Merge commit 'origin/master' into gallium-0.2
[mesa.git] / src / gallium / auxiliary / rtasm / rtasm_ppc_spe.c
1 /*
2 * (C) Copyright IBM Corporation 2008
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 /**
26 * \file
27 * Real-time assembly generation interface for Cell B.E. SPEs.
28 *
29 * \author Ian Romanick <idr@us.ibm.com>
30 * \author Brian Paul
31 */
32
33
34 #include <stdio.h>
35 #include "pipe/p_compiler.h"
36 #include "util/u_memory.h"
37 #include "rtasm_ppc_spe.h"
38
39
40 #ifdef GALLIUM_CELL
41 /**
42 * SPE instruction types
43 *
44 * There are 6 primary instruction encodings used on the Cell's SPEs. Each of
45 * the following unions encodes one type.
46 *
47 * \bug
48 * If, at some point, we start generating SPE code from a little-endian host
49 * these unions will not work.
50 */
51 /*@{*/
52 /**
53 * Encode one output register with two input registers
54 */
55 union spe_inst_RR {
56 uint32_t bits;
57 struct {
58 unsigned op:11;
59 unsigned rB:7;
60 unsigned rA:7;
61 unsigned rT:7;
62 } inst;
63 };
64
65
66 /**
67 * Encode one output register with three input registers
68 */
69 union spe_inst_RRR {
70 uint32_t bits;
71 struct {
72 unsigned op:4;
73 unsigned rT:7;
74 unsigned rB:7;
75 unsigned rA:7;
76 unsigned rC:7;
77 } inst;
78 };
79
80
81 /**
82 * Encode one output register with one input reg. and a 7-bit signed immed
83 */
84 union spe_inst_RI7 {
85 uint32_t bits;
86 struct {
87 unsigned op:11;
88 unsigned i7:7;
89 unsigned rA:7;
90 unsigned rT:7;
91 } inst;
92 };
93
94
95 /**
96 * Encode one output register with one input reg. and an 8-bit signed immed
97 */
98 union spe_inst_RI8 {
99 uint32_t bits;
100 struct {
101 unsigned op:10;
102 unsigned i8:8;
103 unsigned rA:7;
104 unsigned rT:7;
105 } inst;
106 };
107
108
109 /**
110 * Encode one output register with one input reg. and a 10-bit signed immed
111 */
112 union spe_inst_RI10 {
113 uint32_t bits;
114 struct {
115 unsigned op:8;
116 unsigned i10:10;
117 unsigned rA:7;
118 unsigned rT:7;
119 } inst;
120 };
121
122
123 /**
124 * Encode one output register with a 16-bit signed immediate
125 */
126 union spe_inst_RI16 {
127 uint32_t bits;
128 struct {
129 unsigned op:9;
130 unsigned i16:16;
131 unsigned rT:7;
132 } inst;
133 };
134
135
136 /**
137 * Encode one output register with a 18-bit signed immediate
138 */
139 union spe_inst_RI18 {
140 uint32_t bits;
141 struct {
142 unsigned op:7;
143 unsigned i18:18;
144 unsigned rT:7;
145 } inst;
146 };
147 /*@}*/
148
149
150 static void
151 indent(const struct spe_function *p)
152 {
153 int i;
154 for (i = 0; i < p->indent; i++) {
155 putchar(' ');
156 }
157 }
158
159
160 static const char *
161 rem_prefix(const char *longname)
162 {
163 return longname + 4;
164 }
165
166
167 static const char *
168 reg_name(int reg)
169 {
170 switch (reg) {
171 case SPE_REG_SP:
172 return "$sp";
173 case SPE_REG_RA:
174 return "$lr";
175 default:
176 {
177 /* cycle through four buffers to handle multiple calls per printf */
178 static char buf[4][10];
179 static int b = 0;
180 b = (b + 1) % 4;
181 sprintf(buf[b], "$%d", reg);
182 return buf[b];
183 }
184 }
185 }
186
187
188 static void
189 emit_instruction(struct spe_function *p, uint32_t inst_bits)
190 {
191 if (!p->store)
192 return; /* out of memory, drop the instruction */
193
194 if (p->num_inst == p->max_inst) {
195 /* allocate larger buffer */
196 uint32_t *newbuf;
197 p->max_inst *= 2; /* 2x larger */
198 newbuf = align_malloc(p->max_inst * SPE_INST_SIZE, 16);
199 if (newbuf) {
200 memcpy(newbuf, p->store, p->num_inst * SPE_INST_SIZE);
201 }
202 align_free(p->store);
203 p->store = newbuf;
204 if (!p->store) {
205 /* out of memory */
206 p->num_inst = 0;
207 return;
208 }
209 }
210
211 p->store[p->num_inst++] = inst_bits;
212 }
213
214
215
216 static void emit_RR(struct spe_function *p, unsigned op, unsigned rT,
217 unsigned rA, unsigned rB, const char *name)
218 {
219 union spe_inst_RR inst;
220 inst.inst.op = op;
221 inst.inst.rB = rB;
222 inst.inst.rA = rA;
223 inst.inst.rT = rT;
224 emit_instruction(p, inst.bits);
225 if (p->print) {
226 indent(p);
227 printf("%s\t%s, %s, %s\n",
228 rem_prefix(name), reg_name(rT), reg_name(rA), reg_name(rB));
229 }
230 }
231
232
233 static void emit_RRR(struct spe_function *p, unsigned op, unsigned rT,
234 unsigned rA, unsigned rB, unsigned rC, const char *name)
235 {
236 union spe_inst_RRR inst;
237 inst.inst.op = op;
238 inst.inst.rT = rT;
239 inst.inst.rB = rB;
240 inst.inst.rA = rA;
241 inst.inst.rC = rC;
242 emit_instruction(p, inst.bits);
243 if (p->print) {
244 indent(p);
245 printf("%s\t%s, %s, %s, %s\n", rem_prefix(name), reg_name(rT),
246 reg_name(rA), reg_name(rB), reg_name(rC));
247 }
248 }
249
250
251 static void emit_RI7(struct spe_function *p, unsigned op, unsigned rT,
252 unsigned rA, int imm, const char *name)
253 {
254 union spe_inst_RI7 inst;
255 inst.inst.op = op;
256 inst.inst.i7 = imm;
257 inst.inst.rA = rA;
258 inst.inst.rT = rT;
259 emit_instruction(p, inst.bits);
260 if (p->print) {
261 indent(p);
262 printf("%s\t%s, %s, 0x%x\n",
263 rem_prefix(name), reg_name(rT), reg_name(rA), imm);
264 }
265 }
266
267
268
269 static void emit_RI8(struct spe_function *p, unsigned op, unsigned rT,
270 unsigned rA, int imm, const char *name)
271 {
272 union spe_inst_RI8 inst;
273 inst.inst.op = op;
274 inst.inst.i8 = imm;
275 inst.inst.rA = rA;
276 inst.inst.rT = rT;
277 emit_instruction(p, inst.bits);
278 if (p->print) {
279 indent(p);
280 printf("%s\t%s, %s, 0x%x\n",
281 rem_prefix(name), reg_name(rT), reg_name(rA), imm);
282 }
283 }
284
285
286
287 static void emit_RI10(struct spe_function *p, unsigned op, unsigned rT,
288 unsigned rA, int imm, const char *name)
289 {
290 union spe_inst_RI10 inst;
291 inst.inst.op = op;
292 inst.inst.i10 = imm;
293 inst.inst.rA = rA;
294 inst.inst.rT = rT;
295 emit_instruction(p, inst.bits);
296 if (p->print) {
297 indent(p);
298 printf("%s\t%s, %s, 0x%x\n",
299 rem_prefix(name), reg_name(rT), reg_name(rA), imm);
300 }
301 }
302
303
304 /** As above, but do range checking on signed immediate value */
305 static void emit_RI10s(struct spe_function *p, unsigned op, unsigned rT,
306 unsigned rA, int imm, const char *name)
307 {
308 assert(imm <= 511);
309 assert(imm >= -512);
310 emit_RI10(p, op, rT, rA, imm, name);
311 }
312
313
314 static void emit_RI16(struct spe_function *p, unsigned op, unsigned rT,
315 int imm, const char *name)
316 {
317 union spe_inst_RI16 inst;
318 inst.inst.op = op;
319 inst.inst.i16 = imm;
320 inst.inst.rT = rT;
321 emit_instruction(p, inst.bits);
322 if (p->print) {
323 indent(p);
324 printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm);
325 }
326 }
327
328
329 static void emit_RI18(struct spe_function *p, unsigned op, unsigned rT,
330 int imm, const char *name)
331 {
332 union spe_inst_RI18 inst;
333 inst.inst.op = op;
334 inst.inst.i18 = imm;
335 inst.inst.rT = rT;
336 emit_instruction(p, inst.bits);
337 if (p->print) {
338 indent(p);
339 printf("%s\t%s, 0x%x\n", rem_prefix(name), reg_name(rT), imm);
340 }
341 }
342
343
344
345
346 #define EMIT_(_name, _op) \
347 void _name (struct spe_function *p, unsigned rT) \
348 { \
349 emit_RR(p, _op, rT, 0, 0, __FUNCTION__); \
350 }
351
352 #define EMIT_R(_name, _op) \
353 void _name (struct spe_function *p, unsigned rT, unsigned rA) \
354 { \
355 emit_RR(p, _op, rT, rA, 0, __FUNCTION__); \
356 }
357
358 #define EMIT_RR(_name, _op) \
359 void _name (struct spe_function *p, unsigned rT, unsigned rA, unsigned rB) \
360 { \
361 emit_RR(p, _op, rT, rA, rB, __FUNCTION__); \
362 }
363
364 #define EMIT_RRR(_name, _op) \
365 void _name (struct spe_function *p, unsigned rT, unsigned rA, unsigned rB, unsigned rC) \
366 { \
367 emit_RRR(p, _op, rT, rA, rB, rC, __FUNCTION__); \
368 }
369
370 #define EMIT_RI7(_name, _op) \
371 void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \
372 { \
373 emit_RI7(p, _op, rT, rA, imm, __FUNCTION__); \
374 }
375
376 #define EMIT_RI8(_name, _op, bias) \
377 void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \
378 { \
379 emit_RI8(p, _op, rT, rA, bias - imm, __FUNCTION__); \
380 }
381
382 #define EMIT_RI10(_name, _op) \
383 void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \
384 { \
385 emit_RI10(p, _op, rT, rA, imm, __FUNCTION__); \
386 }
387
388 #define EMIT_RI10s(_name, _op) \
389 void _name (struct spe_function *p, unsigned rT, unsigned rA, int imm) \
390 { \
391 emit_RI10s(p, _op, rT, rA, imm, __FUNCTION__); \
392 }
393
394 #define EMIT_RI16(_name, _op) \
395 void _name (struct spe_function *p, unsigned rT, int imm) \
396 { \
397 emit_RI16(p, _op, rT, imm, __FUNCTION__); \
398 }
399
400 #define EMIT_RI18(_name, _op) \
401 void _name (struct spe_function *p, unsigned rT, int imm) \
402 { \
403 emit_RI18(p, _op, rT, imm, __FUNCTION__); \
404 }
405
406 #define EMIT_I16(_name, _op) \
407 void _name (struct spe_function *p, int imm) \
408 { \
409 emit_RI16(p, _op, 0, imm, __FUNCTION__); \
410 }
411
412 #include "rtasm_ppc_spe.h"
413
414
415
416 /**
417 * Initialize an spe_function.
418 * \param code_size initial size of instruction buffer to allocate, in bytes.
419 * If zero, use a default.
420 */
421 void spe_init_func(struct spe_function *p, unsigned code_size)
422 {
423 unsigned int i;
424
425 if (!code_size)
426 code_size = 64;
427
428 p->num_inst = 0;
429 p->max_inst = code_size / SPE_INST_SIZE;
430 p->store = align_malloc(code_size, 16);
431
432 p->set_count = 0;
433 memset(p->regs, 0, SPE_NUM_REGS * sizeof(p->regs[0]));
434
435 /* Conservatively treat R0 - R2 and R80 - R127 as non-volatile.
436 */
437 p->regs[0] = p->regs[1] = p->regs[2] = 1;
438 for (i = 80; i <= 127; i++) {
439 p->regs[i] = 1;
440 }
441
442 p->print = false;
443 p->indent = 0;
444 }
445
446
447 void spe_release_func(struct spe_function *p)
448 {
449 assert(p->num_inst <= p->max_inst);
450 if (p->store != NULL) {
451 align_free(p->store);
452 }
453 p->store = NULL;
454 }
455
456
457 /** Return current code size in bytes. */
458 unsigned spe_code_size(const struct spe_function *p)
459 {
460 return p->num_inst * SPE_INST_SIZE;
461 }
462
463
464 /**
465 * Allocate a SPE register.
466 * \return register index or -1 if none left.
467 */
468 int spe_allocate_available_register(struct spe_function *p)
469 {
470 unsigned i;
471 for (i = 0; i < SPE_NUM_REGS; i++) {
472 if (p->regs[i] == 0) {
473 p->regs[i] = 1;
474 return i;
475 }
476 }
477
478 return -1;
479 }
480
481
482 /**
483 * Mark the given SPE register as "allocated".
484 */
485 int spe_allocate_register(struct spe_function *p, int reg)
486 {
487 assert(reg < SPE_NUM_REGS);
488 assert(p->regs[reg] == 0);
489 p->regs[reg] = 1;
490 return reg;
491 }
492
493
494 /**
495 * Mark the given SPE register as "unallocated". Note that this should
496 * only be used on registers allocated in the current register set; an
497 * assertion will fail if an attempt is made to deallocate a register
498 * allocated in an earlier register set.
499 */
500 void spe_release_register(struct spe_function *p, int reg)
501 {
502 assert(reg < SPE_NUM_REGS);
503 assert(p->regs[reg] == 1);
504
505 p->regs[reg] = 0;
506 }
507
508 /**
509 * Start a new set of registers. This can be called if
510 * it will be difficult later to determine exactly what
511 * registers were actually allocated during a code generation
512 * sequence, and you really just want to deallocate all of them.
513 */
514 void spe_allocate_register_set(struct spe_function *p)
515 {
516 unsigned int i;
517
518 /* Keep track of the set count. If it ever wraps around to 0,
519 * we're in trouble.
520 */
521 p->set_count++;
522 assert(p->set_count > 0);
523
524 /* Increment the allocation count of all registers currently
525 * allocated. Then any registers that are allocated in this set
526 * will be the only ones with a count of 1; they'll all be released
527 * when the register set is released.
528 */
529 for (i = 0; i < SPE_NUM_REGS; i++) {
530 if (p->regs[i] > 0)
531 p->regs[i]++;
532 }
533 }
534
535 void spe_release_register_set(struct spe_function *p)
536 {
537 unsigned int i;
538
539 /* If the set count drops below zero, we're in trouble. */
540 assert(p->set_count > 0);
541 p->set_count--;
542
543 /* Drop the allocation level of all registers. Any allocated
544 * during this register set will drop to 0 and then become
545 * available.
546 */
547 for (i = 0; i < SPE_NUM_REGS; i++) {
548 if (p->regs[i] > 0)
549 p->regs[i]--;
550 }
551 }
552
553
554 unsigned
555 spe_get_registers_used(const struct spe_function *p, ubyte used[])
556 {
557 unsigned i, num = 0;
558 /* only count registers in the range available to callers */
559 for (i = 2; i < 80; i++) {
560 if (p->regs[i]) {
561 used[num++] = i;
562 }
563 }
564 return num;
565 }
566
567
568 void
569 spe_print_code(struct spe_function *p, boolean enable)
570 {
571 p->print = enable;
572 }
573
574
575 void
576 spe_indent(struct spe_function *p, int spaces)
577 {
578 p->indent += spaces;
579 }
580
581
582 void
583 spe_comment(struct spe_function *p, int rel_indent, const char *s)
584 {
585 if (p->print) {
586 p->indent += rel_indent;
587 indent(p);
588 p->indent -= rel_indent;
589 printf("# %s\n", s);
590 }
591 }
592
593
594 /**
595 * Load quad word.
596 * NOTE: offset is in bytes and the least significant 4 bits must be zero!
597 */
598 void spe_lqd(struct spe_function *p, unsigned rT, unsigned rA, int offset)
599 {
600 const boolean pSave = p->print;
601
602 /* offset must be a multiple of 16 */
603 assert(offset % 16 == 0);
604 /* offset must fit in 10-bit signed int field, after shifting */
605 assert((offset >> 4) <= 511);
606 assert((offset >> 4) >= -512);
607
608 p->print = FALSE;
609 emit_RI10(p, 0x034, rT, rA, offset >> 4, "spe_lqd");
610 p->print = pSave;
611
612 if (p->print) {
613 indent(p);
614 printf("lqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA));
615 }
616 }
617
618
619 /**
620 * Store quad word.
621 * NOTE: offset is in bytes and the least significant 4 bits must be zero!
622 */
623 void spe_stqd(struct spe_function *p, unsigned rT, unsigned rA, int offset)
624 {
625 const boolean pSave = p->print;
626
627 /* offset must be a multiple of 16 */
628 assert(offset % 16 == 0);
629 /* offset must fit in 10-bit signed int field, after shifting */
630 assert((offset >> 4) <= 511);
631 assert((offset >> 4) >= -512);
632
633 p->print = FALSE;
634 emit_RI10(p, 0x024, rT, rA, offset >> 4, "spe_stqd");
635 p->print = pSave;
636
637 if (p->print) {
638 indent(p);
639 printf("stqd\t%s, %d(%s)\n", reg_name(rT), offset, reg_name(rA));
640 }
641 }
642
643
644 /**
645 * For branch instructions:
646 * \param d if 1, disable interupts if branch is taken
647 * \param e if 1, enable interupts if branch is taken
648 * If d and e are both zero, don't change interupt status (right?)
649 */
650
651 /** Branch Indirect to address in rA */
652 void spe_bi(struct spe_function *p, unsigned rA, int d, int e)
653 {
654 emit_RI7(p, 0x1a8, 0, rA, (d << 5) | (e << 4), __FUNCTION__);
655 }
656
657 /** Interupt Return */
658 void spe_iret(struct spe_function *p, unsigned rA, int d, int e)
659 {
660 emit_RI7(p, 0x1aa, 0, rA, (d << 5) | (e << 4), __FUNCTION__);
661 }
662
663 /** Branch indirect and set link on external data */
664 void spe_bisled(struct spe_function *p, unsigned rT, unsigned rA, int d,
665 int e)
666 {
667 emit_RI7(p, 0x1ab, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
668 }
669
670 /** Branch indirect and set link. Save PC in rT, jump to rA. */
671 void spe_bisl(struct spe_function *p, unsigned rT, unsigned rA, int d,
672 int e)
673 {
674 emit_RI7(p, 0x1a9, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
675 }
676
677 /** Branch indirect if zero word. If rT.word[0]==0, jump to rA. */
678 void spe_biz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
679 {
680 emit_RI7(p, 0x128, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
681 }
682
683 /** Branch indirect if non-zero word. If rT.word[0]!=0, jump to rA. */
684 void spe_binz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
685 {
686 emit_RI7(p, 0x129, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
687 }
688
689 /** Branch indirect if zero halfword. If rT.halfword[1]==0, jump to rA. */
690 void spe_bihz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
691 {
692 emit_RI7(p, 0x12a, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
693 }
694
695 /** Branch indirect if non-zero halfword. If rT.halfword[1]!=0, jump to rA. */
696 void spe_bihnz(struct spe_function *p, unsigned rT, unsigned rA, int d, int e)
697 {
698 emit_RI7(p, 0x12b, rT, rA, (d << 5) | (e << 4), __FUNCTION__);
699 }
700
701
702 /* Hint-for-branch instructions
703 */
704 #if 0
705 hbr;
706 hbra;
707 hbrr;
708 #endif
709
710
711 /* Control instructions
712 */
713 #if 0
714 stop;
715 EMIT_RR (spe_stopd, 0x140);
716 EMIT_ (spe_lnop, 0x001);
717 EMIT_ (spe_nop, 0x201);
718 sync;
719 EMIT_ (spe_dsync, 0x003);
720 EMIT_R (spe_mfspr, 0x00c);
721 EMIT_R (spe_mtspr, 0x10c);
722 #endif
723
724
725 /**
726 ** Helper / "macro" instructions.
727 ** Use somewhat verbose names as a reminder that these aren't native
728 ** SPE instructions.
729 **/
730
731
732 void
733 spe_load_float(struct spe_function *p, unsigned rT, float x)
734 {
735 if (x == 0.0f) {
736 spe_il(p, rT, 0x0);
737 }
738 else if (x == 0.5f) {
739 spe_ilhu(p, rT, 0x3f00);
740 }
741 else if (x == 1.0f) {
742 spe_ilhu(p, rT, 0x3f80);
743 }
744 else if (x == -1.0f) {
745 spe_ilhu(p, rT, 0xbf80);
746 }
747 else {
748 union {
749 float f;
750 unsigned u;
751 } bits;
752 bits.f = x;
753 spe_ilhu(p, rT, bits.u >> 16);
754 spe_iohl(p, rT, bits.u & 0xffff);
755 }
756 }
757
758
759 void
760 spe_load_int(struct spe_function *p, unsigned rT, int i)
761 {
762 if (-32768 <= i && i <= 32767) {
763 spe_il(p, rT, i);
764 }
765 else {
766 spe_ilhu(p, rT, i >> 16);
767 if (i & 0xffff)
768 spe_iohl(p, rT, i & 0xffff);
769 }
770 }
771
772 void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui)
773 {
774 /* If the whole value is in the lower 18 bits, use ila, which
775 * doesn't sign-extend. Otherwise, if the two halfwords of
776 * the constant are identical, use ilh. Otherwise, if every byte of
777 * the desired value is 0x00 or 0xff, we can use Form Select Mask for
778 * Bytes Immediate (fsmbi) to load the value in a single instruction.
779 * Otherwise, in the general case, we have to use ilhu followed by iohl.
780 */
781 if ((ui & 0x0003ffff) == ui) {
782 spe_ila(p, rT, ui);
783 }
784 else if ((ui >> 16) == (ui & 0xffff)) {
785 spe_ilh(p, rT, ui & 0xffff);
786 }
787 else if (
788 ((ui & 0x000000ff) == 0 || (ui & 0x000000ff) == 0x000000ff) &&
789 ((ui & 0x0000ff00) == 0 || (ui & 0x0000ff00) == 0x0000ff00) &&
790 ((ui & 0x00ff0000) == 0 || (ui & 0x00ff0000) == 0x00ff0000) &&
791 ((ui & 0xff000000) == 0 || (ui & 0xff000000) == 0xff000000)
792 ) {
793 unsigned int mask = 0;
794 /* fsmbi duplicates each bit in the given mask eight times,
795 * using a 16-bit value to initialize a 16-byte quadword.
796 * Each 4-bit nybble of the mask corresponds to a full word
797 * of the result; look at the value and figure out the mask
798 * (replicated for each word in the quadword), and then
799 * form the "select mask" to get the value.
800 */
801 if ((ui & 0x000000ff) == 0x000000ff) mask |= 0x1111;
802 if ((ui & 0x0000ff00) == 0x0000ff00) mask |= 0x2222;
803 if ((ui & 0x00ff0000) == 0x00ff0000) mask |= 0x4444;
804 if ((ui & 0xff000000) == 0xff000000) mask |= 0x8888;
805 spe_fsmbi(p, rT, mask);
806 }
807 else {
808 /* The general case: this usually uses two instructions, but
809 * may use only one if the low-order 16 bits of each word are 0.
810 */
811 spe_ilhu(p, rT, ui >> 16);
812 if (ui & 0xffff)
813 spe_iohl(p, rT, ui & 0xffff);
814 }
815 }
816
817 /**
818 * This function is constructed identically to spe_xor_uint() below.
819 * Changes to one should be made in the other.
820 */
821 void
822 spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
823 {
824 /* If we can, emit a single instruction, either And Byte Immediate
825 * (which uses the same constant across each byte), And Halfword Immediate
826 * (which sign-extends a 10-bit immediate to 16 bits and uses that
827 * across each halfword), or And Word Immediate (which sign-extends
828 * a 10-bit immediate to 32 bits).
829 *
830 * Otherwise, we'll need to use a temporary register.
831 */
832 unsigned int tmp;
833
834 /* If the upper 23 bits are all 0s or all 1s, sign extension
835 * will work and we can use And Word Immediate
836 */
837 tmp = ui & 0xfffffe00;
838 if (tmp == 0xfffffe00 || tmp == 0) {
839 spe_andi(p, rT, rA, ui & 0x000003ff);
840 return;
841 }
842
843 /* If the ui field is symmetric along halfword boundaries and
844 * the upper 7 bits of each halfword are all 0s or 1s, we
845 * can use And Halfword Immediate
846 */
847 tmp = ui & 0xfe00fe00;
848 if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) {
849 spe_andhi(p, rT, rA, ui & 0x000003ff);
850 return;
851 }
852
853 /* If the ui field is symmetric in each byte, then we can use
854 * the And Byte Immediate instruction.
855 */
856 tmp = ui & 0x000000ff;
857 if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) {
858 spe_andbi(p, rT, rA, tmp);
859 return;
860 }
861
862 /* Otherwise, we'll have to use a temporary register. */
863 unsigned int tmp_reg = spe_allocate_available_register(p);
864 spe_load_uint(p, tmp_reg, ui);
865 spe_and(p, rT, rA, tmp_reg);
866 spe_release_register(p, tmp_reg);
867 }
868
869
870 /**
871 * This function is constructed identically to spe_and_uint() above.
872 * Changes to one should be made in the other.
873 */
874 void
875 spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
876 {
877 /* If we can, emit a single instruction, either Exclusive Or Byte
878 * Immediate (which uses the same constant across each byte), Exclusive
879 * Or Halfword Immediate (which sign-extends a 10-bit immediate to
880 * 16 bits and uses that across each halfword), or Exclusive Or Word
881 * Immediate (which sign-extends a 10-bit immediate to 32 bits).
882 *
883 * Otherwise, we'll need to use a temporary register.
884 */
885 unsigned int tmp;
886
887 /* If the upper 23 bits are all 0s or all 1s, sign extension
888 * will work and we can use Exclusive Or Word Immediate
889 */
890 tmp = ui & 0xfffffe00;
891 if (tmp == 0xfffffe00 || tmp == 0) {
892 spe_xori(p, rT, rA, ui & 0x000003ff);
893 return;
894 }
895
896 /* If the ui field is symmetric along halfword boundaries and
897 * the upper 7 bits of each halfword are all 0s or 1s, we
898 * can use Exclusive Or Halfword Immediate
899 */
900 tmp = ui & 0xfe00fe00;
901 if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) {
902 spe_xorhi(p, rT, rA, ui & 0x000003ff);
903 return;
904 }
905
906 /* If the ui field is symmetric in each byte, then we can use
907 * the Exclusive Or Byte Immediate instruction.
908 */
909 tmp = ui & 0x000000ff;
910 if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) {
911 spe_xorbi(p, rT, rA, tmp);
912 return;
913 }
914
915 /* Otherwise, we'll have to use a temporary register. */
916 unsigned int tmp_reg = spe_allocate_available_register(p);
917 spe_load_uint(p, tmp_reg, ui);
918 spe_xor(p, rT, rA, tmp_reg);
919 spe_release_register(p, tmp_reg);
920 }
921
922 void
923 spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
924 {
925 /* If the comparison value is 9 bits or less, it fits inside a
926 * Compare Equal Word Immediate instruction.
927 */
928 if ((ui & 0x000001ff) == ui) {
929 spe_ceqi(p, rT, rA, ui);
930 }
931 /* Otherwise, we're going to have to load a word first. */
932 else {
933 unsigned int tmp_reg = spe_allocate_available_register(p);
934 spe_load_uint(p, tmp_reg, ui);
935 spe_ceq(p, rT, rA, tmp_reg);
936 spe_release_register(p, tmp_reg);
937 }
938 }
939
940 void
941 spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui)
942 {
943 /* If the comparison value is 10 bits or less, it fits inside a
944 * Compare Logical Greater Than Word Immediate instruction.
945 */
946 if ((ui & 0x000003ff) == ui) {
947 spe_clgti(p, rT, rA, ui);
948 }
949 /* Otherwise, we're going to have to load a word first. */
950 else {
951 unsigned int tmp_reg = spe_allocate_available_register(p);
952 spe_load_uint(p, tmp_reg, ui);
953 spe_clgt(p, rT, rA, tmp_reg);
954 spe_release_register(p, tmp_reg);
955 }
956 }
957
958 void
959 spe_splat(struct spe_function *p, unsigned rT, unsigned rA)
960 {
961 /* Use a temporary, just in case rT == rA */
962 unsigned int tmp_reg = spe_allocate_available_register(p);
963 /* Duplicate bytes 0, 1, 2, and 3 across the whole register */
964 spe_ila(p, tmp_reg, 0x00010203);
965 spe_shufb(p, rT, rA, rA, tmp_reg);
966 spe_release_register(p, tmp_reg);
967 }
968
969
970 void
971 spe_complement(struct spe_function *p, unsigned rT, unsigned rA)
972 {
973 spe_nor(p, rT, rA, rA);
974 }
975
976
977 void
978 spe_move(struct spe_function *p, unsigned rT, unsigned rA)
979 {
980 /* Use different instructions depending on the instruction address
981 * to take advantage of the dual pipelines.
982 */
983 if (p->num_inst & 1)
984 spe_shlqbyi(p, rT, rA, 0); /* odd pipe */
985 else
986 spe_ori(p, rT, rA, 0); /* even pipe */
987 }
988
989
990 void
991 spe_zero(struct spe_function *p, unsigned rT)
992 {
993 spe_xor(p, rT, rT, rT);
994 }
995
996
997 void
998 spe_splat_word(struct spe_function *p, unsigned rT, unsigned rA, int word)
999 {
1000 assert(word >= 0);
1001 assert(word <= 3);
1002
1003 if (word == 0) {
1004 int tmp1 = rT;
1005 spe_ila(p, tmp1, 66051);
1006 spe_shufb(p, rT, rA, rA, tmp1);
1007 }
1008 else {
1009 /* XXX review this, we may not need the rotqbyi instruction */
1010 int tmp1 = rT;
1011 int tmp2 = spe_allocate_available_register(p);
1012
1013 spe_ila(p, tmp1, 66051);
1014 spe_rotqbyi(p, tmp2, rA, 4 * word);
1015 spe_shufb(p, rT, tmp2, tmp2, tmp1);
1016
1017 spe_release_register(p, tmp2);
1018 }
1019 }
1020
1021 /**
1022 * For each 32-bit float element of rA and rB, choose the smaller of the
1023 * two, compositing them into the rT register.
1024 *
1025 * The Float Compare Greater Than (fcgt) instruction will put 1s into
1026 * compare_reg where rA > rB, and 0s where rA <= rB.
1027 *
1028 * Then the Select Bits (selb) instruction will take bits from rA where
1029 * compare_reg is 0, and from rB where compare_reg is 1; i.e., from rA
1030 * where rA <= rB and from rB where rB > rA, which is exactly the
1031 * "min" operation.
1032 *
1033 * The compare_reg could in many cases be the same as rT, unless
1034 * rT == rA || rt == rB. But since this is common in constructions
1035 * like "x = min(x, a)", we always allocate a new register to be safe.
1036 */
1037 void
1038 spe_float_min(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB)
1039 {
1040 unsigned int compare_reg = spe_allocate_available_register(p);
1041 spe_fcgt(p, compare_reg, rA, rB);
1042 spe_selb(p, rT, rA, rB, compare_reg);
1043 spe_release_register(p, compare_reg);
1044 }
1045
1046 /**
1047 * For each 32-bit float element of rA and rB, choose the greater of the
1048 * two, compositing them into the rT register.
1049 *
1050 * The logic is similar to that of spe_float_min() above; the only
1051 * difference is that the registers on spe_selb() have been reversed,
1052 * so that the larger of the two is selected instead of the smaller.
1053 */
1054 void
1055 spe_float_max(struct spe_function *p, unsigned rT, unsigned rA, unsigned rB)
1056 {
1057 unsigned int compare_reg = spe_allocate_available_register(p);
1058 spe_fcgt(p, compare_reg, rA, rB);
1059 spe_selb(p, rT, rB, rA, compare_reg);
1060 spe_release_register(p, compare_reg);
1061 }
1062
1063 #endif /* GALLIUM_CELL */