PARAM2);
1253
1254 r.B(0) = satub((int16_t)d->W(0));
1255 r.B(1) = satub((int16_t)d->W(1));
1256 r.B(2) = satub((int16_t)d->W(2));
1257 r.B(3) = satub((int16_t)d->W(3));
1258#if SHIFT == 1
1259 r.B(4) = satub((int16_t)d->W(4));
1260 r.B(5) = satub((int16_t)d->W(5));
1261 r.B(6) = satub((int16_t)d->W(6));
1262 r.B(7) = satub((int16_t)d->W(7));
1263#endif
1264 r.B((4 << SHIFT) + 0) = satub((int16_t)s->W(0));
1265 r.B((4 << SHIFT) + 1) = satub((int16_t)s->W(1));
1266 r.B((4 << SHIFT) + 2) = satub((int16_t)s->W(2));
1267 r.B((4 << SHIFT) + 3) = satub((int16_t)s->W(3));
1268#if SHIFT == 1
1269 r.B(12) = satub((int16_t)s->W(4));
1270 r.B(13) = satub((int16_t)s->W(5));
1271 r.B(14) = satub((int16_t)s->W(6));
1272 r.B(15) = satub((int16_t)s->W(7));
1273#endif
1274 *d = r;
1275}
1276
1277void OPPROTO glue(op_packssdw, SUFFIX) (void)
1278{
1279 Reg r, *d, *s;
1280 d = (Reg *)((char *)env + PARAM1);
1281 s = (Reg *)((char *)env + PARAM2);
1282
1283 r.W(0) = satsw(d->L(0));
1284 r.W(1) = satsw(d->L(1));
1285#if SHIFT == 1
1286 r.W(2) = satsw(d->L(2));
1287 r.W(3) = satsw(d->L(3));
1288#endif
1289 r.W((2 << SHIFT) + 0) = satsw(s->L(0));
1290 r.W((2 << SHIFT) + 1) = satsw(s->L(1));
1291#if SHIFT == 1
1292 r.W(6) = satsw(s->L(2));
1293 r.W(7) = satsw(s->L(3));
1294#endif
1295 *d = r;
1296}
1297
1298#define UNPCK_OP(base_name, base) \
1299 \
1300void OPPROTO glue(op_punpck ## base_name ## bw, SUFFIX) (void) \
1301{ \
1302 Reg r, *d, *s; \
1303 d = (Reg *)((char *)env + PARAM1); \
1304 s = (Reg *)((char *)env + PARAM2); \
1305 \
1306 r.B(0) = d->B((base << (SHIFT + 2)) + 0); \
1307 r.B(1) = s->B((base << (SHIFT + 2)) + 0); \
1308 r.B(2) = d->B((base << (SHIFT + 2)) + 1); \
1309 r.B(3) = s->B((base << (SHIFT + 2)) + 1); \
1310 r.B(4) = d->B((base << (SHIFT + 2)) + 2); \
1311 r.B(5) = s->B((base << (SHIFT + 2)) + 2); \
1312 r.B(6) = d->B((base << (SHIFT + 2)) + 3); \
1313 r.B(7) = s->B((base << (SHIFT + 2)) + 3); \
1314XMM_ONLY( \
1315 r.B(8) = d->B((base << (SHIFT + 2)) + 4); \
1316 r.B(9) = s->B((base << (SHIFT + 2)) + 4); \
1317 r.B(10) = d->B((base << (SHIFT + 2)) + 5); \
1318 r.B(11) = s->B((base << (SHIFT + 2)) + 5); \
1319 r.B(12) = d->B((base << (SHIFT + 2)) + 6); \
1320 r.B(13) = s->B((base << (SHIFT + 2)) + 6); \
1321 r.B(14) = d->B((base << (SHIFT + 2)) + 7); \
1322 r.B(15) = s->B((base << (SHIFT + 2)) + 7); \
1323) \
1324 *d = r; \
1325} \
1326 \
1327void OPPROTO glue(op_punpck ## base_name ## wd, SUFFIX) (void) \
1328{ \
1329 Reg r, *d, *s; \
1330 d = (Reg *)((char *)env + PARAM1); \
1331 s = (Reg *)((char *)env + PARAM2); \
1332 \
1333 r.W(0) = d->W((base << (SHIFT + 1)) + 0); \
1334 r.W(1) = s->W((base << (SHIFT + 1)) + 0); \
1335 r.W(2) = d->W((base << (SHIFT + 1)) + 1); \
1336 r.W(3) = s->W((base << (SHIFT + 1)) + 1); \
1337XMM_ONLY( \
1338 r.W(4) = d->W((base << (SHIFT + 1)) + 2); \
1339 r.W(5) = s->W((base << (SHIFT + 1)) + 2); \
1340 r.W(6) = d->W((base << (SHIFT + 1)) + 3); \
1341 r.W(7) = s->W((base << (SHIFT + 1)) + 3); \
1342) \
1343 *d = r; \
1344} \
1345 \
1346void OPPROTO glue(op_punpck ## base_name ## dq, SUFFIX) (void) \
1347{ \
1348 Reg r, *d, *s; \
1349 d = (Reg *)((char *)env + PARAM1); \
1350 s = (Reg *)((char *)env + PARAM2); \
1351 \
1352 r.L(0) = d->L((base << SHIFT) + 0); \
1353 r.L(1) = s->L((base << SHIFT) + 0); \
1354XMM_ONLY( \
1355 r.L(2) = d->L((base << SHIFT) + 1); \
1356 r.L(3) = s->L((base << SHIFT) + 1); \
1357) \
1358 *d = r; \
1359} \
1360 \
1361XMM_ONLY( \
1362void OPPROTO glue(op_punpck ## base_name ## qdq, SUFFIX) (void) \
1363{ \
1364 Reg r, *d, *s; \
1365 d = (Reg *)((char *)env + PARAM1); \
1366 s = (Reg *)((char *)env + PARAM2); \
1367 \
1368 r.Q(0) = d->Q(base); \
1369 r.Q(1) = s->Q(base); \
1370 *d = r; \
1371} \
1372)
1373
1374UNPCK_OP(l, 0)
1375UNPCK_OP(h, 1)
1376
1377#undef SHIFT
1378#undef XMM_ONLY
1379#undef Reg
1380#undef B
1381#undef W
1382#undef L
1383#undef Q
1384#undef SUFFIX
Qemu copy for testing