?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
glsl2: Add and use new variable mode ir_var_temporary
[mesa/nouveau-pmpeg.git] / src / gallium / drivers / nv50 / nv50_program.c
blob8cb1639013e8ef03501096816bb258c1009c1b24
1 /*
2 * Copyright 2008 Ben Skeggs
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is 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
17 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
20 * SOFTWARE.
21 */
22
23 #include "pipe/p_context.h"
24 #include "pipe/p_defines.h"
25 #include "pipe/p_state.h"
26 #include "util/u_inlines.h"
27
28 #include "pipe/p_shader_tokens.h"
29 #include "tgsi/tgsi_parse.h"
30 #include "tgsi/tgsi_util.h"
31
32 #include "nv50_context.h"
33 #include "nv50_transfer.h"
34
35 #define NV50_SU_MAX_TEMP 127
36 #define NV50_SU_MAX_ADDR 4
37 //#define NV50_PROGRAM_DUMP
38
39 /* $a5 and $a6 always seem to be 0, and using $a7 gives you noise */
40
41 /* ARL - gallium craps itself on progs/vp/arl.txt
42 *
43 * MSB - Like MAD, but MUL+SUB
44 * - Fuck it off, introduce a way to negate args for ops that
45 * support it.
46 *
47 * Look into inlining IMMD for ops other than MOV (make it general?)
48 * - Maybe even relax restrictions a bit, can't do P_RESULT + P_IMMD,
49 * but can emit to P_TEMP first - then MOV later. NVIDIA does this
50 *
51 * In ops such as ADD it's possible to construct a bad opcode in the !is_long()
52 * case, if the emit_src() causes the inst to suddenly become long.
53 *
54 * Verify half-insns work where expected - and force disable them where they
55 * don't work - MUL has it forcibly disabled atm as it fixes POW..
56 *
57 * FUCK! watch dst==src vectors, can overwrite components that are needed.
58 * ie. SUB R0, R0.yzxw, R0
59 *
60 * Things to check with renouveau:
61 * FP attr/result assignment - how?
62 * attrib
63 * - 0x16bc maps vp output onto fp hpos
64 * - 0x16c0 maps vp output onto fp col0
65 * result
66 * - colr always 0-3
67 * - depr always 4
68 * 0x16bc->0x16e8 --> some binding between vp/fp regs
69 * 0x16b8 --> VP output count
70 *
71 * 0x1298 --> "MOV rcol.x, fcol.y" "MOV depr, fcol.y" = 0x00000005
72 * "MOV rcol.x, fcol.y" = 0x00000004
73 * 0x19a8 --> as above but 0x00000100 and 0x00000000
74 * - 0x00100000 used when KIL used
75 * 0x196c --> as above but 0x00000011 and 0x00000000
76 *
77 * 0x1988 --> 0xXXNNNNNN
78 * - XX == FP high something
79 */
80 struct nv50_reg {
81 enum {
82 P_TEMP,
83 P_ATTR,
84 P_RESULT,
85 P_CONST,
86 P_IMMD,
87 P_ADDR
88 } type;
89 int index;
90
91 int hw;
92 int mod;
93
94 int rhw; /* result hw for FP outputs, or interpolant index */
95 int acc; /* instruction where this reg is last read (first insn == 1) */
96
97 int vtx; /* vertex index, for GP inputs (TGSI Dimension.Index) */
98 int indirect[2]; /* index into pc->addr, or -1 */
99
100 ubyte buf_index; /* c{0 .. 15}[] or g{0 .. 15}[] */
101 };
102
103 #define NV50_MOD_NEG 1
104 #define NV50_MOD_ABS 2
105 #define NV50_MOD_NEG_ABS (NV50_MOD_NEG | NV50_MOD_ABS)
106 #define NV50_MOD_SAT 4
107 #define NV50_MOD_I32 8
108
109 /* NV50_MOD_I32 is used to indicate integer mode for neg/abs */
110
111 /* STACK: Conditionals and loops have to use the (per warp) stack.
112 * Stack entries consist of an entry type (divergent path, join at),
113 * a mask indicating the active threads of the warp, and an address.
114 * MPs can store 12 stack entries internally, if we need more (and
115 * we probably do), we have to create a stack buffer in VRAM.
116 */
117 /* impose low limits for now */
118 #define NV50_MAX_COND_NESTING 4
119 #define NV50_MAX_LOOP_NESTING 3
120
121 #define JOIN_ON(e) e; pc->p->exec_tail->inst[1] |= 2
122
123 struct nv50_pc {
124 struct nv50_program *p;
125
126 /* hw resources */
127 struct nv50_reg *r_temp[NV50_SU_MAX_TEMP];
128 struct nv50_reg r_addr[NV50_SU_MAX_ADDR];
129
130 /* tgsi resources */
131 struct nv50_reg *temp;
132 int temp_nr;
133 struct nv50_reg *attr;
134 int attr_nr;
135 struct nv50_reg *result;
136 int result_nr;
137 struct nv50_reg *param;
138 int param_nr;
139 struct nv50_reg *immd;
140 uint32_t *immd_buf;
141 int immd_nr;
142 struct nv50_reg **addr;
143 int addr_nr;
144 struct nv50_reg *sysval;
145 int sysval_nr;
146
147 struct nv50_reg *temp_temp[16];
148 struct nv50_program_exec *temp_temp_exec[16];
149 unsigned temp_temp_nr;
150
151 /* broadcast and destination replacement regs */
152 struct nv50_reg *r_brdc;
153 struct nv50_reg *r_dst[4];
154
155 struct nv50_reg reg_instances[16];
156 unsigned reg_instance_nr;
157
158 unsigned interp_mode[32];
159 /* perspective interpolation registers */
160 struct nv50_reg *iv_p;
161 struct nv50_reg *iv_c;
162
163 struct nv50_program_exec *if_insn[NV50_MAX_COND_NESTING];
164 struct nv50_program_exec *if_join[NV50_MAX_COND_NESTING];
165 struct nv50_program_exec *loop_brka[NV50_MAX_LOOP_NESTING];
166 int if_lvl, loop_lvl;
167 unsigned loop_pos[NV50_MAX_LOOP_NESTING];
168
169 unsigned *insn_pos; /* actual program offset of each TGSI insn */
170 boolean in_subroutine;
171
172 /* current instruction and total number of insns */
173 unsigned insn_cur;
174 unsigned insn_nr;
175
176 boolean allow32;
177
178 uint8_t edgeflag_out;
179 };
180
181 static struct nv50_reg *get_address_reg(struct nv50_pc *, struct nv50_reg *);
182
183 static INLINE void
184 ctor_reg(struct nv50_reg *reg, unsigned type, int index, int hw)
185 {
186 reg->type = type;
187 reg->index = index;
188 reg->hw = hw;
189 reg->mod = 0;
190 reg->rhw = -1;
191 reg->vtx = -1;
192 reg->acc = 0;
193 reg->indirect[0] = reg->indirect[1] = -1;
194 reg->buf_index = (type == P_CONST) ? 1 : 0;
195 }
196
197 static INLINE unsigned
198 popcnt4(uint32_t val)
199 {
200 static const unsigned cnt[16]
201 = { 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4 };
202 return cnt[val & 0xf];
203 }
204
205 static void
206 terminate_mbb(struct nv50_pc *pc)
207 {
208 int i;
209
210 /* remove records of temporary address register values */
211 for (i = 0; i < NV50_SU_MAX_ADDR; ++i)
212 if (pc->r_addr[i].index < 0)
213 pc->r_addr[i].acc = 0;
214 }
215
216 static void
217 alloc_reg(struct nv50_pc *pc, struct nv50_reg *reg)
218 {
219 int i = 0;
220
221 if (reg->type == P_RESULT) {
222 if (pc->p->cfg.high_result < (reg->hw + 1))
223 pc->p->cfg.high_result = reg->hw + 1;
224 }
225
226 if (reg->type != P_TEMP)
227 return;
228
229 if (reg->hw >= 0) {
230 /*XXX: do this here too to catch FP temp-as-attr usage..
231 * not clean, but works */
232 if (pc->p->cfg.high_temp < (reg->hw + 1))
233 pc->p->cfg.high_temp = reg->hw + 1;
234 return;
235 }
236
237 if (reg->rhw != -1) {
238 /* try to allocate temporary with index rhw first */
239 if (!(pc->r_temp[reg->rhw])) {
240 pc->r_temp[reg->rhw] = reg;
241 reg->hw = reg->rhw;
242 if (pc->p->cfg.high_temp < (reg->rhw + 1))
243 pc->p->cfg.high_temp = reg->rhw + 1;
244 return;
245 }
246 /* make sure we don't get things like $r0 needs to go
247 * in $r1 and $r1 in $r0
248 */
249 i = pc->result_nr * 4;
250 }
251
252 for (; i < NV50_SU_MAX_TEMP; i++) {
253 if (!(pc->r_temp[i])) {
254 pc->r_temp[i] = reg;
255 reg->hw = i;
256 if (pc->p->cfg.high_temp < (i + 1))
257 pc->p->cfg.high_temp = i + 1;
258 return;
259 }
260 }
261
262 NOUVEAU_ERR("out of registers\n");
263 abort();
264 }
265
266 static INLINE struct nv50_reg *
267 reg_instance(struct nv50_pc *pc, struct nv50_reg *reg)
268 {
269 struct nv50_reg *ri;
270
271 assert(pc->reg_instance_nr < 16);
272 ri = &pc->reg_instances[pc->reg_instance_nr++];
273 if (reg) {
274 alloc_reg(pc, reg);
275 *ri = *reg;
276 reg->indirect[0] = reg->indirect[1] = -1;
277 reg->mod = 0;
278 }
279 return ri;
280 }
281
282 /* XXX: For shaders that aren't executed linearly (e.g. shaders that
283 * contain loops), we need to assign all hw regs to TGSI TEMPs early,
284 * lest we risk temp_temps overwriting regs alloc'd "later".
285 */
286 static struct nv50_reg *
287 alloc_temp(struct nv50_pc *pc, struct nv50_reg *dst)
288 {
289 struct nv50_reg *r;
290 int i;
291
292 if (dst && dst->type == P_TEMP && dst->hw == -1)
293 return dst;
294
295 for (i = 0; i < NV50_SU_MAX_TEMP; i++) {
296 if (!pc->r_temp[i]) {
297 r = MALLOC_STRUCT(nv50_reg);
298 ctor_reg(r, P_TEMP, -1, i);
299 pc->r_temp[i] = r;
300 return r;
301 }
302 }
303
304 NOUVEAU_ERR("out of registers\n");
305 abort();
306 return NULL;
307 }
308
309 /* release the hardware resource held by r */
310 static void
311 release_hw(struct nv50_pc *pc, struct nv50_reg *r)
312 {
313 assert(r->type == P_TEMP);
314 if (r->hw == -1)
315 return;
316
317 assert(pc->r_temp[r->hw] == r);
318 pc->r_temp[r->hw] = NULL;
319
320 r->acc = 0;
321 if (r->index == -1)
322 FREE(r);
323 }
324
325 static void
326 free_temp(struct nv50_pc *pc, struct nv50_reg *r)
327 {
328 if (r->index == -1) {
329 unsigned hw = r->hw;
330
331 FREE(pc->r_temp[hw]);
332 pc->r_temp[hw] = NULL;
333 }
334 }
335
336 static int
337 alloc_temp4(struct nv50_pc *pc, struct nv50_reg *dst[4], int idx)
338 {
339 int i;
340
341 if ((idx + 4) >= NV50_SU_MAX_TEMP)
342 return 1;
343
344 if (pc->r_temp[idx] || pc->r_temp[idx + 1] ||
345 pc->r_temp[idx + 2] || pc->r_temp[idx + 3])
346 return alloc_temp4(pc, dst, idx + 4);
347
348 for (i = 0; i < 4; i++) {
349 dst[i] = MALLOC_STRUCT(nv50_reg);
350 ctor_reg(dst[i], P_TEMP, -1, idx + i);
351 pc->r_temp[idx + i] = dst[i];
352 }
353
354 return 0;
355 }
356
357 static void
358 free_temp4(struct nv50_pc *pc, struct nv50_reg *reg[4])
359 {
360 int i;
361
362 for (i = 0; i < 4; i++)
363 free_temp(pc, reg[i]);
364 }
365
366 static struct nv50_reg *
367 temp_temp(struct nv50_pc *pc, struct nv50_program_exec *e)
368 {
369 if (pc->temp_temp_nr >= 16)
370 assert(0);
371
372 pc->temp_temp[pc->temp_temp_nr] = alloc_temp(pc, NULL);
373 pc->temp_temp_exec[pc->temp_temp_nr] = e;
374 return pc->temp_temp[pc->temp_temp_nr++];
375 }
376
377 /* This *must* be called for all nv50_program_exec that have been
378 * given as argument to temp_temp, or the temps will be leaked !
379 */
380 static void
381 kill_temp_temp(struct nv50_pc *pc, struct nv50_program_exec *e)
382 {
383 int i;
384
385 for (i = 0; i < pc->temp_temp_nr; i++)
386 if (pc->temp_temp_exec[i] == e)
387 free_temp(pc, pc->temp_temp[i]);
388 if (!e)
389 pc->temp_temp_nr = 0;
390 }
391
392 static int
393 ctor_immd_4u32(struct nv50_pc *pc,
394 uint32_t x, uint32_t y, uint32_t z, uint32_t w)
395 {
396 unsigned size = pc->immd_nr * 4 * sizeof(uint32_t);
397
398 pc->immd_buf = REALLOC(pc->immd_buf, size, size + 4 * sizeof(uint32_t));
399
400 pc->immd_buf[(pc->immd_nr * 4) + 0] = x;
401 pc->immd_buf[(pc->immd_nr * 4) + 1] = y;
402 pc->immd_buf[(pc->immd_nr * 4) + 2] = z;
403 pc->immd_buf[(pc->immd_nr * 4) + 3] = w;
404
405 return pc->immd_nr++;
406 }
407
408 static INLINE int
409 ctor_immd_4f32(struct nv50_pc *pc, float x, float y, float z, float w)
410 {
411 return ctor_immd_4u32(pc, fui(x), fui(y), fui(z), fui(w));
412 }
413
414 static struct nv50_reg *
415 alloc_immd(struct nv50_pc *pc, float f)
416 {
417 struct nv50_reg *r = MALLOC_STRUCT(nv50_reg);
418 unsigned hw;
419
420 for (hw = 0; hw < pc->immd_nr * 4; hw++)
421 if (pc->immd_buf[hw] == fui(f))
422 break;
423
424 if (hw == pc->immd_nr * 4)
425 hw = ctor_immd_4f32(pc, f, -f, 0.5 * f, 0) * 4;
426
427 ctor_reg(r, P_IMMD, -1, hw);
428 return r;
429 }
430
431 static struct nv50_program_exec *
432 exec(struct nv50_pc *pc)
433 {
434 struct nv50_program_exec *e = CALLOC_STRUCT(nv50_program_exec);
435
436 e->param.index = -1;
437 return e;
438 }
439
440 static void
441 emit(struct nv50_pc *pc, struct nv50_program_exec *e)
442 {
443 struct nv50_program *p = pc->p;
444
445 if (p->exec_tail)
446 p->exec_tail->next = e;
447 if (!p->exec_head)
448 p->exec_head = e;
449 p->exec_tail = e;
450 p->exec_size += (e->inst[0] & 1) ? 2 : 1;
451
452 kill_temp_temp(pc, e);
453 }
454
455 static INLINE void set_long(struct nv50_pc *, struct nv50_program_exec *);
456
457 static boolean
458 is_long(struct nv50_program_exec *e)
459 {
460 if (e->inst[0] & 1)
461 return TRUE;
462 return FALSE;
463 }
464
465 static boolean
466 is_immd(struct nv50_program_exec *e)
467 {
468 if (is_long(e) && (e->inst[1] & 3) == 3)
469 return TRUE;
470 return FALSE;
471 }
472
473 static boolean
474 is_join(struct nv50_program_exec *e)
475 {
476 if (is_long(e) && (e->inst[1] & 3) == 2)
477 return TRUE;
478 return FALSE;
479 }
480
481 static INLINE boolean
482 is_control_flow(struct nv50_program_exec *e)
483 {
484 return (e->inst[0] & 2);
485 }
486
487 static INLINE void
488 set_pred(struct nv50_pc *pc, unsigned pred, unsigned idx,
489 struct nv50_program_exec *e)
490 {
491 assert(!is_immd(e));
492 set_long(pc, e);
493 e->inst[1] &= ~((0x1f << 7) | (0x3 << 12));
494 e->inst[1] |= (pred << 7) | (idx << 12);
495 }
496
497 static INLINE void
498 set_pred_wr(struct nv50_pc *pc, unsigned on, unsigned idx,
499 struct nv50_program_exec *e)
500 {
501 set_long(pc, e);
502 e->inst[1] &= ~((0x3 << 4) | (1 << 6));
503 e->inst[1] |= (idx << 4) | (on << 6);
504 }
505
506 static INLINE void
507 set_long(struct nv50_pc *pc, struct nv50_program_exec *e)
508 {
509 if (is_long(e))
510 return;
511
512 e->inst[0] |= 1;
513 set_pred(pc, 0xf, 0, e);
514 set_pred_wr(pc, 0, 0, e);
515 }
516
517 static INLINE void
518 set_dst(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_program_exec *e)
519 {
520 if (dst->type == P_RESULT) {
521 set_long(pc, e);
522 e->inst[1] |= 0x00000008;
523 }
524
525 alloc_reg(pc, dst);
526 if (dst->hw > 63)
527 set_long(pc, e);
528 e->inst[0] |= (dst->hw << 2);
529 }
530
531 static INLINE void
532 set_immd(struct nv50_pc *pc, struct nv50_reg *imm, struct nv50_program_exec *e)
533 {
534 set_long(pc, e);
535 /* XXX: can't be predicated - bits overlap; cases where both
536 * are required should be avoided by using pc->allow32 */
537 set_pred(pc, 0, 0, e);
538 set_pred_wr(pc, 0, 0, e);
539
540 e->inst[1] |= 0x00000002 | 0x00000001;
541 e->inst[0] |= (pc->immd_buf[imm->hw] & 0x3f) << 16;
542 e->inst[1] |= (pc->immd_buf[imm->hw] >> 6) << 2;
543 }
544
545 static INLINE void
546 set_addr(struct nv50_program_exec *e, struct nv50_reg *a)
547 {
548 assert(a->type == P_ADDR);
549
550 assert(!(e->inst[0] & 0x0c000000));
551 assert(!(e->inst[1] & 0x00000004));
552
553 e->inst[0] |= (a->hw & 3) << 26;
554 e->inst[1] |= a->hw & 4;
555 }
556
557 static void
558 emit_arl(struct nv50_pc *, struct nv50_reg *, struct nv50_reg *, uint8_t);
559
560 static void
561 emit_shl_imm(struct nv50_pc *, struct nv50_reg *, struct nv50_reg *, int);
562
563 static void
564 emit_mov_from_addr(struct nv50_pc *pc, struct nv50_reg *dst,
565 struct nv50_reg *src)
566 {
567 struct nv50_program_exec *e = exec(pc);
568
569 e->inst[1] = 0x40000000;
570 set_long(pc, e);
571 set_dst(pc, dst, e);
572 set_addr(e, src);
573
574 emit(pc, e);
575 }
576
577 static void
578 emit_add_addr_imm(struct nv50_pc *pc, struct nv50_reg *dst,
579 struct nv50_reg *src0, uint16_t src1_val)
580 {
581 struct nv50_program_exec *e = exec(pc);
582
583 e->inst[0] = 0xd0000000 | (src1_val << 9);
584 e->inst[1] = 0x20000000;
585 set_long(pc, e);
586 e->inst[0] |= dst->hw << 2;
587 if (src0) /* otherwise will add to $a0, which is always 0 */
588 set_addr(e, src0);
589
590 emit(pc, e);
591 }
592
593 #define INTERP_LINEAR 0
594 #define INTERP_FLAT 1
595 #define INTERP_PERSPECTIVE 2
596 #define INTERP_CENTROID 4
597
598 /* interpolant index has been stored in dst->rhw */
599 static void
600 emit_interp(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *iv,
601 unsigned mode)
602 {
603 struct nv50_program_exec *e = exec(pc);
604 assert(dst->rhw != -1);
605
606 e->inst[0] |= 0x80000000;
607 set_dst(pc, dst, e);
608 e->inst[0] |= (dst->rhw << 16);
609
610 if (mode & INTERP_FLAT) {
611 e->inst[0] |= (1 << 8);
612 } else {
613 if (mode & INTERP_PERSPECTIVE) {
614 e->inst[0] |= (1 << 25);
615 alloc_reg(pc, iv);
616 e->inst[0] |= (iv->hw << 9);
617 }
618
619 if (mode & INTERP_CENTROID)
620 e->inst[0] |= (1 << 24);
621 }
622
623 emit(pc, e);
624 }
625
626 static void
627 set_data(struct nv50_pc *pc, struct nv50_reg *src, unsigned m, unsigned s,
628 struct nv50_program_exec *e)
629 {
630 set_long(pc, e);
631
632 e->param.index = src->hw & 127;
633 e->param.shift = s;
634 e->param.mask = m << (s % 32);
635
636 if (src->hw < 0 || src->hw > 127) /* need (additional) address reg */
637 set_addr(e, get_address_reg(pc, src));
638 else
639 if (src->acc < 0) {
640 assert(src->type == P_CONST);
641 set_addr(e, pc->addr[src->indirect[0]]);
642 }
643
644 e->inst[1] |= (src->buf_index << 22);
645 }
646
647 /* Never apply nv50_reg::mod in emit_mov, or carefully check the code !!! */
648 static void
649 emit_mov(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
650 {
651 struct nv50_program_exec *e = exec(pc);
652
653 e->inst[0] = 0x10000000;
654 if (!pc->allow32)
655 set_long(pc, e);
656
657 set_dst(pc, dst, e);
658
659 if (!is_long(e) && src->type == P_IMMD) {
660 set_immd(pc, src, e);
661 /*XXX: 32-bit, but steals part of "half" reg space - need to
662 * catch and handle this case if/when we do half-regs
663 */
664 } else
665 if (src->type == P_IMMD || src->type == P_CONST) {
666 set_long(pc, e);
667 set_data(pc, src, 0x7f, 9, e);
668 e->inst[1] |= 0x20000000; /* mov from c[] */
669 } else {
670 if (src->type == P_ATTR) {
671 set_long(pc, e);
672 e->inst[1] |= 0x00200000;
673
674 if (src->vtx >= 0) {
675 /* indirect (vertex base + c) load from p[] */
676 e->inst[0] |= 0x01800000;
677 set_addr(e, get_address_reg(pc, src));
678 }
679 }
680
681 alloc_reg(pc, src);
682 if (src->hw > 63)
683 set_long(pc, e);
684 e->inst[0] |= (src->hw << 9);
685 }
686
687 if (is_long(e) && !is_immd(e)) {
688 e->inst[1] |= 0x04000000; /* 32-bit */
689 e->inst[1] |= 0x0000c000; /* 32-bit c[] load / lane mask 0:1 */
690 if (!(e->inst[1] & 0x20000000))
691 e->inst[1] |= 0x00030000; /* lane mask 2:3 */
692 } else
693 e->inst[0] |= 0x00008000;
694
695 emit(pc, e);
696 }
697
698 static INLINE void
699 emit_mov_immdval(struct nv50_pc *pc, struct nv50_reg *dst, float f)
700 {
701 struct nv50_reg *imm = alloc_immd(pc, f);
702 emit_mov(pc, dst, imm);
703 FREE(imm);
704 }
705
706 /* Assign the hw of the discarded temporary register src
707 * to the tgsi register dst and free src.
708 */
709 static void
710 assimilate_temp(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
711 {
712 assert(src->index == -1 && src->hw != -1);
713
714 if (pc->if_lvl || pc->loop_lvl ||
715 (dst->type != P_TEMP) ||
716 (src->hw < pc->result_nr * 4 &&
717 pc->p->type == PIPE_SHADER_FRAGMENT) ||
718 pc->p->info.opcode_count[TGSI_OPCODE_CAL] ||
719 pc->p->info.opcode_count[TGSI_OPCODE_BRA]) {
720
721 emit_mov(pc, dst, src);
722 free_temp(pc, src);
723 return;
724 }
725
726 if (dst->hw != -1)
727 pc->r_temp[dst->hw] = NULL;
728 pc->r_temp[src->hw] = dst;
729 dst->hw = src->hw;
730
731 FREE(src);
732 }
733
734 static void
735 emit_nop(struct nv50_pc *pc)
736 {
737 struct nv50_program_exec *e = exec(pc);
738
739 e->inst[0] = 0xf0000000;
740 set_long(pc, e);
741 e->inst[1] = 0xe0000000;
742 emit(pc, e);
743 }
744
745 static boolean
746 check_swap_src_0_1(struct nv50_pc *pc,
747 struct nv50_reg **s0, struct nv50_reg **s1)
748 {
749 struct nv50_reg *src0 = *s0, *src1 = *s1;
750
751 if (src0->type == P_CONST) {
752 if (src1->type != P_CONST) {
753 *s0 = src1;
754 *s1 = src0;
755 return TRUE;
756 }
757 } else
758 if (src1->type == P_ATTR) {
759 if (src0->type != P_ATTR) {
760 *s0 = src1;
761 *s1 = src0;
762 return TRUE;
763 }
764 }
765
766 return FALSE;
767 }
768
769 static void
770 set_src_0_restricted(struct nv50_pc *pc, struct nv50_reg *src,
771 struct nv50_program_exec *e)
772 {
773 struct nv50_reg *temp;
774
775 if (src->type != P_TEMP) {
776 temp = temp_temp(pc, e);
777 emit_mov(pc, temp, src);
778 src = temp;
779 }
780
781 alloc_reg(pc, src);
782 if (src->hw > 63)
783 set_long(pc, e);
784 e->inst[0] |= (src->hw << 9);
785 }
786
787 static void
788 set_src_0(struct nv50_pc *pc, struct nv50_reg *src, struct nv50_program_exec *e)
789 {
790 if (src->type == P_ATTR) {
791 set_long(pc, e);
792 e->inst[1] |= 0x00200000;
793
794 if (src->vtx >= 0) {
795 e->inst[0] |= 0x01800000; /* src from p[] */
796 set_addr(e, get_address_reg(pc, src));
797 }
798 } else
799 if (src->type == P_CONST || src->type == P_IMMD) {
800 struct nv50_reg *temp = temp_temp(pc, e);
801
802 emit_mov(pc, temp, src);
803 src = temp;
804 }
805
806 alloc_reg(pc, src);
807 if (src->hw > 63)
808 set_long(pc, e);
809 e->inst[0] |= (src->hw << 9);
810 }
811
812 static void
813 set_src_1(struct nv50_pc *pc, struct nv50_reg *src, struct nv50_program_exec *e)
814 {
815 if (src->type == P_ATTR) {
816 struct nv50_reg *temp = temp_temp(pc, e);
817
818 emit_mov(pc, temp, src);
819 src = temp;
820 } else
821 if (src->type == P_CONST || src->type == P_IMMD) {
822 if (e->inst[0] & 0x01800000) {
823 struct nv50_reg *temp = temp_temp(pc, e);
824
825 emit_mov(pc, temp, src);
826 src = temp;
827 } else {
828 assert(!(e->inst[0] & 0x00800000));
829 set_data(pc, src, 0x7f, 16, e);
830 e->inst[0] |= 0x00800000;
831 }
832 }
833
834 alloc_reg(pc, src);
835 if (src->hw > 63)
836 set_long(pc, e);
837 e->inst[0] |= ((src->hw & 127) << 16);
838 }
839
840 static void
841 set_src_2(struct nv50_pc *pc, struct nv50_reg *src, struct nv50_program_exec *e)
842 {
843 set_long(pc, e);
844
845 if (src->type == P_ATTR) {
846 struct nv50_reg *temp = temp_temp(pc, e);
847
848 emit_mov(pc, temp, src);
849 src = temp;
850 } else
851 if (src->type == P_CONST || src->type == P_IMMD) {
852 if (e->inst[0] & 0x01800000) {
853 struct nv50_reg *temp = temp_temp(pc, e);
854
855 emit_mov(pc, temp, src);
856 src = temp;
857 } else {
858 assert(!(e->inst[0] & 0x01000000));
859 set_data(pc, src, 0x7f, 32+14, e);
860 e->inst[0] |= 0x01000000;
861 }
862 }
863
864 alloc_reg(pc, src);
865 e->inst[1] |= ((src->hw & 127) << 14);
866 }
867
868 static void
869 set_half_src(struct nv50_pc *pc, struct nv50_reg *src, int lh,
870 struct nv50_program_exec *e, int pos)
871 {
872 struct nv50_reg *r = src;
873
874 alloc_reg(pc, r);
875 if (r->type != P_TEMP) {
876 r = temp_temp(pc, e);
877 emit_mov(pc, r, src);
878 }
879
880 if (r->hw > (NV50_SU_MAX_TEMP / 2)) {
881 NOUVEAU_ERR("out of low GPRs\n");
882 abort();
883 }
884
885 e->inst[pos / 32] |= ((src->hw * 2) + lh) << (pos % 32);
886 }
887
888 static void
889 emit_mov_from_pred(struct nv50_pc *pc, struct nv50_reg *dst, int pred)
890 {
891 struct nv50_program_exec *e = exec(pc);
892
893 assert(dst->type == P_TEMP);
894 e->inst[1] = 0x20000000 | (pred << 12);
895 set_long(pc, e);
896 set_dst(pc, dst, e);
897
898 emit(pc, e);
899 }
900
901 static void
902 emit_mov_to_pred(struct nv50_pc *pc, int pred, struct nv50_reg *src)
903 {
904 struct nv50_program_exec *e = exec(pc);
905
906 e->inst[0] = 0x000001fc;
907 e->inst[1] = 0xa0000008;
908 set_long(pc, e);
909 set_pred_wr(pc, 1, pred, e);
910 set_src_0_restricted(pc, src, e);
911
912 emit(pc, e);
913 }
914
915 static void
916 emit_mul(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
917 struct nv50_reg *src1)
918 {
919 struct nv50_program_exec *e = exec(pc);
920
921 e->inst[0] |= 0xc0000000;
922
923 if (!pc->allow32)
924 set_long(pc, e);
925
926 check_swap_src_0_1(pc, &src0, &src1);
927 set_dst(pc, dst, e);
928 set_src_0(pc, src0, e);
929 if (src1->type == P_IMMD && !is_long(e)) {
930 if (src0->mod ^ src1->mod)
931 e->inst[0] |= 0x00008000;
932 set_immd(pc, src1, e);
933 } else {
934 set_src_1(pc, src1, e);
935 if ((src0->mod ^ src1->mod) & NV50_MOD_NEG) {
936 if (is_long(e))
937 e->inst[1] |= 0x08000000;
938 else
939 e->inst[0] |= 0x00008000;
940 }
941 }
942
943 emit(pc, e);
944 }
945
946 static void
947 emit_add(struct nv50_pc *pc, struct nv50_reg *dst,
948 struct nv50_reg *src0, struct nv50_reg *src1)
949 {
950 struct nv50_program_exec *e = exec(pc);
951
952 e->inst[0] = 0xb0000000;
953
954 alloc_reg(pc, src1);
955 check_swap_src_0_1(pc, &src0, &src1);
956
957 if (!pc->allow32 || (src0->mod | src1->mod) || src1->hw > 63) {
958 set_long(pc, e);
959 e->inst[1] |= ((src0->mod & NV50_MOD_NEG) << 26) |
960 ((src1->mod & NV50_MOD_NEG) << 27);
961 }
962
963 set_dst(pc, dst, e);
964 set_src_0(pc, src0, e);
965 if (src1->type == P_CONST || src1->type == P_ATTR || is_long(e))
966 set_src_2(pc, src1, e);
967 else
968 if (src1->type == P_IMMD)
969 set_immd(pc, src1, e);
970 else
971 set_src_1(pc, src1, e);
972
973 emit(pc, e);
974 }
975
976 static void
977 emit_arl(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src,
978 uint8_t s)
979 {
980 struct nv50_program_exec *e = exec(pc);
981
982 set_long(pc, e);
983 e->inst[1] |= 0xc0000000;
984
985 e->inst[0] |= dst->hw << 2;
986 e->inst[0] |= s << 16; /* shift left */
987 set_src_0(pc, src, e);
988
989 emit(pc, e);
990 }
991
992 static boolean
993 address_reg_suitable(struct nv50_reg *a, struct nv50_reg *r)
994 {
995 if (!r)
996 return FALSE;
997
998 if (r->vtx != a->vtx)
999 return FALSE;
1000 if (r->vtx >= 0)
1001 return (r->indirect[1] == a->indirect[1]);
1002
1003 if (r->hw < a->rhw || (r->hw - a->rhw) >= 128)
1004 return FALSE;
1005
1006 if (a->index >= 0)
1007 return (a->index == r->indirect[0]);
1008 return (a->indirect[0] == r->indirect[0]);
1009 }
1010
1011 static void
1012 load_vertex_base(struct nv50_pc *pc, struct nv50_reg *dst,
1013 struct nv50_reg *a, int shift)
1014 {
1015 struct nv50_reg mem, *temp;
1016
1017 ctor_reg(&mem, P_ATTR, -1, dst->vtx);
1018
1019 assert(dst->type == P_ADDR);
1020 if (!a) {
1021 emit_arl(pc, dst, &mem, 0);
1022 return;
1023 }
1024 temp = alloc_temp(pc, NULL);
1025
1026 if (shift) {
1027 emit_mov_from_addr(pc, temp, a);
1028 if (shift < 0)
1029 emit_shl_imm(pc, temp, temp, shift);
1030 emit_arl(pc, dst, temp, MAX2(shift, 0));
1031 }
1032 emit_mov(pc, temp, &mem);
1033 set_addr(pc->p->exec_tail, dst);
1034
1035 emit_arl(pc, dst, temp, 0);
1036 free_temp(pc, temp);
1037 }
1038
1039 /* case (ref == NULL): allocate address register for TGSI_FILE_ADDRESS
1040 * case (vtx >= 0, acc >= 0): load vertex base from a[vtx * 4] to $aX
1041 * case (vtx >= 0, acc < 0): load vertex base from s[$aY + vtx * 4] to $aX
1042 * case (vtx < 0, acc >= 0): memory address too high to encode
1043 * case (vtx < 0, acc < 0): get source register for TGSI_FILE_ADDRESS
1044 */
1045 static struct nv50_reg *
1046 get_address_reg(struct nv50_pc *pc, struct nv50_reg *ref)
1047 {
1048 int i;
1049 struct nv50_reg *a_ref, *a = NULL;
1050
1051 for (i = 0; i < NV50_SU_MAX_ADDR; ++i) {
1052 if (pc->r_addr[i].acc == 0)
1053 a = &pc->r_addr[i]; /* an unused address reg */
1054 else
1055 if (address_reg_suitable(&pc->r_addr[i], ref)) {
1056 pc->r_addr[i].acc = pc->insn_cur;
1057 return &pc->r_addr[i];
1058 } else
1059 if (!a && pc->r_addr[i].index < 0 &&
1060 pc->r_addr[i].acc < pc->insn_cur)
1061 a = &pc->r_addr[i];
1062 }
1063 if (!a) {
1064 /* We'll be able to spill address regs when this
1065 * mess is replaced with a proper compiler ...
1066 */
1067 NOUVEAU_ERR("out of address regs\n");
1068 abort();
1069 return NULL;
1070 }
1071
1072 /* initialize and reserve for this TGSI instruction */
1073 a->rhw = 0;
1074 a->index = a->indirect[0] = a->indirect[1] = -1;
1075 a->acc = pc->insn_cur;
1076
1077 if (!ref) {
1078 a->vtx = -1;
1079 return a;
1080 }
1081 a->vtx = ref->vtx;
1082
1083 /* now put in the correct value ... */
1084
1085 if (ref->vtx >= 0) {
1086 a->indirect[1] = ref->indirect[1];
1087
1088 /* For an indirect vertex index, we need to shift address right
1089 * by 2, the address register will contain vtx * 16, we need to
1090 * load from a[vtx * 4].
1091 */
1092 load_vertex_base(pc, a, (ref->acc < 0) ?
1093 pc->addr[ref->indirect[1]] : NULL, -2);
1094 } else {
1095 assert(ref->acc < 0 || ref->indirect[0] < 0);
1096
1097 a->rhw = ref->hw & ~0x7f;
1098 a->indirect[0] = ref->indirect[0];
1099 a_ref = (ref->acc < 0) ? pc->addr[ref->indirect[0]] : NULL;
1100
1101 emit_add_addr_imm(pc, a, a_ref, a->rhw * 4);
1102 }
1103 return a;
1104 }
1105
1106 #define NV50_MAX_F32 0x880
1107 #define NV50_MAX_S32 0x08c
1108 #define NV50_MAX_U32 0x084
1109 #define NV50_MIN_F32 0x8a0
1110 #define NV50_MIN_S32 0x0ac
1111 #define NV50_MIN_U32 0x0a4
1112
1113 static void
1114 emit_minmax(struct nv50_pc *pc, unsigned sub, struct nv50_reg *dst,
1115 struct nv50_reg *src0, struct nv50_reg *src1)
1116 {
1117 struct nv50_program_exec *e = exec(pc);
1118
1119 set_long(pc, e);
1120 e->inst[0] |= 0x30000000 | ((sub & 0x800) << 20);
1121 e->inst[1] |= (sub << 24);
1122
1123 check_swap_src_0_1(pc, &src0, &src1);
1124 set_dst(pc, dst, e);
1125 set_src_0(pc, src0, e);
1126 set_src_1(pc, src1, e);
1127
1128 if (src0->mod & NV50_MOD_ABS)
1129 e->inst[1] |= 0x00100000;
1130 if (src1->mod & NV50_MOD_ABS)
1131 e->inst[1] |= 0x00080000;
1132
1133 emit(pc, e);
1134 }
1135
1136 static INLINE void
1137 emit_sub(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
1138 struct nv50_reg *src1)
1139 {
1140 src1->mod ^= NV50_MOD_NEG;
1141 emit_add(pc, dst, src0, src1);
1142 src1->mod ^= NV50_MOD_NEG;
1143 }
1144
1145 static void
1146 emit_bitop2(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
1147 struct nv50_reg *src1, unsigned op)
1148 {
1149 struct nv50_program_exec *e = exec(pc);
1150
1151 e->inst[0] = 0xd0000000;
1152 set_long(pc, e);
1153
1154 check_swap_src_0_1(pc, &src0, &src1);
1155 set_dst(pc, dst, e);
1156 set_src_0(pc, src0, e);
1157
1158 if (op != TGSI_OPCODE_AND && op != TGSI_OPCODE_OR &&
1159 op != TGSI_OPCODE_XOR)
1160 assert(!"invalid bit op");
1161
1162 assert(!(src0->mod | src1->mod));
1163
1164 if (src1->type == P_IMMD && src0->type == P_TEMP && pc->allow32) {
1165 set_immd(pc, src1, e);
1166 if (op == TGSI_OPCODE_OR)
1167 e->inst[0] |= 0x0100;
1168 else
1169 if (op == TGSI_OPCODE_XOR)
1170 e->inst[0] |= 0x8000;
1171 } else {
1172 set_src_1(pc, src1, e);
1173 e->inst[1] |= 0x04000000; /* 32 bit */
1174 if (op == TGSI_OPCODE_OR)
1175 e->inst[1] |= 0x4000;
1176 else
1177 if (op == TGSI_OPCODE_XOR)
1178 e->inst[1] |= 0x8000;
1179 }
1180
1181 emit(pc, e);
1182 }
1183
1184 static void
1185 emit_not(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
1186 {
1187 struct nv50_program_exec *e = exec(pc);
1188
1189 e->inst[0] = 0xd0000000;
1190 e->inst[1] = 0x0402c000;
1191 set_long(pc, e);
1192 set_dst(pc, dst, e);
1193 set_src_1(pc, src, e);
1194
1195 emit(pc, e);
1196 }
1197
1198 static void
1199 emit_shift(struct nv50_pc *pc, struct nv50_reg *dst,
1200 struct nv50_reg *src0, struct nv50_reg *src1, unsigned dir)
1201 {
1202 struct nv50_program_exec *e = exec(pc);
1203
1204 e->inst[0] = 0x30000000;
1205 e->inst[1] = 0xc4000000;
1206
1207 set_long(pc, e);
1208 set_dst(pc, dst, e);
1209 set_src_0(pc, src0, e);
1210
1211 if (src1->type == P_IMMD) {
1212 e->inst[1] |= (1 << 20);
1213 e->inst[0] |= (pc->immd_buf[src1->hw] & 0x7f) << 16;
1214 } else
1215 set_src_1(pc, src1, e);
1216
1217 if (dir != TGSI_OPCODE_SHL)
1218 e->inst[1] |= (1 << 29);
1219
1220 if (dir == TGSI_OPCODE_ISHR)
1221 e->inst[1] |= (1 << 27);
1222
1223 emit(pc, e);
1224 }
1225
1226 static void
1227 emit_shl_imm(struct nv50_pc *pc, struct nv50_reg *dst,
1228 struct nv50_reg *src, int s)
1229 {
1230 struct nv50_program_exec *e = exec(pc);
1231
1232 e->inst[0] = 0x30000000;
1233 e->inst[1] = 0xc4100000;
1234 if (s < 0) {
1235 e->inst[1] |= 1 << 29;
1236 s = -s;
1237 }
1238 e->inst[1] |= ((s & 0x7f) << 16);
1239
1240 set_long(pc, e);
1241 set_dst(pc, dst, e);
1242 set_src_0(pc, src, e);
1243
1244 emit(pc, e);
1245 }
1246
1247 static void
1248 emit_mad(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
1249 struct nv50_reg *src1, struct nv50_reg *src2)
1250 {
1251 struct nv50_program_exec *e = exec(pc);
1252
1253 e->inst[0] |= 0xe0000000;
1254
1255 check_swap_src_0_1(pc, &src0, &src1);
1256 set_dst(pc, dst, e);
1257 set_src_0(pc, src0, e);
1258 set_src_1(pc, src1, e);
1259 set_src_2(pc, src2, e);
1260
1261 if ((src0->mod ^ src1->mod) & NV50_MOD_NEG)
1262 e->inst[1] |= 0x04000000;
1263 if (src2->mod & NV50_MOD_NEG)
1264 e->inst[1] |= 0x08000000;
1265
1266 emit(pc, e);
1267 }
1268
1269 static INLINE void
1270 emit_msb(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src0,
1271 struct nv50_reg *src1, struct nv50_reg *src2)
1272 {
1273 src2->mod ^= NV50_MOD_NEG;
1274 emit_mad(pc, dst, src0, src1, src2);
1275 src2->mod ^= NV50_MOD_NEG;
1276 }
1277
1278 #define NV50_FLOP_RCP 0
1279 #define NV50_FLOP_RSQ 2
1280 #define NV50_FLOP_LG2 3
1281 #define NV50_FLOP_SIN 4
1282 #define NV50_FLOP_COS 5
1283 #define NV50_FLOP_EX2 6
1284
1285 /* rcp, rsqrt, lg2 support neg and abs */
1286 static void
1287 emit_flop(struct nv50_pc *pc, unsigned sub,
1288 struct nv50_reg *dst, struct nv50_reg *src)
1289 {
1290 struct nv50_program_exec *e = exec(pc);
1291
1292 e->inst[0] |= 0x90000000;
1293 if (sub || src->mod) {
1294 set_long(pc, e);
1295 e->inst[1] |= (sub << 29);
1296 }
1297
1298 set_dst(pc, dst, e);
1299 set_src_0_restricted(pc, src, e);
1300
1301 assert(!src->mod || sub < 4);
1302
1303 if (src->mod & NV50_MOD_NEG)
1304 e->inst[1] |= 0x04000000;
1305 if (src->mod & NV50_MOD_ABS)
1306 e->inst[1] |= 0x00100000;
1307
1308 emit(pc, e);
1309 }
1310
1311 static void
1312 emit_preex2(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
1313 {
1314 struct nv50_program_exec *e = exec(pc);
1315
1316 e->inst[0] |= 0xb0000000;
1317
1318 set_dst(pc, dst, e);
1319 set_src_0(pc, src, e);
1320 set_long(pc, e);
1321 e->inst[1] |= (6 << 29) | 0x00004000;
1322
1323 if (src->mod & NV50_MOD_NEG)
1324 e->inst[1] |= 0x04000000;
1325 if (src->mod & NV50_MOD_ABS)
1326 e->inst[1] |= 0x00100000;
1327
1328 emit(pc, e);
1329 }
1330
1331 static void
1332 emit_precossin(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
1333 {
1334 struct nv50_program_exec *e = exec(pc);
1335
1336 e->inst[0] |= 0xb0000000;
1337
1338 set_dst(pc, dst, e);
1339 set_src_0(pc, src, e);
1340 set_long(pc, e);
1341 e->inst[1] |= (6 << 29);
1342
1343 if (src->mod & NV50_MOD_NEG)
1344 e->inst[1] |= 0x04000000;
1345 if (src->mod & NV50_MOD_ABS)
1346 e->inst[1] |= 0x00100000;
1347
1348 emit(pc, e);
1349 }
1350
1351 #define CVT_RN (0x00 << 16)
1352 #define CVT_FLOOR (0x02 << 16)
1353 #define CVT_CEIL (0x04 << 16)
1354 #define CVT_TRUNC (0x06 << 16)
1355 #define CVT_SAT (0x08 << 16)
1356 #define CVT_ABS (0x10 << 16)
1357
1358 #define CVT_X32_X32 0x04004000
1359 #define CVT_X32_S32 0x04014000
1360 #define CVT_F32_F32 ((0xc0 << 24) | CVT_X32_X32)
1361 #define CVT_S32_F32 ((0x88 << 24) | CVT_X32_X32)
1362 #define CVT_U32_F32 ((0x80 << 24) | CVT_X32_X32)
1363 #define CVT_F32_S32 ((0x40 << 24) | CVT_X32_S32)
1364 #define CVT_F32_U32 ((0x40 << 24) | CVT_X32_X32)
1365 #define CVT_S32_S32 ((0x08 << 24) | CVT_X32_S32)
1366 #define CVT_S32_U32 ((0x08 << 24) | CVT_X32_X32)
1367 #define CVT_U32_S32 ((0x00 << 24) | CVT_X32_S32)
1368
1369 #define CVT_NEG 0x20000000
1370 #define CVT_RI 0x08000000
1371
1372 static void
1373 emit_cvt(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src,
1374 int wp, uint32_t cvn)
1375 {
1376 struct nv50_program_exec *e;
1377
1378 e = exec(pc);
1379
1380 if (src->mod & NV50_MOD_NEG) cvn |= CVT_NEG;
1381 if (src->mod & NV50_MOD_ABS) cvn |= CVT_ABS;
1382
1383 e->inst[0] = 0xa0000000;
1384 e->inst[1] = cvn;
1385 set_long(pc, e);
1386 set_src_0(pc, src, e);
1387
1388 if (wp >= 0)
1389 set_pred_wr(pc, 1, wp, e);
1390
1391 if (dst)
1392 set_dst(pc, dst, e);
1393 else {
1394 e->inst[0] |= 0x000001fc;
1395 e->inst[1] |= 0x00000008;
1396 }
1397
1398 emit(pc, e);
1399 }
1400
1401 /* nv50 Condition codes:
1402 * 0x1 = LT
1403 * 0x2 = EQ
1404 * 0x3 = LE
1405 * 0x4 = GT
1406 * 0x5 = NE
1407 * 0x6 = GE
1408 * 0x7 = set condition code ? (used before bra.lt/le/gt/ge)
1409 * 0x8 = unordered bit (allows NaN)
1410 *
1411 * mode = 0x04 (u32), 0x0c (s32), 0x80 (f32)
1412 */
1413 static void
1414 emit_set(struct nv50_pc *pc, unsigned ccode, struct nv50_reg *dst, int wp,
1415 struct nv50_reg *src0, struct nv50_reg *src1, uint8_t mode)
1416 {
1417 static const unsigned cc_swapped[8] = { 0, 4, 2, 6, 1, 5, 3, 7 };
1418
1419 struct nv50_program_exec *e = exec(pc);
1420 struct nv50_reg *rdst;
1421
1422 assert(ccode < 16);
1423 if (check_swap_src_0_1(pc, &src0, &src1))
1424 ccode = cc_swapped[ccode & 7] | (ccode & 8);
1425
1426 rdst = dst;
1427 if (dst && dst->type != P_TEMP)
1428 dst = alloc_temp(pc, NULL);
1429
1430 set_long(pc, e);
1431 e->inst[0] |= 0x30000000 | (mode << 24);
1432 e->inst[1] |= 0x60000000 | (ccode << 14);
1433
1434 if (wp >= 0)
1435 set_pred_wr(pc, 1, wp, e);
1436 if (dst)
1437 set_dst(pc, dst, e);
1438 else {
1439 e->inst[0] |= 0x000001fc;
1440 e->inst[1] |= 0x00000008;
1441 }
1442
1443 set_src_0(pc, src0, e);
1444 set_src_1(pc, src1, e);
1445
1446 emit(pc, e);
1447
1448 if (rdst && mode == 0x80) /* convert to float ? */
1449 emit_cvt(pc, rdst, dst, -1, CVT_ABS | CVT_F32_S32);
1450 if (rdst && rdst != dst)
1451 free_temp(pc, dst);
1452 }
1453
1454 static INLINE void
1455 map_tgsi_setop_hw(unsigned op, uint8_t *cc, uint8_t *ty)
1456 {
1457 switch (op) {
1458 case TGSI_OPCODE_SLT: *cc = 0x1; *ty = 0x80; break;
1459 case TGSI_OPCODE_SGE: *cc = 0x6; *ty = 0x80; break;
1460 case TGSI_OPCODE_SEQ: *cc = 0x2; *ty = 0x80; break;
1461 case TGSI_OPCODE_SGT: *cc = 0x4; *ty = 0x80; break;
1462 case TGSI_OPCODE_SLE: *cc = 0x3; *ty = 0x80; break;
1463 case TGSI_OPCODE_SNE: *cc = 0xd; *ty = 0x80; break;
1464
1465 case TGSI_OPCODE_ISLT: *cc = 0x1; *ty = 0x0c; break;
1466 case TGSI_OPCODE_ISGE: *cc = 0x6; *ty = 0x0c; break;
1467 case TGSI_OPCODE_USEQ: *cc = 0x2; *ty = 0x04; break;
1468 case TGSI_OPCODE_USGE: *cc = 0x6; *ty = 0x04; break;
1469 case TGSI_OPCODE_USLT: *cc = 0x1; *ty = 0x04; break;
1470 case TGSI_OPCODE_USNE: *cc = 0x5; *ty = 0x04; break;
1471 default:
1472 assert(0);
1473 return;
1474 }
1475 }
1476
1477 static void
1478 emit_add_b32(struct nv50_pc *pc, struct nv50_reg *dst,
1479 struct nv50_reg *src0, struct nv50_reg *rsrc1)
1480 {
1481 struct nv50_program_exec *e = exec(pc);
1482 struct nv50_reg *src1;
1483
1484 e->inst[0] = 0x20000000;
1485
1486 alloc_reg(pc, rsrc1);
1487 check_swap_src_0_1(pc, &src0, &rsrc1);
1488
1489 src1 = rsrc1;
1490 if (src0->mod & rsrc1->mod & NV50_MOD_NEG) {
1491 src1 = temp_temp(pc, e);
1492 emit_cvt(pc, src1, rsrc1, -1, CVT_S32_S32);
1493 }
1494
1495 if (!pc->allow32 || src1->hw > 63 ||
1496 (src1->type != P_TEMP && src1->type != P_IMMD))
1497 set_long(pc, e);
1498
1499 set_dst(pc, dst, e);
1500 set_src_0(pc, src0, e);
1501
1502 if (is_long(e)) {
1503 e->inst[1] |= 1 << 26;
1504 set_src_2(pc, src1, e);
1505 } else {
1506 e->inst[0] |= 0x8000;
1507 if (src1->type == P_IMMD)
1508 set_immd(pc, src1, e);
1509 else
1510 set_src_1(pc, src1, e);
1511 }
1512
1513 if (src0->mod & NV50_MOD_NEG)
1514 e->inst[0] |= 1 << 28;
1515 else
1516 if (src1->mod & NV50_MOD_NEG)
1517 e->inst[0] |= 1 << 22;
1518
1519 emit(pc, e);
1520 }
1521
1522 static void
1523 emit_mad_u16(struct nv50_pc *pc, struct nv50_reg *dst,
1524 struct nv50_reg *src0, int lh_0, struct nv50_reg *src1, int lh_1,
1525 struct nv50_reg *src2)
1526 {
1527 struct nv50_program_exec *e = exec(pc);
1528
1529 e->inst[0] = 0x60000000;
1530 if (!pc->allow32)
1531 set_long(pc, e);
1532 set_dst(pc, dst, e);
1533
1534 set_half_src(pc, src0, lh_0, e, 9);
1535 set_half_src(pc, src1, lh_1, e, 16);
1536 alloc_reg(pc, src2);
1537 if (is_long(e) || (src2->type != P_TEMP) || (src2->hw != dst->hw))
1538 set_src_2(pc, src2, e);
1539
1540 emit(pc, e);
1541 }
1542
1543 static void
1544 emit_mul_u16(struct nv50_pc *pc, struct nv50_reg *dst,
1545 struct nv50_reg *src0, int lh_0, struct nv50_reg *src1, int lh_1)
1546 {
1547 struct nv50_program_exec *e = exec(pc);
1548
1549 e->inst[0] = 0x40000000;
1550 set_long(pc, e);
1551 set_dst(pc, dst, e);
1552
1553 set_half_src(pc, src0, lh_0, e, 9);
1554 set_half_src(pc, src1, lh_1, e, 16);
1555
1556 emit(pc, e);
1557 }
1558
1559 static void
1560 emit_sad(struct nv50_pc *pc, struct nv50_reg *dst,
1561 struct nv50_reg *src0, struct nv50_reg *src1, struct nv50_reg *src2)
1562 {
1563 struct nv50_program_exec *e = exec(pc);
1564
1565 e->inst[0] = 0x50000000;
1566 if (!pc->allow32)
1567 set_long(pc, e);
1568 check_swap_src_0_1(pc, &src0, &src1);
1569 set_dst(pc, dst, e);
1570 set_src_0(pc, src0, e);
1571 set_src_1(pc, src1, e);
1572 alloc_reg(pc, src2);
1573 if (is_long(e) || (src2->type != dst->type) || (src2->hw != dst->hw))
1574 set_src_2(pc, src2, e);
1575
1576 if (is_long(e))
1577 e->inst[1] |= 0x0c << 24;
1578 else
1579 e->inst[0] |= 0x81 << 8;
1580
1581 emit(pc, e);
1582 }
1583
1584 static INLINE void
1585 emit_flr(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
1586 {
1587 emit_cvt(pc, dst, src, -1, CVT_FLOOR | CVT_F32_F32 | CVT_RI);
1588 }
1589
1590 static void
1591 emit_pow(struct nv50_pc *pc, struct nv50_reg *dst,
1592 struct nv50_reg *v, struct nv50_reg *e)
1593 {
1594 struct nv50_reg *temp = alloc_temp(pc, NULL);
1595
1596 emit_flop(pc, NV50_FLOP_LG2, temp, v);
1597 emit_mul(pc, temp, temp, e);
1598 emit_preex2(pc, temp, temp);
1599 emit_flop(pc, NV50_FLOP_EX2, dst, temp);
1600
1601 free_temp(pc, temp);
1602 }
1603
1604 static INLINE void
1605 emit_sat(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
1606 {
1607 emit_cvt(pc, dst, src, -1, CVT_SAT | CVT_F32_F32);
1608 }
1609
1610 static void
1611 emit_lit(struct nv50_pc *pc, struct nv50_reg **dst, unsigned mask,
1612 struct nv50_reg **src)
1613 {
1614 struct nv50_reg *one = alloc_immd(pc, 1.0);
1615 struct nv50_reg *zero = alloc_immd(pc, 0.0);
1616 struct nv50_reg *neg128 = alloc_immd(pc, -127.999999);
1617 struct nv50_reg *pos128 = alloc_immd(pc, 127.999999);
1618 struct nv50_reg *tmp[4] = { 0 };
1619 boolean allow32 = pc->allow32;
1620
1621 pc->allow32 = FALSE;
1622
1623 if (mask & (3 << 1)) {
1624 tmp[0] = alloc_temp(pc, NULL);
1625 emit_minmax(pc, NV50_MAX_F32, tmp[0], src[0], zero);
1626 }
1627
1628 if (mask & (1 << 2)) {
1629 set_pred_wr(pc, 1, 0, pc->p->exec_tail);
1630
1631 tmp[1] = temp_temp(pc, NULL);
1632 emit_minmax(pc, NV50_MAX_F32, tmp[1], src[1], zero);
1633
1634 tmp[3] = temp_temp(pc, NULL);
1635 emit_minmax(pc, NV50_MAX_F32, tmp[3], src[3], neg128);
1636 emit_minmax(pc, NV50_MIN_F32, tmp[3], tmp[3], pos128);
1637
1638 emit_pow(pc, dst[2], tmp[1], tmp[3]);
1639 emit_mov(pc, dst[2], zero);
1640 set_pred(pc, 3, 0, pc->p->exec_tail);
1641 }
1642
1643 if (mask & (1 << 1))
1644 assimilate_temp(pc, dst[1], tmp[0]);
1645 else
1646 if (mask & (1 << 2))
1647 free_temp(pc, tmp[0]);
1648
1649 pc->allow32 = allow32;
1650
1651 /* do this last, in case src[i,j] == dst[0,3] */
1652 if (mask & (1 << 0))
1653 emit_mov(pc, dst[0], one);
1654
1655 if (mask & (1 << 3))
1656 emit_mov(pc, dst[3], one);
1657
1658 FREE(pos128);
1659 FREE(neg128);
1660 FREE(zero);
1661 FREE(one);
1662 }
1663
1664 static void
1665 emit_kil(struct nv50_pc *pc, struct nv50_reg *src)
1666 {
1667 struct nv50_program_exec *e;
1668 const int r_pred = 1;
1669
1670 e = exec(pc);
1671 e->inst[0] = 0x00000002; /* discard */
1672 set_long(pc, e); /* sets cond code to ALWAYS */
1673
1674 if (src) {
1675 set_pred(pc, 0x1 /* cc = LT */, r_pred, e);
1676 /* write to predicate reg */
1677 emit_cvt(pc, NULL, src, r_pred, CVT_F32_F32);
1678 }
1679
1680 emit(pc, e);
1681 }
1682
1683 static struct nv50_program_exec *
1684 emit_control_flow(struct nv50_pc *pc, unsigned op, int pred, unsigned cc)
1685 {
1686 struct nv50_program_exec *e = exec(pc);
1687
1688 e->inst[0] = (op << 28) | 2;
1689 set_long(pc, e);
1690 if (pred >= 0)
1691 set_pred(pc, cc, pred, e);
1692
1693 emit(pc, e);
1694 return e;
1695 }
1696
1697 static INLINE struct nv50_program_exec *
1698 emit_breakaddr(struct nv50_pc *pc)
1699 {
1700 return emit_control_flow(pc, 0x4, -1, 0);
1701 }
1702
1703 static INLINE void
1704 emit_break(struct nv50_pc *pc, int pred, unsigned cc)
1705 {
1706 emit_control_flow(pc, 0x5, pred, cc);
1707 }
1708
1709 static INLINE struct nv50_program_exec *
1710 emit_joinat(struct nv50_pc *pc)
1711 {
1712 return emit_control_flow(pc, 0xa, -1, 0);
1713 }
1714
1715 static INLINE struct nv50_program_exec *
1716 emit_branch(struct nv50_pc *pc, int pred, unsigned cc)
1717 {
1718 return emit_control_flow(pc, 0x1, pred, cc);
1719 }
1720
1721 static INLINE struct nv50_program_exec *
1722 emit_call(struct nv50_pc *pc, int pred, unsigned cc)
1723 {
1724 return emit_control_flow(pc, 0x2, pred, cc);
1725 }
1726
1727 static INLINE void
1728 emit_ret(struct nv50_pc *pc, int pred, unsigned cc)
1729 {
1730 emit_control_flow(pc, 0x3, pred, cc);
1731 }
1732
1733 static void
1734 emit_prim_cmd(struct nv50_pc *pc, unsigned cmd)
1735 {
1736 struct nv50_program_exec *e = exec(pc);
1737
1738 e->inst[0] = 0xf0000000 | (cmd << 9);
1739 e->inst[1] = 0xc0000000;
1740 set_long(pc, e);
1741
1742 emit(pc, e);
1743 }
1744
1745 #define QOP_ADD 0
1746 #define QOP_SUBR 1
1747 #define QOP_SUB 2
1748 #define QOP_MOV_SRC1 3
1749
1750 /* For a quad of threads / top left, top right, bottom left, bottom right
1751 * pixels, do a different operation, and take src0 from a specific thread.
1752 */
1753 static void
1754 emit_quadop(struct nv50_pc *pc, struct nv50_reg *dst, int wp, int lane_src0,
1755 struct nv50_reg *src0, struct nv50_reg *src1, ubyte qop)
1756 {
1757 struct nv50_program_exec *e = exec(pc);
1758
1759 e->inst[0] = 0xc0000000;
1760 e->inst[1] = 0x80000000;
1761 set_long(pc, e);
1762 e->inst[0] |= lane_src0 << 16;
1763 set_src_0(pc, src0, e);
1764 set_src_2(pc, src1, e);
1765
1766 if (wp >= 0)
1767 set_pred_wr(pc, 1, wp, e);
1768
1769 if (dst)
1770 set_dst(pc, dst, e);
1771 else {
1772 e->inst[0] |= 0x000001fc;
1773 e->inst[1] |= 0x00000008;
1774 }
1775
1776 e->inst[0] |= (qop & 3) << 20;
1777 e->inst[1] |= (qop >> 2) << 22;
1778
1779 emit(pc, e);
1780 }
1781
1782 static void
1783 load_cube_tex_coords(struct nv50_pc *pc, struct nv50_reg *t[4],
1784 struct nv50_reg **src, unsigned arg, boolean proj)
1785 {
1786 int mod[3] = { src[0]->mod, src[1]->mod, src[2]->mod };
1787
1788 src[0]->mod |= NV50_MOD_ABS;
1789 src[1]->mod |= NV50_MOD_ABS;
1790 src[2]->mod |= NV50_MOD_ABS;
1791
1792 emit_minmax(pc, NV50_MAX_F32, t[2], src[0], src[1]);
1793 emit_minmax(pc, NV50_MAX_F32, t[2], src[2], t[2]);
1794
1795 src[0]->mod = mod[0];
1796 src[1]->mod = mod[1];
1797 src[2]->mod = mod[2];
1798
1799 if (proj && 0 /* looks more correct without this */)
1800 emit_mul(pc, t[2], t[2], src[3]);
1801 else
1802 if (arg == 4) /* there is no textureProj(samplerCubeShadow) */
1803 emit_mov(pc, t[3], src[3]);
1804
1805 emit_flop(pc, NV50_FLOP_RCP, t[2], t[2]);
1806
1807 emit_mul(pc, t[0], src[0], t[2]);
1808 emit_mul(pc, t[1], src[1], t[2]);
1809 emit_mul(pc, t[2], src[2], t[2]);
1810 }
1811
1812 static void
1813 load_proj_tex_coords(struct nv50_pc *pc, struct nv50_reg *t[4],
1814 struct nv50_reg **src, unsigned dim, unsigned arg)
1815 {
1816 unsigned c, mode;
1817
1818 if (src[0]->type == P_TEMP && src[0]->rhw != -1) {
1819 mode = pc->interp_mode[src[0]->index] | INTERP_PERSPECTIVE;
1820
1821 t[3]->rhw = src[3]->rhw;
1822 emit_interp(pc, t[3], NULL, (mode & INTERP_CENTROID));
1823 emit_flop(pc, NV50_FLOP_RCP, t[3], t[3]);
1824
1825 for (c = 0; c < dim; ++c) {
1826 t[c]->rhw = src[c]->rhw;
1827 emit_interp(pc, t[c], t[3], mode);
1828 }
1829 if (arg != dim) { /* depth reference value */
1830 t[dim]->rhw = src[2]->rhw;
1831 emit_interp(pc, t[dim], t[3], mode);
1832 }
1833 } else {
1834 /* XXX: for some reason the blob sometimes uses MAD
1835 * (mad f32 $rX $rY $rZ neg $r63)
1836 */
1837 emit_flop(pc, NV50_FLOP_RCP, t[3], src[3]);
1838 for (c = 0; c < dim; ++c)
1839 emit_mul(pc, t[c], src[c], t[3]);
1840 if (arg != dim) /* depth reference value */
1841 emit_mul(pc, t[dim], src[2], t[3]);
1842 }
1843 }
1844
1845 static INLINE void
1846 get_tex_dim(unsigned type, unsigned *dim, unsigned *arg)
1847 {
1848 switch (type) {
1849 case TGSI_TEXTURE_1D:
1850 *arg = *dim = 1;
1851 break;
1852 case TGSI_TEXTURE_SHADOW1D:
1853 *dim = 1;
1854 *arg = 2;
1855 break;
1856 case TGSI_TEXTURE_UNKNOWN:
1857 case TGSI_TEXTURE_2D:
1858 case TGSI_TEXTURE_RECT:
1859 *arg = *dim = 2;
1860 break;
1861 case TGSI_TEXTURE_SHADOW2D:
1862 case TGSI_TEXTURE_SHADOWRECT:
1863 *dim = 2;
1864 *arg = 3;
1865 break;
1866 case TGSI_TEXTURE_3D:
1867 case TGSI_TEXTURE_CUBE:
1868 *dim = *arg = 3;
1869 break;
1870 default:
1871 assert(0);
1872 break;
1873 }
1874 }
1875
1876 /* We shouldn't execute TEXLOD if any of the pixels in a quad have
1877 * different LOD values, so branch off groups of equal LOD.
1878 */
1879 static void
1880 emit_texlod_sequence(struct nv50_pc *pc, struct nv50_reg *tlod,
1881 struct nv50_reg *src, struct nv50_program_exec *tex)
1882 {
1883 struct nv50_program_exec *join_at;
1884 unsigned i, target = pc->p->exec_size + 9 * 2;
1885
1886 if (pc->p->type != PIPE_SHADER_FRAGMENT) {
1887 emit(pc, tex);
1888 return;
1889 }
1890 pc->allow32 = FALSE;
1891
1892 /* Subtract lod of each pixel from lod of top left pixel, jump
1893 * texlod insn if result is 0, then repeat for 2 other pixels.
1894 */
1895 join_at = emit_joinat(pc);
1896 emit_quadop(pc, NULL, 0, 0, tlod, tlod, 0x55);
1897 emit_branch(pc, 0, 2)->param.index = target;
1898
1899 for (i = 1; i < 4; ++i) {
1900 emit_quadop(pc, NULL, 0, i, tlod, tlod, 0x55);
1901 emit_branch(pc, 0, 2)->param.index = target;
1902 }
1903
1904 emit_mov(pc, tlod, src); /* target */
1905 emit(pc, tex); /* texlod */
1906
1907 join_at->param.index = target + 2 * 2;
1908 JOIN_ON(emit_nop(pc)); /* join _after_ tex */
1909 }
1910
1911 static void
1912 emit_texbias_sequence(struct nv50_pc *pc, struct nv50_reg *t[4], unsigned arg,
1913 struct nv50_program_exec *tex)
1914 {
1915 struct nv50_program_exec *e;
1916 struct nv50_reg imm_1248, *t123[4][4], *r_bits = alloc_temp(pc, NULL);
1917 int r_pred = 0;
1918 unsigned n, c, i, cc[4] = { 0x0a, 0x13, 0x11, 0x10 };
1919
1920 pc->allow32 = FALSE;
1921 ctor_reg(&imm_1248, P_IMMD, -1, ctor_immd_4u32(pc, 1, 2, 4, 8) * 4);
1922
1923 /* Subtract bias value of thread i from bias values of each thread,
1924 * store result in r_pred, and set bit i in r_bits if result was 0.
1925 */
1926 assert(arg < 4);
1927 for (i = 0; i < 4; ++i, ++imm_1248.hw) {
1928 emit_quadop(pc, NULL, r_pred, i, t[arg], t[arg], 0x55);
1929 emit_mov(pc, r_bits, &imm_1248);
1930 set_pred(pc, 2, r_pred, pc->p->exec_tail);
1931 }
1932 emit_mov_to_pred(pc, r_pred, r_bits);
1933
1934 /* The lanes of a quad are now grouped by the bit in r_pred they have
1935 * set. Put the input values for TEX into a new register set for each
1936 * group and execute TEX only for a specific group.
1937 * We cannot use the same register set for each group because we need
1938 * the derivatives, which are implicitly calculated, to be correct.
1939 */
1940 for (i = 1; i < 4; ++i) {
1941 alloc_temp4(pc, t123[i], 0);
1942
1943 for (c = 0; c <= arg; ++c)
1944 emit_mov(pc, t123[i][c], t[c]);
1945
1946 *(e = exec(pc)) = *(tex);
1947 e->inst[0] &= ~0x01fc;
1948 set_dst(pc, t123[i][0], e);
1949 set_pred(pc, cc[i], r_pred, e);
1950 emit(pc, e);
1951 }
1952 /* finally TEX on the original regs (where we kept the input) */
1953 set_pred(pc, cc[0], r_pred, tex);
1954 emit(pc, tex);
1955
1956 /* put the 3 * n other results into regs for lane 0 */
1957 n = popcnt4(((e->inst[0] >> 25) & 0x3) | ((e->inst[1] >> 12) & 0xc));
1958 for (i = 1; i < 4; ++i) {
1959 for (c = 0; c < n; ++c) {
1960 emit_mov(pc, t[c], t123[i][c]);
1961 set_pred(pc, cc[i], r_pred, pc->p->exec_tail);
1962 }
1963 free_temp4(pc, t123[i]);
1964 }
1965
1966 emit_nop(pc);
1967 free_temp(pc, r_bits);
1968 }
1969
1970 static void
1971 emit_tex(struct nv50_pc *pc, struct nv50_reg **dst, unsigned mask,
1972 struct nv50_reg **src, unsigned unit, unsigned type,
1973 boolean proj, int bias_lod)
1974 {
1975 struct nv50_reg *t[4];
1976 struct nv50_program_exec *e;
1977 unsigned c, dim, arg;
1978
1979 /* t[i] must be within a single 128 bit super-reg */
1980 alloc_temp4(pc, t, 0);
1981
1982 e = exec(pc);
1983 e->inst[0] = 0xf0000000;
1984 set_long(pc, e);
1985 set_dst(pc, t[0], e);
1986
1987 /* TIC and TSC binding indices (TSC is ignored as TSC_LINKED = TRUE): */
1988 e->inst[0] |= (unit << 9) /* | (unit << 17) */;
1989
1990 /* live flag (don't set if TEX results affect input to another TEX): */
1991 /* e->inst[0] |= 0x00000004; */
1992
1993 get_tex_dim(type, &dim, &arg);
1994
1995 if (type == TGSI_TEXTURE_CUBE) {
1996 e->inst[0] |= 0x08000000;
1997 load_cube_tex_coords(pc, t, src, arg, proj);
1998 } else
1999 if (proj)
2000 load_proj_tex_coords(pc, t, src, dim, arg);
2001 else {
2002 for (c = 0; c < dim; c++)
2003 emit_mov(pc, t[c], src[c]);
2004 if (arg != dim) /* depth reference value (always src.z here) */
2005 emit_mov(pc, t[dim], src[2]);
2006 }
2007
2008 e->inst[0] |= (mask & 0x3) << 25;
2009 e->inst[1] |= (mask & 0xc) << 12;
2010
2011 if (!bias_lod) {
2012 e->inst[0] |= (arg - 1) << 22;
2013 emit(pc, e);
2014 } else
2015 if (bias_lod < 0) {
2016 assert(pc->p->type == PIPE_SHADER_FRAGMENT);
2017 e->inst[0] |= arg << 22;
2018 e->inst[1] |= 0x20000000; /* texbias */
2019 emit_mov(pc, t[arg], src[3]);
2020 emit_texbias_sequence(pc, t, arg, e);
2021 } else {
2022 e->inst[0] |= arg << 22;
2023 e->inst[1] |= 0x40000000; /* texlod */
2024 emit_mov(pc, t[arg], src[3]);
2025 emit_texlod_sequence(pc, t[arg], src[3], e);
2026 }
2027
2028 #if 1
2029 c = 0;
2030 if (mask & 1) emit_mov(pc, dst[0], t[c++]);
2031 if (mask & 2) emit_mov(pc, dst[1], t[c++]);
2032 if (mask & 4) emit_mov(pc, dst[2], t[c++]);
2033 if (mask & 8) emit_mov(pc, dst[3], t[c]);
2034
2035 free_temp4(pc, t);
2036 #else
2037 /* XXX: if p.e. MUL is used directly after TEX, it would still use
2038 * the texture coordinates, not the fetched values: latency ? */
2039
2040 for (c = 0; c < 4; c++) {
2041 if (mask & (1 << c))
2042 assimilate_temp(pc, dst[c], t[c]);
2043 else
2044 free_temp(pc, t[c]);
2045 }
2046 #endif
2047 }
2048
2049 static void
2050 emit_ddx(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
2051 {
2052 struct nv50_program_exec *e = exec(pc);
2053
2054 assert(src->type == P_TEMP);
2055
2056 e->inst[0] = (src->mod & NV50_MOD_NEG) ? 0xc0240000 : 0xc0140000;
2057 e->inst[1] = (src->mod & NV50_MOD_NEG) ? 0x86400000 : 0x89800000;
2058 set_long(pc, e);
2059 set_dst(pc, dst, e);
2060 set_src_0(pc, src, e);
2061 set_src_2(pc, src, e);
2062
2063 emit(pc, e);
2064 }
2065
2066 static void
2067 emit_ddy(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
2068 {
2069 struct nv50_program_exec *e = exec(pc);
2070
2071 assert(src->type == P_TEMP);
2072
2073 e->inst[0] = (src->mod & NV50_MOD_NEG) ? 0xc0250000 : 0xc0150000;
2074 e->inst[1] = (src->mod & NV50_MOD_NEG) ? 0x85800000 : 0x8a400000;
2075 set_long(pc, e);
2076 set_dst(pc, dst, e);
2077 set_src_0(pc, src, e);
2078 set_src_2(pc, src, e);
2079
2080 emit(pc, e);
2081 }
2082
2083 static void
2084 convert_to_long(struct nv50_pc *pc, struct nv50_program_exec *e)
2085 {
2086 unsigned q = 0, m = ~0;
2087
2088 assert(!is_long(e));
2089
2090 switch (e->inst[0] >> 28) {
2091 case 0x1:
2092 /* MOV */
2093 q = 0x0403c000;
2094 m = 0xffff7fff;
2095 break;
2096 case 0x2:
2097 case 0x3:
2098 /* ADD, SUB, SUBR b32 */
2099 m = ~(0x8000 | (127 << 16));
2100 q = ((e->inst[0] & (~m)) >> 2) | (1 << 26);
2101 break;
2102 case 0x5:
2103 /* SAD */
2104 m = ~(0x81 << 8);
2105 q = (0x0c << 24) | ((e->inst[0] & (0x7f << 2)) << 12);
2106 break;
2107 case 0x6:
2108 /* MAD u16 */
2109 q = (e->inst[0] & (0x7f << 2)) << 12;
2110 break;
2111 case 0x8:
2112 /* INTERP (move centroid, perspective and flat bits) */
2113 m = ~0x03000100;
2114 q = (e->inst[0] & (3 << 24)) >> (24 - 16);
2115 q |= (e->inst[0] & (1 << 8)) << (18 - 8);
2116 break;
2117 case 0x9:
2118 /* RCP */
2119 break;
2120 case 0xB:
2121 /* ADD */
2122 m = ~(127 << 16);
2123 q = ((e->inst[0] & (~m)) >> 2);
2124 break;
2125 case 0xC:
2126 /* MUL */
2127 m = ~0x00008000;
2128 q = ((e->inst[0] & (~m)) << 12);
2129 break;
2130 case 0xE:
2131 /* MAD (if src2 == dst) */
2132 q = ((e->inst[0] & 0x1fc) << 12);
2133 break;
2134 default:
2135 assert(0);
2136 break;
2137 }
2138
2139 set_long(pc, e);
2140 pc->p->exec_size++;
2141
2142 e->inst[0] &= m;
2143 e->inst[1] |= q;
2144 }
2145
2146 /* Some operations support an optional negation flag. */
2147 static int
2148 get_supported_mods(const struct tgsi_full_instruction *insn, int i)
2149 {
2150 switch (insn->Instruction.Opcode) {
2151 case TGSI_OPCODE_ADD:
2152 case TGSI_OPCODE_COS:
2153 case TGSI_OPCODE_DDX:
2154 case TGSI_OPCODE_DDY:
2155 case TGSI_OPCODE_DP3:
2156 case TGSI_OPCODE_DP4:
2157 case TGSI_OPCODE_EX2:
2158 case TGSI_OPCODE_KIL:
2159 case TGSI_OPCODE_LG2:
2160 case TGSI_OPCODE_MAD:
2161 case TGSI_OPCODE_MUL:
2162 case TGSI_OPCODE_POW:
2163 case TGSI_OPCODE_RCP:
2164 case TGSI_OPCODE_RSQ: /* ignored, RSQ = rsqrt(abs(src.x)) */
2165 case TGSI_OPCODE_SCS:
2166 case TGSI_OPCODE_SIN:
2167 case TGSI_OPCODE_SUB:
2168 return NV50_MOD_NEG;
2169 case TGSI_OPCODE_MAX:
2170 case TGSI_OPCODE_MIN:
2171 case TGSI_OPCODE_INEG: /* tgsi src sign toggle/set would be stupid */
2172 return NV50_MOD_ABS;
2173 case TGSI_OPCODE_CEIL:
2174 case TGSI_OPCODE_FLR:
2175 case TGSI_OPCODE_TRUNC:
2176 return NV50_MOD_NEG | NV50_MOD_ABS;
2177 case TGSI_OPCODE_F2I:
2178 case TGSI_OPCODE_F2U:
2179 case TGSI_OPCODE_I2F:
2180 case TGSI_OPCODE_U2F:
2181 return NV50_MOD_NEG | NV50_MOD_ABS | NV50_MOD_I32;
2182 case TGSI_OPCODE_UADD:
2183 return NV50_MOD_NEG | NV50_MOD_I32;
2184 case TGSI_OPCODE_SAD:
2185 case TGSI_OPCODE_SHL:
2186 case TGSI_OPCODE_IMAX:
2187 case TGSI_OPCODE_IMIN:
2188 case TGSI_OPCODE_ISHR:
2189 case TGSI_OPCODE_NOT:
2190 case TGSI_OPCODE_UMAD:
2191 case TGSI_OPCODE_UMAX:
2192 case TGSI_OPCODE_UMIN:
2193 case TGSI_OPCODE_UMUL:
2194 case TGSI_OPCODE_USHR:
2195 return NV50_MOD_I32;
2196 default:
2197 return 0;
2198 }
2199 }
2200
2201 /* Return a read mask for source registers deduced from opcode & write mask. */
2202 static unsigned
2203 nv50_tgsi_src_mask(const struct tgsi_full_instruction *insn, int c)
2204 {
2205 unsigned x, mask = insn->Dst[0].Register.WriteMask;
2206
2207 switch (insn->Instruction.Opcode) {
2208 case TGSI_OPCODE_COS:
2209 case TGSI_OPCODE_SIN:
2210 return (mask & 0x8) | ((mask & 0x7) ? 0x1 : 0x0);
2211 case TGSI_OPCODE_DP3:
2212 return 0x7;
2213 case TGSI_OPCODE_DP4:
2214 case TGSI_OPCODE_DPH:
2215 case TGSI_OPCODE_KIL: /* WriteMask ignored */
2216 return 0xf;
2217 case TGSI_OPCODE_DST:
2218 return mask & (c ? 0xa : 0x6);
2219 case TGSI_OPCODE_EX2:
2220 case TGSI_OPCODE_EXP:
2221 case TGSI_OPCODE_LG2:
2222 case TGSI_OPCODE_LOG:
2223 case TGSI_OPCODE_POW:
2224 case TGSI_OPCODE_RCP:
2225 case TGSI_OPCODE_RSQ:
2226 case TGSI_OPCODE_SCS:
2227 return 0x1;
2228 case TGSI_OPCODE_IF:
2229 return 0x1;
2230 case TGSI_OPCODE_LIT:
2231 return 0xb;
2232 case TGSI_OPCODE_TEX:
2233 case TGSI_OPCODE_TXB:
2234 case TGSI_OPCODE_TXL:
2235 case TGSI_OPCODE_TXP:
2236 {
2237 const struct tgsi_instruction_texture *tex;
2238
2239 assert(insn->Instruction.Texture);
2240 tex = &insn->Texture;
2241
2242 mask = 0x7;
2243 if (insn->Instruction.Opcode != TGSI_OPCODE_TEX &&
2244 insn->Instruction.Opcode != TGSI_OPCODE_TXD)
2245 mask |= 0x8; /* bias, lod or proj */
2246
2247 switch (tex->Texture) {
2248 case TGSI_TEXTURE_1D:
2249 mask &= 0x9;
2250 break;
2251 case TGSI_TEXTURE_SHADOW1D:
2252 mask &= 0x5;
2253 break;
2254 case TGSI_TEXTURE_2D:
2255 mask &= 0xb;
2256 break;
2257 default:
2258 break;
2259 }
2260 }
2261 return mask;
2262 case TGSI_OPCODE_XPD:
2263 x = 0;
2264 if (mask & 1) x |= 0x6;
2265 if (mask & 2) x |= 0x5;
2266 if (mask & 4) x |= 0x3;
2267 return x;
2268 default:
2269 break;
2270 }
2271
2272 return mask;
2273 }
2274
2275 static struct nv50_reg *
2276 tgsi_dst(struct nv50_pc *pc, int c, const struct tgsi_full_dst_register *dst)
2277 {
2278 switch (dst->Register.File) {
2279 case TGSI_FILE_TEMPORARY:
2280 return &pc->temp[dst->Register.Index * 4 + c];
2281 case TGSI_FILE_OUTPUT:
2282 return &pc->result[dst->Register.Index * 4 + c];
2283 case TGSI_FILE_ADDRESS:
2284 {
2285 struct nv50_reg *r = pc->addr[dst->Register.Index * 4 + c];
2286 if (!r) {
2287 r = get_address_reg(pc, NULL);
2288 r->index = dst->Register.Index * 4 + c;
2289 pc->addr[r->index] = r;
2290 }
2291 assert(r);
2292 return r;
2293 }
2294 case TGSI_FILE_NULL:
2295 return NULL;
2296 case TGSI_FILE_SYSTEM_VALUE:
2297 assert(pc->sysval[dst->Register.Index].type == P_RESULT);
2298 assert(c == 0);
2299 return &pc->sysval[dst->Register.Index];
2300 default:
2301 break;
2302 }
2303
2304 return NULL;
2305 }
2306
2307 static struct nv50_reg *
2308 tgsi_src(struct nv50_pc *pc, int chan, const struct tgsi_full_src_register *src,
2309 int mod)
2310 {
2311 struct nv50_reg *r = NULL;
2312 struct nv50_reg *temp = NULL;
2313 unsigned sgn, c, swz, cvn;
2314
2315 if (src->Register.File != TGSI_FILE_CONSTANT)
2316 assert(!src->Register.Indirect);
2317
2318 sgn = tgsi_util_get_full_src_register_sign_mode(src, chan);
2319
2320 c = tgsi_util_get_full_src_register_swizzle(src, chan);
2321 switch (c) {
2322 case TGSI_SWIZZLE_X:
2323 case TGSI_SWIZZLE_Y:
2324 case TGSI_SWIZZLE_Z:
2325 case TGSI_SWIZZLE_W:
2326 switch (src->Register.File) {
2327 case TGSI_FILE_INPUT:
2328 r = &pc->attr[src->Register.Index * 4 + c];
2329
2330 if (!src->Dimension.Dimension)
2331 break;
2332 r = reg_instance(pc, r);
2333 r->vtx = src->Dimension.Index;
2334
2335 if (!src->Dimension.Indirect)
2336 break;
2337 swz = tgsi_util_get_src_register_swizzle(
2338 &src->DimIndirect, 0);
2339 r->acc = -1;
2340 r->indirect[1] = src->DimIndirect.Index * 4 + swz;
2341 break;
2342 case TGSI_FILE_TEMPORARY:
2343 r = &pc->temp[src->Register.Index * 4 + c];
2344 break;
2345 case TGSI_FILE_CONSTANT:
2346 if (!src->Register.Indirect) {
2347 r = &pc->param[src->Register.Index * 4 + c];
2348 break;
2349 }
2350 /* Indicate indirection by setting r->acc < 0 and
2351 * use the index field to select the address reg.
2352 */
2353 r = reg_instance(pc, NULL);
2354 ctor_reg(r, P_CONST, -1, src->Register.Index * 4 + c);
2355
2356 swz = tgsi_util_get_src_register_swizzle(
2357 &src->Indirect, 0);
2358 r->acc = -1;
2359 r->indirect[0] = src->Indirect.Index * 4 + swz;
2360 break;
2361 case TGSI_FILE_IMMEDIATE:
2362 r = &pc->immd[src->Register.Index * 4 + c];
2363 break;
2364 case TGSI_FILE_SAMPLER:
2365 return NULL;
2366 case TGSI_FILE_ADDRESS:
2367 r = pc->addr[src->Register.Index * 4 + c];
2368 assert(r);
2369 break;
2370 case TGSI_FILE_SYSTEM_VALUE:
2371 assert(c == 0);
2372 r = &pc->sysval[src->Register.Index];
2373 break;
2374 default:
2375 assert(0);
2376 break;
2377 }
2378 break;
2379 default:
2380 assert(0);
2381 break;
2382 }
2383
2384 cvn = (mod & NV50_MOD_I32) ? CVT_S32_S32 : CVT_F32_F32;
2385
2386 switch (sgn) {
2387 case TGSI_UTIL_SIGN_CLEAR:
2388 r->mod = NV50_MOD_ABS;
2389 break;
2390 case TGSI_UTIL_SIGN_SET:
2391 r->mod = NV50_MOD_NEG_ABS;
2392 break;
2393 case TGSI_UTIL_SIGN_TOGGLE:
2394 r->mod = NV50_MOD_NEG;
2395 break;
2396 default:
2397 assert(!r->mod && sgn == TGSI_UTIL_SIGN_KEEP);
2398 break;
2399 }
2400
2401 if ((r->mod & mod) != r->mod) {
2402 temp = temp_temp(pc, NULL);
2403 emit_cvt(pc, temp, r, -1, cvn);
2404 r->mod = 0;
2405 r = temp;
2406 } else
2407 r->mod |= mod & NV50_MOD_I32;
2408
2409 assert(r);
2410 if (r->acc >= 0 && r->vtx < 0 && r != temp)
2411 return reg_instance(pc, r); /* will clear r->mod */
2412 return r;
2413 }
2414
2415 /* return TRUE for ops that produce only a single result */
2416 static boolean
2417 is_scalar_op(unsigned op)
2418 {
2419 switch (op) {
2420 case TGSI_OPCODE_COS:
2421 case TGSI_OPCODE_DP2:
2422 case TGSI_OPCODE_DP3:
2423 case TGSI_OPCODE_DP4:
2424 case TGSI_OPCODE_DPH:
2425 case TGSI_OPCODE_EX2:
2426 case TGSI_OPCODE_LG2:
2427 case TGSI_OPCODE_POW:
2428 case TGSI_OPCODE_RCP:
2429 case TGSI_OPCODE_RSQ:
2430 case TGSI_OPCODE_SIN:
2431 /*
2432 case TGSI_OPCODE_KIL:
2433 case TGSI_OPCODE_LIT:
2434 case TGSI_OPCODE_SCS:
2435 */
2436 return TRUE;
2437 default:
2438 return FALSE;
2439 }
2440 }
2441
2442 /* Returns a bitmask indicating which dst components depend
2443 * on source s, component c (reverse of nv50_tgsi_src_mask).
2444 */
2445 static unsigned
2446 nv50_tgsi_dst_revdep(unsigned op, int s, int c)
2447 {
2448 if (is_scalar_op(op))
2449 return 0x1;
2450
2451 switch (op) {
2452 case TGSI_OPCODE_DST:
2453 return (1 << c) & (s ? 0xa : 0x6);
2454 case TGSI_OPCODE_XPD:
2455 switch (c) {
2456 case 0: return 0x6;
2457 case 1: return 0x5;
2458 case 2: return 0x3;
2459 case 3: return 0x0;
2460 default:
2461 assert(0);
2462 return 0x0;
2463 }
2464 case TGSI_OPCODE_EXP:
2465 case TGSI_OPCODE_LOG:
2466 case TGSI_OPCODE_LIT:
2467 case TGSI_OPCODE_SCS:
2468 case TGSI_OPCODE_TEX:
2469 case TGSI_OPCODE_TXB:
2470 case TGSI_OPCODE_TXL:
2471 case TGSI_OPCODE_TXP:
2472 /* these take care of dangerous swizzles themselves */
2473 return 0x0;
2474 case TGSI_OPCODE_IF:
2475 case TGSI_OPCODE_KIL:
2476 /* don't call this function for these ops */
2477 assert(0);
2478 return 0;
2479 default:
2480 /* linear vector instruction */
2481 return (1 << c);
2482 }
2483 }
2484
2485 static INLINE boolean
2486 has_pred(struct nv50_program_exec *e, unsigned cc)
2487 {
2488 if (!is_long(e) || is_immd(e))
2489 return FALSE;
2490 return ((e->inst[1] & 0x780) == (cc << 7));
2491 }
2492
2493 /* on ENDIF see if we can do "@p0.neu single_op" instead of:
2494 * join_at ENDIF
2495 * @p0.eq bra ENDIF
2496 * single_op
2497 * ENDIF: nop.join
2498 */
2499 static boolean
2500 nv50_kill_branch(struct nv50_pc *pc)
2501 {
2502 int lvl = pc->if_lvl;
2503
2504 if (pc->if_insn[lvl]->next != pc->p->exec_tail)
2505 return FALSE;
2506 if (is_immd(pc->p->exec_tail))
2507 return FALSE;
2508
2509 /* if ccode == 'true', the BRA is from an ELSE and the predicate
2510 * reg may no longer be valid, since we currently always use $p0
2511 */
2512 if (has_pred(pc->if_insn[lvl], 0xf))
2513 return FALSE;
2514 assert(pc->if_insn[lvl] && pc->if_join[lvl]);
2515
2516 /* We'll use the exec allocated for JOIN_AT (we can't easily
2517 * access nv50_program_exec's prev).
2518 */
2519 pc->p->exec_size -= 4; /* remove JOIN_AT and BRA */
2520
2521 *pc->if_join[lvl] = *pc->p->exec_tail;
2522
2523 FREE(pc->if_insn[lvl]);
2524 FREE(pc->p->exec_tail);
2525
2526 pc->p->exec_tail = pc->if_join[lvl];
2527 pc->p->exec_tail->next = NULL;
2528 set_pred(pc, 0xd, 0, pc->p->exec_tail);
2529
2530 return TRUE;
2531 }
2532
2533 static void
2534 nv50_fp_move_results(struct nv50_pc *pc)
2535 {
2536 struct nv50_reg reg;
2537 unsigned i;
2538
2539 ctor_reg(&reg, P_TEMP, -1, -1);
2540
2541 for (i = 0; i < pc->result_nr * 4; ++i) {
2542 if (pc->result[i].rhw < 0 || pc->result[i].hw < 0)
2543 continue;
2544 if (pc->result[i].rhw != pc->result[i].hw) {
2545 reg.hw = pc->result[i].rhw;
2546 emit_mov(pc, &reg, &pc->result[i]);
2547 }
2548 }
2549 }
2550
2551 static boolean
2552 nv50_program_tx_insn(struct nv50_pc *pc,
2553 const struct tgsi_full_instruction *inst)
2554 {
2555 struct nv50_reg *rdst[4], *dst[4], *brdc, *src[3][4], *temp;
2556 unsigned mask, sat, unit = 0;
2557 int i, c;
2558
2559 mask = inst->Dst[0].Register.WriteMask;
2560 sat = inst->Instruction.Saturate == TGSI_SAT_ZERO_ONE;
2561
2562 memset(src, 0, sizeof(src));
2563
2564 for (c = 0; c < 4; c++) {
2565 if ((mask & (1 << c)) && !pc->r_dst[c])
2566 dst[c] = tgsi_dst(pc, c, &inst->Dst[0]);
2567 else
2568 dst[c] = pc->r_dst[c];
2569 rdst[c] = dst[c];
2570 }
2571
2572 for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
2573 const struct tgsi_full_src_register *fs = &inst->Src[i];
2574 unsigned src_mask;
2575 int mod_supp;
2576
2577 src_mask = nv50_tgsi_src_mask(inst, i);
2578 mod_supp = get_supported_mods(inst, i);
2579
2580 if (fs->Register.File == TGSI_FILE_SAMPLER)
2581 unit = fs->Register.Index;
2582
2583 for (c = 0; c < 4; c++)
2584 if (src_mask & (1 << c))
2585 src[i][c] = tgsi_src(pc, c, fs, mod_supp);
2586 }
2587
2588 brdc = temp = pc->r_brdc;
2589 if (brdc && brdc->type != P_TEMP) {
2590 temp = temp_temp(pc, NULL);
2591 if (sat)
2592 brdc = temp;
2593 } else
2594 if (sat) {
2595 for (c = 0; c < 4; c++) {
2596 if (!(mask & (1 << c)) || dst[c]->type == P_TEMP)
2597 continue;
2598 /* rdst[c] = dst[c]; */ /* done above */
2599 dst[c] = temp_temp(pc, NULL);
2600 }
2601 }
2602
2603 assert(brdc || !is_scalar_op(inst->Instruction.Opcode));
2604
2605 switch (inst->Instruction.Opcode) {
2606 case TGSI_OPCODE_ABS:
2607 for (c = 0; c < 4; c++) {
2608 if (!(mask & (1 << c)))
2609 continue;
2610 emit_cvt(pc, dst[c], src[0][c], -1,
2611 CVT_ABS | CVT_F32_F32);
2612 }
2613 break;
2614 case TGSI_OPCODE_ADD:
2615 for (c = 0; c < 4; c++) {
2616 if (!(mask & (1 << c)))
2617 continue;
2618 emit_add(pc, dst[c], src[0][c], src[1][c]);
2619 }
2620 break;
2621 case TGSI_OPCODE_AND:
2622 case TGSI_OPCODE_XOR:
2623 case TGSI_OPCODE_OR:
2624 for (c = 0; c < 4; c++) {
2625 if (!(mask & (1 << c)))
2626 continue;
2627 emit_bitop2(pc, dst[c], src[0][c], src[1][c],
2628 inst->Instruction.Opcode);
2629 }
2630 break;
2631 case TGSI_OPCODE_ARL:
2632 temp = temp_temp(pc, NULL);
2633 for (c = 0; c < 4; c++) {
2634 if (!(mask & (1 << c)))
2635 continue;
2636 emit_cvt(pc, temp, src[0][c], -1,
2637 CVT_FLOOR | CVT_S32_F32);
2638 emit_arl(pc, dst[c], temp, 4);
2639 }
2640 break;
2641 case TGSI_OPCODE_BGNLOOP:
2642 pc->loop_brka[pc->loop_lvl] = emit_breakaddr(pc);
2643 pc->loop_pos[pc->loop_lvl++] = pc->p->exec_size;
2644 terminate_mbb(pc);
2645 break;
2646 case TGSI_OPCODE_BGNSUB:
2647 assert(!pc->in_subroutine);
2648 pc->in_subroutine = TRUE;
2649 /* probably not necessary, but align to 8 byte boundary */
2650 if (!is_long(pc->p->exec_tail))
2651 convert_to_long(pc, pc->p->exec_tail);
2652 break;
2653 case TGSI_OPCODE_BRK:
2654 assert(pc->loop_lvl > 0);
2655 emit_break(pc, -1, 0);
2656 break;
2657 case TGSI_OPCODE_CAL:
2658 assert(inst->Label.Label < pc->insn_nr);
2659 emit_call(pc, -1, 0)->param.index = inst->Label.Label;
2660 /* replaced by actual offset in nv50_program_fixup_insns */
2661 break;
2662 case TGSI_OPCODE_CEIL:
2663 for (c = 0; c < 4; c++) {
2664 if (!(mask & (1 << c)))
2665 continue;
2666 emit_cvt(pc, dst[c], src[0][c], -1,
2667 CVT_CEIL | CVT_F32_F32 | CVT_RI);
2668 }
2669 break;
2670 case TGSI_OPCODE_CMP:
2671 pc->allow32 = FALSE;
2672 for (c = 0; c < 4; c++) {
2673 if (!(mask & (1 << c)))
2674 continue;
2675 emit_cvt(pc, NULL, src[0][c], 1, CVT_F32_F32);
2676 emit_mov(pc, dst[c], src[1][c]);
2677 set_pred(pc, 0x1, 1, pc->p->exec_tail); /* @SF */
2678 emit_mov(pc, dst[c], src[2][c]);
2679 set_pred(pc, 0x6, 1, pc->p->exec_tail); /* @NSF */
2680 }
2681 break;
2682 case TGSI_OPCODE_CONT:
2683 assert(pc->loop_lvl > 0);
2684 emit_branch(pc, -1, 0)->param.index =
2685 pc->loop_pos[pc->loop_lvl - 1];
2686 break;
2687 case TGSI_OPCODE_COS:
2688 if (mask & 8) {
2689 emit_precossin(pc, temp, src[0][3]);
2690 emit_flop(pc, NV50_FLOP_COS, dst[3], temp);
2691 if (!(mask &= 7))
2692 break;
2693 if (temp == dst[3])
2694 temp = brdc = temp_temp(pc, NULL);
2695 }
2696 emit_precossin(pc, temp, src[0][0]);
2697 emit_flop(pc, NV50_FLOP_COS, brdc, temp);
2698 break;
2699 case TGSI_OPCODE_DDX:
2700 for (c = 0; c < 4; c++) {
2701 if (!(mask & (1 << c)))
2702 continue;
2703 emit_ddx(pc, dst[c], src[0][c]);
2704 }
2705 break;
2706 case TGSI_OPCODE_DDY:
2707 for (c = 0; c < 4; c++) {
2708 if (!(mask & (1 << c)))
2709 continue;
2710 emit_ddy(pc, dst[c], src[0][c]);
2711 }
2712 break;
2713 case TGSI_OPCODE_DP3:
2714 emit_mul(pc, temp, src[0][0], src[1][0]);
2715 emit_mad(pc, temp, src[0][1], src[1][1], temp);
2716 emit_mad(pc, brdc, src[0][2], src[1][2], temp);
2717 break;
2718 case TGSI_OPCODE_DP4:
2719 emit_mul(pc, temp, src[0][0], src[1][0]);
2720 emit_mad(pc, temp, src[0][1], src[1][1], temp);
2721 emit_mad(pc, temp, src[0][2], src[1][2], temp);
2722 emit_mad(pc, brdc, src[0][3], src[1][3], temp);
2723 break;
2724 case TGSI_OPCODE_DPH:
2725 emit_mul(pc, temp, src[0][0], src[1][0]);
2726 emit_mad(pc, temp, src[0][1], src[1][1], temp);
2727 emit_mad(pc, temp, src[0][2], src[1][2], temp);
2728 emit_add(pc, brdc, src[1][3], temp);
2729 break;
2730 case TGSI_OPCODE_DST:
2731 if (mask & (1 << 1))
2732 emit_mul(pc, dst[1], src[0][1], src[1][1]);
2733 if (mask & (1 << 2))
2734 emit_mov(pc, dst[2], src[0][2]);
2735 if (mask & (1 << 3))
2736 emit_mov(pc, dst[3], src[1][3]);
2737 if (mask & (1 << 0))
2738 emit_mov_immdval(pc, dst[0], 1.0f);
2739 break;
2740 case TGSI_OPCODE_ELSE:
2741 emit_branch(pc, -1, 0);
2742 pc->if_insn[--pc->if_lvl]->param.index = pc->p->exec_size;
2743 pc->if_insn[pc->if_lvl++] = pc->p->exec_tail;
2744 terminate_mbb(pc);
2745 break;
2746 case TGSI_OPCODE_EMIT:
2747 emit_prim_cmd(pc, 1);
2748 break;
2749 case TGSI_OPCODE_ENDIF:
2750 pc->if_insn[--pc->if_lvl]->param.index = pc->p->exec_size;
2751
2752 /* try to replace branch over 1 insn with a predicated insn */
2753 if (nv50_kill_branch(pc) == TRUE)
2754 break;
2755
2756 if (pc->if_join[pc->if_lvl]) {
2757 pc->if_join[pc->if_lvl]->param.index = pc->p->exec_size;
2758 pc->if_join[pc->if_lvl] = NULL;
2759 }
2760 terminate_mbb(pc);
2761 /* emit a NOP as join point, we could set it on the next
2762 * one, but would have to make sure it is long and !immd
2763 */
2764 JOIN_ON(emit_nop(pc));
2765 break;
2766 case TGSI_OPCODE_ENDLOOP:
2767 emit_branch(pc, -1, 0)->param.index =
2768 pc->loop_pos[--pc->loop_lvl];
2769 pc->loop_brka[pc->loop_lvl]->param.index = pc->p->exec_size;
2770 terminate_mbb(pc);
2771 break;
2772 case TGSI_OPCODE_ENDPRIM:
2773 emit_prim_cmd(pc, 2);
2774 break;
2775 case TGSI_OPCODE_ENDSUB:
2776 assert(pc->in_subroutine);
2777 terminate_mbb(pc);
2778 pc->in_subroutine = FALSE;
2779 break;
2780 case TGSI_OPCODE_EX2:
2781 emit_preex2(pc, temp, src[0][0]);
2782 emit_flop(pc, NV50_FLOP_EX2, brdc, temp);
2783 break;
2784 case TGSI_OPCODE_EXP:
2785 {
2786 struct nv50_reg *t[2];
2787
2788 assert(!temp);
2789 t[0] = temp_temp(pc, NULL);
2790 t[1] = temp_temp(pc, NULL);
2791
2792 if (mask & 0x6)
2793 emit_mov(pc, t[0], src[0][0]);
2794 if (mask & 0x3)
2795 emit_flr(pc, t[1], src[0][0]);
2796
2797 if (mask & (1 << 1))
2798 emit_sub(pc, dst[1], t[0], t[1]);
2799 if (mask & (1 << 0)) {
2800 emit_preex2(pc, t[1], t[1]);
2801 emit_flop(pc, NV50_FLOP_EX2, dst[0], t[1]);
2802 }
2803 if (mask & (1 << 2)) {
2804 emit_preex2(pc, t[0], t[0]);
2805 emit_flop(pc, NV50_FLOP_EX2, dst[2], t[0]);
2806 }
2807 if (mask & (1 << 3))
2808 emit_mov_immdval(pc, dst[3], 1.0f);
2809 }
2810 break;
2811 case TGSI_OPCODE_F2I:
2812 for (c = 0; c < 4; c++) {
2813 if (!(mask & (1 << c)))
2814 continue;
2815 emit_cvt(pc, dst[c], src[0][c], -1,
2816 CVT_TRUNC | CVT_S32_F32);
2817 }
2818 break;
2819 case TGSI_OPCODE_F2U:
2820 for (c = 0; c < 4; c++) {
2821 if (!(mask & (1 << c)))
2822 continue;
2823 emit_cvt(pc, dst[c], src[0][c], -1,
2824 CVT_TRUNC | CVT_U32_F32);
2825 }
2826 break;
2827 case TGSI_OPCODE_FLR:
2828 for (c = 0; c < 4; c++) {
2829 if (!(mask & (1 << c)))
2830 continue;
2831 emit_flr(pc, dst[c], src[0][c]);
2832 }
2833 break;
2834 case TGSI_OPCODE_FRC:
2835 temp = temp_temp(pc, NULL);
2836 for (c = 0; c < 4; c++) {
2837 if (!(mask & (1 << c)))
2838 continue;
2839 emit_flr(pc, temp, src[0][c]);
2840 emit_sub(pc, dst[c], src[0][c], temp);
2841 }
2842 break;
2843 case TGSI_OPCODE_I2F:
2844 for (c = 0; c < 4; c++) {
2845 if (!(mask & (1 << c)))
2846 continue;
2847 emit_cvt(pc, dst[c], src[0][c], -1, CVT_F32_S32);
2848 }
2849 break;
2850 case TGSI_OPCODE_IF:
2851 assert(pc->if_lvl < NV50_MAX_COND_NESTING);
2852 emit_cvt(pc, NULL, src[0][0], 0, CVT_ABS | CVT_F32_F32);
2853 pc->if_join[pc->if_lvl] = emit_joinat(pc);
2854 pc->if_insn[pc->if_lvl++] = emit_branch(pc, 0, 2);;
2855 terminate_mbb(pc);
2856 break;
2857 case TGSI_OPCODE_IMAX:
2858 for (c = 0; c < 4; c++) {
2859 if (!(mask & (1 << c)))
2860 continue;
2861 emit_minmax(pc, 0x08c, dst[c], src[0][c], src[1][c]);
2862 }
2863 break;
2864 case TGSI_OPCODE_IMIN:
2865 for (c = 0; c < 4; c++) {
2866 if (!(mask & (1 << c)))
2867 continue;
2868 emit_minmax(pc, 0x0ac, dst[c], src[0][c], src[1][c]);
2869 }
2870 break;
2871 case TGSI_OPCODE_INEG:
2872 for (c = 0; c < 4; c++) {
2873 if (!(mask & (1 << c)))
2874 continue;
2875 emit_cvt(pc, dst[c], src[0][c], -1,
2876 CVT_S32_S32 | CVT_NEG);
2877 }
2878 break;
2879 case TGSI_OPCODE_KIL:
2880 assert(src[0][0] && src[0][1] && src[0][2] && src[0][3]);
2881 emit_kil(pc, src[0][0]);
2882 emit_kil(pc, src[0][1]);
2883 emit_kil(pc, src[0][2]);
2884 emit_kil(pc, src[0][3]);
2885 break;
2886 case TGSI_OPCODE_KILP:
2887 emit_kil(pc, NULL);
2888 break;
2889 case TGSI_OPCODE_LIT:
2890 emit_lit(pc, &dst[0], mask, &src[0][0]);
2891 break;
2892 case TGSI_OPCODE_LG2:
2893 emit_flop(pc, NV50_FLOP_LG2, brdc, src[0][0]);
2894 break;
2895 case TGSI_OPCODE_LOG:
2896 {
2897 struct nv50_reg *t[2];
2898
2899 t[0] = temp_temp(pc, NULL);
2900 if (mask & (1 << 1))
2901 t[1] = temp_temp(pc, NULL);
2902 else
2903 t[1] = t[0];
2904
2905 emit_cvt(pc, t[0], src[0][0], -1, CVT_ABS | CVT_F32_F32);
2906 emit_flop(pc, NV50_FLOP_LG2, t[1], t[0]);
2907 if (mask & (1 << 2))
2908 emit_mov(pc, dst[2], t[1]);
2909 emit_flr(pc, t[1], t[1]);
2910 if (mask & (1 << 0))
2911 emit_mov(pc, dst[0], t[1]);
2912 if (mask & (1 << 1)) {
2913 t[1]->mod = NV50_MOD_NEG;
2914 emit_preex2(pc, t[1], t[1]);
2915 t[1]->mod = 0;
2916 emit_flop(pc, NV50_FLOP_EX2, t[1], t[1]);
2917 emit_mul(pc, dst[1], t[0], t[1]);
2918 }
2919 if (mask & (1 << 3))
2920 emit_mov_immdval(pc, dst[3], 1.0f);
2921 }
2922 break;
2923 case TGSI_OPCODE_LRP:
2924 temp = temp_temp(pc, NULL);
2925 for (c = 0; c < 4; c++) {
2926 if (!(mask & (1 << c)))
2927 continue;
2928 emit_sub(pc, temp, src[1][c], src[2][c]);
2929 emit_mad(pc, dst[c], temp, src[0][c], src[2][c]);
2930 }
2931 break;
2932 case TGSI_OPCODE_MAD:
2933 for (c = 0; c < 4; c++) {
2934 if (!(mask & (1 << c)))
2935 continue;
2936 emit_mad(pc, dst[c], src[0][c], src[1][c], src[2][c]);
2937 }
2938 break;
2939 case TGSI_OPCODE_MAX:
2940 for (c = 0; c < 4; c++) {
2941 if (!(mask & (1 << c)))
2942 continue;
2943 emit_minmax(pc, 0x880, dst[c], src[0][c], src[1][c]);
2944 }
2945 break;
2946 case TGSI_OPCODE_MIN:
2947 for (c = 0; c < 4; c++) {
2948 if (!(mask & (1 << c)))
2949 continue;
2950 emit_minmax(pc, 0x8a0, dst[c], src[0][c], src[1][c]);
2951 }
2952 break;
2953 case TGSI_OPCODE_MOV:
2954 for (c = 0; c < 4; c++) {
2955 if (!(mask & (1 << c)))
2956 continue;
2957 emit_mov(pc, dst[c], src[0][c]);
2958 }
2959 break;
2960 case TGSI_OPCODE_MUL:
2961 for (c = 0; c < 4; c++) {
2962 if (!(mask & (1 << c)))
2963 continue;
2964 emit_mul(pc, dst[c], src[0][c], src[1][c]);
2965 }
2966 break;
2967 case TGSI_OPCODE_NOT:
2968 for (c = 0; c < 4; c++) {
2969 if (!(mask & (1 << c)))
2970 continue;
2971 emit_not(pc, dst[c], src[0][c]);
2972 }
2973 break;
2974 case TGSI_OPCODE_POW:
2975 emit_pow(pc, brdc, src[0][0], src[1][0]);
2976 break;
2977 case TGSI_OPCODE_RCP:
2978 if (!sat && popcnt4(mask) == 1)
2979 brdc = dst[ffs(mask) - 1];
2980 emit_flop(pc, NV50_FLOP_RCP, brdc, src[0][0]);
2981 break;
2982 case TGSI_OPCODE_RET:
2983 if (pc->p->type == PIPE_SHADER_FRAGMENT && !pc->in_subroutine)
2984 nv50_fp_move_results(pc);
2985 emit_ret(pc, -1, 0);
2986 break;
2987 case TGSI_OPCODE_RSQ:
2988 if (!sat && popcnt4(mask) == 1)
2989 brdc = dst[ffs(mask) - 1];
2990 src[0][0]->mod |= NV50_MOD_ABS;
2991 emit_flop(pc, NV50_FLOP_RSQ, brdc, src[0][0]);
2992 break;
2993 case TGSI_OPCODE_SAD:
2994 for (c = 0; c < 4; c++) {
2995 if (!(mask & (1 << c)))
2996 continue;
2997 emit_sad(pc, dst[c], src[0][c], src[1][c], src[2][c]);
2998 }
2999 break;
3000 case TGSI_OPCODE_SCS:
3001 temp = temp_temp(pc, NULL);
3002 if (mask & 3)
3003 emit_precossin(pc, temp, src[0][0]);
3004 if (mask & (1 << 0))
3005 emit_flop(pc, NV50_FLOP_COS, dst[0], temp);
3006 if (mask & (1 << 1))
3007 emit_flop(pc, NV50_FLOP_SIN, dst[1], temp);
3008 if (mask & (1 << 2))
3009 emit_mov_immdval(pc, dst[2], 0.0);
3010 if (mask & (1 << 3))
3011 emit_mov_immdval(pc, dst[3], 1.0);
3012 break;
3013 case TGSI_OPCODE_SHL:
3014 case TGSI_OPCODE_ISHR:
3015 case TGSI_OPCODE_USHR:
3016 for (c = 0; c < 4; c++) {
3017 if (!(mask & (1 << c)))
3018 continue;
3019 emit_shift(pc, dst[c], src[0][c], src[1][c],
3020 inst->Instruction.Opcode);
3021 }
3022 break;
3023 case TGSI_OPCODE_SIN:
3024 if (mask & 8) {
3025 emit_precossin(pc, temp, src[0][3]);
3026 emit_flop(pc, NV50_FLOP_SIN, dst[3], temp);
3027 if (!(mask &= 7))
3028 break;
3029 if (temp == dst[3])
3030 temp = brdc = temp_temp(pc, NULL);
3031 }
3032 emit_precossin(pc, temp, src[0][0]);
3033 emit_flop(pc, NV50_FLOP_SIN, brdc, temp);
3034 break;
3035 case TGSI_OPCODE_SLT:
3036 case TGSI_OPCODE_SGE:
3037 case TGSI_OPCODE_SEQ:
3038 case TGSI_OPCODE_SGT:
3039 case TGSI_OPCODE_SLE:
3040 case TGSI_OPCODE_SNE:
3041 case TGSI_OPCODE_ISLT:
3042 case TGSI_OPCODE_ISGE:
3043 case TGSI_OPCODE_USEQ:
3044 case TGSI_OPCODE_USGE:
3045 case TGSI_OPCODE_USLT:
3046 case TGSI_OPCODE_USNE:
3047 {
3048 uint8_t cc, ty;
3049
3050 map_tgsi_setop_hw(inst->Instruction.Opcode, &cc, &ty);
3051
3052 for (c = 0; c < 4; c++) {
3053 if (!(mask & (1 << c)))
3054 continue;
3055 emit_set(pc, cc, dst[c], -1, src[0][c], src[1][c], ty);
3056 }
3057 }
3058 break;
3059 case TGSI_OPCODE_SUB:
3060 for (c = 0; c < 4; c++) {
3061 if (!(mask & (1 << c)))
3062 continue;
3063 emit_sub(pc, dst[c], src[0][c], src[1][c]);
3064 }
3065 break;
3066 case TGSI_OPCODE_TEX:
3067 emit_tex(pc, dst, mask, src[0], unit,
3068 inst->Texture.Texture, FALSE, 0);
3069 break;
3070 case TGSI_OPCODE_TXB:
3071 emit_tex(pc, dst, mask, src[0], unit,
3072 inst->Texture.Texture, FALSE, -1);
3073 break;
3074 case TGSI_OPCODE_TXL:
3075 emit_tex(pc, dst, mask, src[0], unit,
3076 inst->Texture.Texture, FALSE, 1);
3077 break;
3078 case TGSI_OPCODE_TXP:
3079 emit_tex(pc, dst, mask, src[0], unit,
3080 inst->Texture.Texture, TRUE, 0);
3081 break;
3082 case TGSI_OPCODE_TRUNC:
3083 for (c = 0; c < 4; c++) {
3084 if (!(mask & (1 << c)))
3085 continue;
3086 emit_cvt(pc, dst[c], src[0][c], -1,
3087 CVT_TRUNC | CVT_F32_F32 | CVT_RI);
3088 }
3089 break;
3090 case TGSI_OPCODE_U2F:
3091 for (c = 0; c < 4; c++) {
3092 if (!(mask & (1 << c)))
3093 continue;
3094 emit_cvt(pc, dst[c], src[0][c], -1, CVT_F32_U32);
3095 }
3096 break;
3097 case TGSI_OPCODE_UADD:
3098 for (c = 0; c < 4; c++) {
3099 if (!(mask & (1 << c)))
3100 continue;
3101 emit_add_b32(pc, dst[c], src[0][c], src[1][c]);
3102 }
3103 break;
3104 case TGSI_OPCODE_UMAX:
3105 for (c = 0; c < 4; c++) {
3106 if (!(mask & (1 << c)))
3107 continue;
3108 emit_minmax(pc, 0x084, dst[c], src[0][c], src[1][c]);
3109 }
3110 break;
3111 case TGSI_OPCODE_UMIN:
3112 for (c = 0; c < 4; c++) {
3113 if (!(mask & (1 << c)))
3114 continue;
3115 emit_minmax(pc, 0x0a4, dst[c], src[0][c], src[1][c]);
3116 }
3117 break;
3118 case TGSI_OPCODE_UMAD:
3119 {
3120 assert(!temp);
3121 temp = temp_temp(pc, NULL);
3122 for (c = 0; c < 4; c++) {
3123 if (!(mask & (1 << c)))
3124 continue;
3125 emit_mul_u16(pc, temp, src[0][c], 0, src[1][c], 1);
3126 emit_mad_u16(pc, temp, src[0][c], 1, src[1][c], 0,
3127 temp);
3128 emit_shl_imm(pc, temp, temp, 16);
3129 emit_mad_u16(pc, temp, src[0][c], 0, src[1][c], 0,
3130 temp);
3131 emit_add_b32(pc, dst[c], temp, src[2][c]);
3132 }
3133 }
3134 break;
3135 case TGSI_OPCODE_UMUL:
3136 {
3137 assert(!temp);
3138 temp = temp_temp(pc, NULL);
3139 for (c = 0; c < 4; c++) {
3140 if (!(mask & (1 << c)))
3141 continue;
3142 emit_mul_u16(pc, temp, src[0][c], 0, src[1][c], 1);
3143 emit_mad_u16(pc, temp, src[0][c], 1, src[1][c], 0,
3144 temp);
3145 emit_shl_imm(pc, temp, temp, 16);
3146 emit_mad_u16(pc, dst[c], src[0][c], 0, src[1][c], 0,
3147 temp);
3148 }
3149 }
3150 break;
3151 case TGSI_OPCODE_XPD:
3152 temp = temp_temp(pc, NULL);
3153 if (mask & (1 << 0)) {
3154 emit_mul(pc, temp, src[0][2], src[1][1]);
3155 emit_msb(pc, dst[0], src[0][1], src[1][2], temp);
3156 }
3157 if (mask & (1 << 1)) {
3158 emit_mul(pc, temp, src[0][0], src[1][2]);
3159 emit_msb(pc, dst[1], src[0][2], src[1][0], temp);
3160 }
3161 if (mask & (1 << 2)) {
3162 emit_mul(pc, temp, src[0][1], src[1][0]);
3163 emit_msb(pc, dst[2], src[0][0], src[1][1], temp);
3164 }
3165 if (mask & (1 << 3))
3166 emit_mov_immdval(pc, dst[3], 1.0);
3167 break;
3168 case TGSI_OPCODE_END:
3169 if (pc->p->type == PIPE_SHADER_FRAGMENT)
3170 nv50_fp_move_results(pc);
3171
3172 if (!pc->p->exec_tail ||
3173 is_immd(pc->p->exec_tail) ||
3174 is_join(pc->p->exec_tail) ||
3175 is_control_flow(pc->p->exec_tail))
3176 emit_nop(pc);
3177
3178 /* last insn must be long so it can have the exit bit set */
3179 if (!is_long(pc->p->exec_tail))
3180 convert_to_long(pc, pc->p->exec_tail);
3181
3182 pc->p->exec_tail->inst[1] |= 1; /* set exit bit */
3183
3184 terminate_mbb(pc);
3185 break;
3186 default:
3187 NOUVEAU_ERR("invalid opcode %d\n", inst->Instruction.Opcode);
3188 return FALSE;
3189 }
3190
3191 if (brdc) {
3192 if (sat)
3193 emit_sat(pc, brdc, brdc);
3194 for (c = 0; c < 4; c++)
3195 if ((mask & (1 << c)) && dst[c] != brdc)
3196 emit_mov(pc, dst[c], brdc);
3197 } else
3198 if (sat) {
3199 for (c = 0; c < 4; c++) {
3200 if (!(mask & (1 << c)))
3201 continue;
3202 /* In this case we saturate later, and dst[c] won't
3203 * be another temp_temp (and thus lost), since rdst
3204 * already is TEMP (see above). */
3205 if (rdst[c]->type == P_TEMP && rdst[c]->index < 0)
3206 continue;
3207 emit_sat(pc, rdst[c], dst[c]);
3208 }
3209 }
3210
3211 kill_temp_temp(pc, NULL);
3212 pc->reg_instance_nr = 0;
3213
3214 return TRUE;
3215 }
3216
3217 static void
3218 prep_inspect_insn(struct nv50_pc *pc, const struct tgsi_full_instruction *insn)
3219 {
3220 struct nv50_reg *r, *reg = NULL;
3221 const struct tgsi_full_src_register *src;
3222 const struct tgsi_dst_register *dst;
3223 unsigned i, c, k, mask;
3224
3225 dst = &insn->Dst[0].Register;
3226 mask = dst->WriteMask;
3227
3228 if (dst->File == TGSI_FILE_TEMPORARY)
3229 reg = pc->temp;
3230 else
3231 if (dst->File == TGSI_FILE_OUTPUT) {
3232 reg = pc->result;
3233
3234 if (insn->Instruction.Opcode == TGSI_OPCODE_MOV &&
3235 dst->Index == pc->edgeflag_out &&
3236 insn->Src[0].Register.File == TGSI_FILE_INPUT)
3237 pc->p->cfg.edgeflag_in = insn->Src[0].Register.Index;
3238 }
3239
3240 if (reg) {
3241 for (c = 0; c < 4; c++) {
3242 if (!(mask & (1 << c)))
3243 continue;
3244 reg[dst->Index * 4 + c].acc = pc->insn_nr;
3245 }
3246 }
3247
3248 for (i = 0; i < insn->Instruction.NumSrcRegs; i++) {
3249 src = &insn->Src[i];
3250
3251 if (src->Register.File == TGSI_FILE_TEMPORARY)
3252 reg = pc->temp;
3253 else
3254 if (src->Register.File == TGSI_FILE_INPUT)
3255 reg = pc->attr;
3256 else
3257 continue;
3258
3259 mask = nv50_tgsi_src_mask(insn, i);
3260
3261 for (c = 0; c < 4; c++) {
3262 if (!(mask & (1 << c)))
3263 continue;
3264 k = tgsi_util_get_full_src_register_swizzle(src, c);
3265
3266 r = &reg[src->Register.Index * 4 + k];
3267
3268 /* If used before written, pre-allocate the reg,
3269 * lest we overwrite results from a subroutine.
3270 */
3271 if (!r->acc && r->type == P_TEMP)
3272 alloc_reg(pc, r);
3273
3274 r->acc = pc->insn_nr;
3275 }
3276 }
3277 }
3278
3279 /* Returns a bitmask indicating which dst components need to be
3280 * written to temporaries first to avoid 'corrupting' sources.
3281 *
3282 * m[i] (out) indicate component to write in the i-th position
3283 * rdep[c] (in) bitmasks of dst[i] that require dst[c] as source
3284 */
3285 static unsigned
3286 nv50_revdep_reorder(unsigned m[4], unsigned rdep[4])
3287 {
3288 unsigned i, c, x, unsafe = 0;
3289
3290 for (c = 0; c < 4; c++)
3291 m[c] = c;
3292
3293 /* Swap as long as a dst component written earlier is depended on
3294 * by one written later, but the next one isn't depended on by it.
3295 */
3296 for (c = 0; c < 3; c++) {
3297 if (rdep[m[c + 1]] & (1 << m[c]))
3298 continue; /* if next one is depended on by us */
3299 for (i = c + 1; i < 4; i++)
3300 /* if we are depended on by a later one */
3301 if (rdep[m[c]] & (1 << m[i]))
3302 break;
3303 if (i == 4)
3304 continue;
3305 /* now, swap */
3306 x = m[c];
3307 m[c] = m[c + 1];
3308 m[c + 1] = x;
3309
3310 /* restart */
3311 c = 0;
3312 }
3313
3314 /* mark dependencies that could not be resolved by reordering */
3315 for (i = 0; i < 3; ++i)
3316 for (c = i + 1; c < 4; ++c)
3317 if (rdep[m[i]] & (1 << m[c]))
3318 unsafe |= (1 << i);
3319
3320 /* NOTE: $unsafe is with respect to order, not component */
3321 return unsafe;
3322 }
3323
3324 /* Select a suitable dst register for broadcasting scalar results,
3325 * or return NULL if we have to allocate an extra TEMP.
3326 *
3327 * If e.g. only 1 component is written, we may also emit the final
3328 * result to a write-only register.
3329 */
3330 static struct nv50_reg *
3331 tgsi_broadcast_dst(struct nv50_pc *pc,
3332 const struct tgsi_full_dst_register *fd, unsigned mask)
3333 {
3334 if (fd->Register.File == TGSI_FILE_TEMPORARY) {
3335 int c = ffs(~mask & fd->Register.WriteMask);
3336 if (c)
3337 return tgsi_dst(pc, c - 1, fd);
3338 } else {
3339 int c = ffs(fd->Register.WriteMask) - 1;
3340 if ((1 << c) == fd->Register.WriteMask)
3341 return tgsi_dst(pc, c, fd);
3342 }
3343
3344 return NULL;
3345 }
3346
3347 /* Scan source swizzles and return a bitmask indicating dst regs that
3348 * also occur among the src regs, and fill rdep for nv50_revdep_reoder.
3349 */
3350 static unsigned
3351 nv50_tgsi_scan_swizzle(const struct tgsi_full_instruction *insn,
3352 unsigned rdep[4])
3353 {
3354 const struct tgsi_full_dst_register *fd = &insn->Dst[0];
3355 const struct tgsi_full_src_register *fs;
3356 unsigned i, deqs = 0;
3357
3358 for (i = 0; i < 4; ++i)
3359 rdep[i] = 0;
3360
3361 for (i = 0; i < insn->Instruction.NumSrcRegs; i++) {
3362 unsigned chn, mask = nv50_tgsi_src_mask(insn, i);
3363 int ms = get_supported_mods(insn, i);
3364
3365 fs = &insn->Src[i];
3366 if (fs->Register.File != fd->Register.File ||
3367 fs->Register.Index != fd->Register.Index)
3368 continue;
3369
3370 for (chn = 0; chn < 4; ++chn) {
3371 unsigned s, c;
3372
3373 if (!(mask & (1 << chn))) /* src is not read */
3374 continue;
3375 c = tgsi_util_get_full_src_register_swizzle(fs, chn);
3376 s = tgsi_util_get_full_src_register_sign_mode(fs, chn);
3377
3378 if (!(fd->Register.WriteMask & (1 << c)))
3379 continue;
3380
3381 if (s == TGSI_UTIL_SIGN_TOGGLE && !(ms & NV50_MOD_NEG))
3382 continue;
3383 if (s == TGSI_UTIL_SIGN_CLEAR && !(ms & NV50_MOD_ABS))
3384 continue;
3385 if ((s == TGSI_UTIL_SIGN_SET) && ((ms & 3) != 3))
3386 continue;
3387
3388 rdep[c] |= nv50_tgsi_dst_revdep(
3389 insn->Instruction.Opcode, i, chn);
3390 deqs |= (1 << c);
3391 }
3392 }
3393
3394 return deqs;
3395 }
3396
3397 static boolean
3398 nv50_tgsi_insn(struct nv50_pc *pc, const union tgsi_full_token *tok)
3399 {
3400 struct tgsi_full_instruction insn = tok->FullInstruction;
3401 const struct tgsi_full_dst_register *fd;
3402 unsigned i, deqs, rdep[4], m[4];
3403
3404 fd = &tok->FullInstruction.Dst[0];
3405 deqs = nv50_tgsi_scan_swizzle(&insn, rdep);
3406
3407 if (is_scalar_op(insn.Instruction.Opcode)) {
3408 pc->r_brdc = tgsi_broadcast_dst(pc, fd, deqs);
3409 if (!pc->r_brdc)
3410 pc->r_brdc = temp_temp(pc, NULL);
3411 return nv50_program_tx_insn(pc, &insn);
3412 }
3413 pc->r_brdc = NULL;
3414
3415 if (!deqs || (!rdep[0] && !rdep[1] && !rdep[2] && !rdep[3]))
3416 return nv50_program_tx_insn(pc, &insn);
3417
3418 deqs = nv50_revdep_reorder(m, rdep);
3419
3420 for (i = 0; i < 4; ++i) {
3421 assert(pc->r_dst[m[i]] == NULL);
3422
3423 insn.Dst[0].Register.WriteMask =
3424 fd->Register.WriteMask & (1 << m[i]);
3425
3426 if (!insn.Dst[0].Register.WriteMask)
3427 continue;
3428
3429 if (deqs & (1 << i))
3430 pc->r_dst[m[i]] = alloc_temp(pc, NULL);
3431
3432 if (!nv50_program_tx_insn(pc, &insn))
3433 return FALSE;
3434 }
3435
3436 for (i = 0; i < 4; i++) {
3437 struct nv50_reg *reg = pc->r_dst[i];
3438 if (!reg)
3439 continue;
3440 pc->r_dst[i] = NULL;
3441
3442 if (insn.Instruction.Saturate == TGSI_SAT_ZERO_ONE)
3443 emit_sat(pc, tgsi_dst(pc, i, fd), reg);
3444 else
3445 emit_mov(pc, tgsi_dst(pc, i, fd), reg);
3446 free_temp(pc, reg);
3447 }
3448
3449 return TRUE;
3450 }
3451
3452 static void
3453 load_interpolant(struct nv50_pc *pc, struct nv50_reg *reg)
3454 {
3455 struct nv50_reg *iv, **ppiv;
3456 unsigned mode = pc->interp_mode[reg->index];
3457
3458 ppiv = (mode & INTERP_CENTROID) ? &pc->iv_c : &pc->iv_p;
3459 iv = *ppiv;
3460
3461 if ((mode & INTERP_PERSPECTIVE) && !iv) {
3462 iv = *ppiv = alloc_temp(pc, NULL);
3463 iv->rhw = popcnt4(pc->p->cfg.regs[1] >> 24) - 1;
3464
3465 emit_interp(pc, iv, NULL, mode & INTERP_CENTROID);
3466 emit_flop(pc, NV50_FLOP_RCP, iv, iv);
3467
3468 /* XXX: when loading interpolants dynamically, move these
3469 * to the program head, or make sure it can't be skipped.
3470 */
3471 }
3472
3473 emit_interp(pc, reg, iv, mode);
3474 }
3475
3476 /* The face input is always at v[255] (varying space), with a
3477 * value of 0 for back-facing, and 0xffffffff for front-facing.
3478 */
3479 static void
3480 load_frontfacing(struct nv50_pc *pc, struct nv50_reg *sv)
3481 {
3482 struct nv50_reg *temp = alloc_temp(pc, NULL);
3483 int r_pred = 0;
3484
3485 temp->rhw = 255;
3486 emit_interp(pc, temp, NULL, INTERP_FLAT);
3487
3488 emit_cvt(pc, sv, temp, r_pred, CVT_ABS | CVT_F32_S32);
3489
3490 emit_not(pc, temp, temp);
3491 set_pred(pc, 0x2, r_pred, pc->p->exec_tail);
3492 emit_cvt(pc, sv, temp, -1, CVT_F32_S32);
3493 set_pred(pc, 0x2, r_pred, pc->p->exec_tail);
3494
3495 free_temp(pc, temp);
3496 }
3497
3498 static void
3499 load_instance_id(struct nv50_pc *pc, unsigned index)
3500 {
3501 struct nv50_reg reg, mem;
3502
3503 ctor_reg(&reg, P_TEMP, -1, -1);
3504 ctor_reg(&mem, P_CONST, -1, 24); /* startInstance */
3505 mem.buf_index = 2;
3506
3507 emit_add_b32(pc, &reg, &pc->sysval[index], &mem);
3508 pc->sysval[index] = reg;
3509 }
3510
3511 static void
3512 copy_semantic_info(struct nv50_program *p)
3513 {
3514 unsigned i, id;
3515
3516 for (i = 0; i < p->cfg.in_nr; ++i) {
3517 id = p->cfg.in[i].id;
3518 p->cfg.in[i].sn = p->info.input_semantic_name[id];
3519 p->cfg.in[i].si = p->info.input_semantic_index[id];
3520 }
3521
3522 for (i = 0; i < p->cfg.out_nr; ++i) {
3523 id = p->cfg.out[i].id;
3524 p->cfg.out[i].sn = p->info.output_semantic_name[id];
3525 p->cfg.out[i].si = p->info.output_semantic_index[id];
3526 }
3527 }
3528
3529 static boolean
3530 nv50_program_tx_prep(struct nv50_pc *pc)
3531 {
3532 struct tgsi_parse_context tp;
3533 struct nv50_program *p = pc->p;
3534 boolean ret = FALSE;
3535 unsigned i, c, instance_id = 0, vertex_id = 0, flat_nr = 0;
3536
3537 tgsi_parse_init(&tp, pc->p->pipe.tokens);
3538 while (!tgsi_parse_end_of_tokens(&tp)) {
3539 const union tgsi_full_token *tok = &tp.FullToken;
3540
3541 tgsi_parse_token(&tp);
3542 switch (tok->Token.Type) {
3543 case TGSI_TOKEN_TYPE_IMMEDIATE:
3544 {
3545 const struct tgsi_full_immediate *imm =
3546 &tp.FullToken.FullImmediate;
3547
3548 ctor_immd_4f32(pc, imm->u[0].Float,
3549 imm->u[1].Float,
3550 imm->u[2].Float,
3551 imm->u[3].Float);
3552 }
3553 break;
3554 case TGSI_TOKEN_TYPE_DECLARATION:
3555 {
3556 const struct tgsi_full_declaration *d;
3557 unsigned si, last, first, mode;
3558
3559 d = &tp.FullToken.FullDeclaration;
3560 first = d->Range.First;
3561 last = d->Range.Last;
3562
3563 switch (d->Declaration.File) {
3564 case TGSI_FILE_TEMPORARY:
3565 break;
3566 case TGSI_FILE_OUTPUT:
3567 if (!d->Declaration.Semantic ||
3568 p->type == PIPE_SHADER_FRAGMENT)
3569 break;
3570
3571 si = d->Semantic.Index;
3572 switch (d->Semantic.Name) {
3573 case TGSI_SEMANTIC_BCOLOR:
3574 p->cfg.two_side[si].hw = first;
3575 if (p->cfg.out_nr > first)
3576 p->cfg.out_nr = first;
3577 break;
3578 case TGSI_SEMANTIC_PSIZE:
3579 p->cfg.psiz = first;
3580 if (p->cfg.out_nr > first)
3581 p->cfg.out_nr = first;
3582 break;
3583 case TGSI_SEMANTIC_EDGEFLAG:
3584 pc->edgeflag_out = first;
3585 break;
3586 /*
3587 case TGSI_SEMANTIC_CLIP_DISTANCE:
3588 p->cfg.clpd = MIN2(p->cfg.clpd, first);
3589 break;
3590 */
3591 default:
3592 break;
3593 }
3594 break;
3595 case TGSI_FILE_INPUT:
3596 {
3597 if (p->type != PIPE_SHADER_FRAGMENT)
3598 break;
3599
3600 switch (d->Declaration.Interpolate) {
3601 case TGSI_INTERPOLATE_CONSTANT:
3602 mode = INTERP_FLAT;
3603 flat_nr++;
3604 break;
3605 case TGSI_INTERPOLATE_PERSPECTIVE:
3606 mode = INTERP_PERSPECTIVE;
3607 p->cfg.regs[1] |= 0x08 << 24;
3608 break;
3609 default:
3610 mode = INTERP_LINEAR;
3611 break;
3612 }
3613 if (d->Declaration.Centroid)
3614 mode |= INTERP_CENTROID;
3615
3616 assert(last < 32);
3617 for (i = first; i <= last; i++)
3618 pc->interp_mode[i] = mode;
3619 }
3620 break;
3621 case TGSI_FILE_SYSTEM_VALUE:
3622 assert(d->Declaration.Semantic);
3623 switch (d->Semantic.Name) {
3624 case TGSI_SEMANTIC_FACE:
3625 assert(p->type == PIPE_SHADER_FRAGMENT);
3626 load_frontfacing(pc,
3627 &pc->sysval[first]);
3628 break;
3629 case TGSI_SEMANTIC_INSTANCEID:
3630 assert(p->type == PIPE_SHADER_VERTEX);
3631 instance_id = first;
3632 p->cfg.regs[0] |= (1 << 4);
3633 break;
3634 case TGSI_SEMANTIC_PRIMID:
3635 assert(p->type != PIPE_SHADER_VERTEX);
3636 p->cfg.prim_id = first;
3637 break;
3638 /*
3639 case TGSI_SEMANTIC_PRIMIDIN:
3640 assert(p->type == PIPE_SHADER_GEOMETRY);
3641 pc->sysval[first].hw = 6;
3642 p->cfg.regs[0] |= (1 << 8);
3643 break;
3644 case TGSI_SEMANTIC_VERTEXID:
3645 assert(p->type == PIPE_SHADER_VERTEX);
3646 vertex_id = first;
3647 p->cfg.regs[0] |= (1 << 12) | (1 << 0);
3648 break;
3649 */
3650 }
3651 break;
3652 case TGSI_FILE_ADDRESS:
3653 case TGSI_FILE_CONSTANT:
3654 case TGSI_FILE_SAMPLER:
3655 break;
3656 default:
3657 NOUVEAU_ERR("bad decl file %d\n",
3658 d->Declaration.File);
3659 goto out_err;
3660 }
3661 }
3662 break;
3663 case TGSI_TOKEN_TYPE_INSTRUCTION:
3664 pc->insn_nr++;
3665 prep_inspect_insn(pc, &tok->FullInstruction);
3666 break;
3667 default:
3668 break;
3669 }
3670 }
3671
3672 if (p->type == PIPE_SHADER_VERTEX || p->type == PIPE_SHADER_GEOMETRY) {
3673 int rid = 0;
3674
3675 if (p->type == PIPE_SHADER_GEOMETRY) {
3676 for (i = 0; i < pc->attr_nr; ++i) {
3677 p->cfg.in[i].hw = rid;
3678 p->cfg.in[i].id = i;
3679
3680 for (c = 0; c < 4; ++c) {
3681 int n = i * 4 + c;
3682 if (!pc->attr[n].acc)
3683 continue;
3684 pc->attr[n].hw = rid++;
3685 p->cfg.in[i].mask |= 1 << c;
3686 }
3687 }
3688 } else {
3689 for (i = 0; i < pc->attr_nr * 4; ++i) {
3690 if (pc->attr[i].acc) {
3691 pc->attr[i].hw = rid++;
3692 p->cfg.attr[i / 32] |= 1 << (i % 32);
3693 }
3694 }
3695 if (p->cfg.regs[0] & (1 << 0))
3696 pc->sysval[vertex_id].hw = rid++;
3697 if (p->cfg.regs[0] & (1 << 4)) {
3698 pc->sysval[instance_id].hw = rid++;
3699 load_instance_id(pc, instance_id);
3700 }
3701 }
3702
3703 for (i = 0, rid = 0; i < pc->result_nr; ++i) {
3704 p->cfg.out[i].hw = rid;
3705 p->cfg.out[i].id = i;
3706
3707 for (c = 0; c < 4; ++c) {
3708 int n = i * 4 + c;
3709 if (!pc->result[n].acc)
3710 continue;
3711 pc->result[n].hw = rid++;
3712 p->cfg.out[i].mask |= 1 << c;
3713 }
3714 }
3715 if (p->cfg.prim_id < 0x40) {
3716 /* GP has to write to PrimitiveID */
3717 ctor_reg(&pc->sysval[p->cfg.prim_id],
3718 P_RESULT, p->cfg.prim_id, rid);
3719 p->cfg.prim_id = rid++;
3720 }
3721
3722 for (c = 0; c < 2; ++c)
3723 if (p->cfg.two_side[c].hw < 0x40)
3724 p->cfg.two_side[c] = p->cfg.out[
3725 p->cfg.two_side[c].hw];
3726
3727 if (p->cfg.psiz < 0x40)
3728 p->cfg.psiz = p->cfg.out[p->cfg.psiz].hw;
3729
3730 copy_semantic_info(p);
3731 } else
3732 if (p->type == PIPE_SHADER_FRAGMENT) {
3733 int rid = 0, aid;
3734 unsigned n = 0, m = pc->attr_nr - flat_nr;
3735
3736 pc->allow32 = TRUE;
3737
3738 /* do we read FragCoord ? */
3739 if (pc->attr_nr &&
3740 p->info.input_semantic_name[0] == TGSI_SEMANTIC_POSITION) {
3741 /* select FCRD components we want accessible */
3742 for (c = 0; c < 4; ++c)
3743 if (pc->attr[c].acc)
3744 p->cfg.regs[1] |= 1 << (24 + c);
3745 aid = 0;
3746 } else /* offset by 1 if FCRD.w is needed for pinterp */
3747 aid = popcnt4(p->cfg.regs[1] >> 24);
3748
3749 /* non-flat interpolants have to be mapped to
3750 * the lower hardware IDs, so sort them:
3751 */
3752 for (i = 0; i < pc->attr_nr; i++) {
3753 if (pc->interp_mode[i] == INTERP_FLAT)
3754 p->cfg.in[m++].id = i;
3755 else {
3756 if (!(pc->interp_mode[i] & INTERP_PERSPECTIVE))
3757 p->cfg.in[n].linear = TRUE;
3758 p->cfg.in[n++].id = i;
3759 }
3760 }
3761 copy_semantic_info(p);
3762
3763 for (n = 0; n < pc->attr_nr; ++n) {
3764 p->cfg.in[n].hw = rid = aid;
3765 i = p->cfg.in[n].id;
3766
3767 if (p->info.input_semantic_name[i] ==
3768 TGSI_SEMANTIC_FACE) {
3769 load_frontfacing(pc, &pc->attr[i * 4]);
3770 continue;
3771 }
3772
3773 for (c = 0; c < 4; ++c) {
3774 if (!pc->attr[i * 4 + c].acc)
3775 continue;
3776 pc->attr[i * 4 + c].rhw = rid++;
3777 p->cfg.in[n].mask |= 1 << c;
3778
3779 load_interpolant(pc, &pc->attr[i * 4 + c]);
3780 }
3781 aid += popcnt4(p->cfg.in[n].mask);
3782 }
3783
3784 m = popcnt4(p->cfg.regs[1] >> 24);
3785
3786 /* set count of non-position inputs and of non-flat
3787 * non-position inputs for FP_INTERPOLANT_CTRL
3788 */
3789 p->cfg.regs[1] |= aid - m;
3790
3791 if (flat_nr) {
3792 i = p->cfg.in[pc->attr_nr - flat_nr].hw;
3793 p->cfg.regs[1] |= (i - m) << 16;
3794 } else
3795 p->cfg.regs[1] |= p->cfg.regs[1] << 16;
3796
3797 /* mark color semantic for light-twoside */
3798 n = 0x80;
3799 for (i = 0; i < p->cfg.in_nr; i++) {
3800 if (p->cfg.in[i].sn == TGSI_SEMANTIC_COLOR) {
3801 n = MIN2(n, p->cfg.in[i].hw - m);
3802 p->cfg.two_side[p->cfg.in[i].si] = p->cfg.in[i];
3803
3804 p->cfg.regs[0] += /* increase colour count */
3805 popcnt4(p->cfg.in[i].mask) << 16;
3806 }
3807 }
3808 if (n < 0x80)
3809 p->cfg.regs[0] += n;
3810
3811 if (p->cfg.prim_id < 0x40) {
3812 pc->sysval[p->cfg.prim_id].rhw = rid++;
3813 emit_interp(pc, &pc->sysval[p->cfg.prim_id], NULL,
3814 INTERP_FLAT);
3815 /* increase FP_INTERPOLANT_CTRL_COUNT */
3816 p->cfg.regs[1] += 1;
3817 }
3818
3819 /* Initialize FP results:
3820 * FragDepth is always first TGSI and last hw output
3821 */
3822 i = p->info.writes_z ? 4 : 0;
3823 for (rid = 0; i < pc->result_nr * 4; i++)
3824 pc->result[i].rhw = rid++;
3825 if (p->info.writes_z)
3826 pc->result[2].rhw = rid++;
3827
3828 p->cfg.high_result = rid;
3829
3830 /* separate/different colour results for MRTs ? */
3831 if (pc->result_nr - (p->info.writes_z ? 1 : 0) > 1)
3832 p->cfg.regs[2] |= 1;
3833 }
3834
3835 if (pc->immd_nr) {
3836 int rid = 0;
3837
3838 pc->immd = MALLOC(pc->immd_nr * 4 * sizeof(struct nv50_reg));
3839 if (!pc->immd)
3840 goto out_err;
3841
3842 for (i = 0; i < pc->immd_nr; i++) {
3843 for (c = 0; c < 4; c++, rid++)
3844 ctor_reg(&pc->immd[rid], P_IMMD, i, rid);
3845 }
3846 }
3847
3848 ret = TRUE;
3849 out_err:
3850 if (pc->iv_p)
3851 free_temp(pc, pc->iv_p);
3852 if (pc->iv_c)
3853 free_temp(pc, pc->iv_c);
3854
3855 tgsi_parse_free(&tp);
3856 return ret;
3857 }
3858
3859 static void
3860 free_nv50_pc(struct nv50_pc *pc)
3861 {
3862 if (pc->immd)
3863 FREE(pc->immd);
3864 if (pc->param)
3865 FREE(pc->param);
3866 if (pc->result)
3867 FREE(pc->result);
3868 if (pc->attr)
3869 FREE(pc->attr);
3870 if (pc->temp)
3871 FREE(pc->temp);
3872 if (pc->sysval)
3873 FREE(pc->sysval);
3874 if (pc->insn_pos)
3875 FREE(pc->insn_pos);
3876
3877 FREE(pc);
3878 }
3879
3880 static INLINE uint32_t
3881 nv50_map_gs_output_prim(unsigned pprim)
3882 {
3883 switch (pprim) {
3884 case PIPE_PRIM_POINTS:
3885 return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_POINTS;
3886 case PIPE_PRIM_LINE_STRIP:
3887 return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_LINE_STRIP;
3888 case PIPE_PRIM_TRIANGLE_STRIP:
3889 return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_TRIANGLE_STRIP;
3890 default:
3891 NOUVEAU_ERR("invalid GS_OUTPUT_PRIMITIVE: %u\n", pprim);
3892 abort();
3893 return 0;
3894 }
3895 }
3896
3897 static boolean
3898 ctor_nv50_pc(struct nv50_pc *pc, struct nv50_program *p)
3899 {
3900 int i, c;
3901 unsigned rtype[2] = { P_ATTR, P_RESULT };
3902
3903 pc->p = p;
3904 pc->temp_nr = p->info.file_max[TGSI_FILE_TEMPORARY] + 1;
3905 pc->attr_nr = p->info.file_max[TGSI_FILE_INPUT] + 1;
3906 pc->result_nr = p->info.file_max[TGSI_FILE_OUTPUT] + 1;
3907 pc->param_nr = p->info.file_max[TGSI_FILE_CONSTANT] + 1;
3908 pc->addr_nr = p->info.file_max[TGSI_FILE_ADDRESS] + 1;
3909 assert(pc->addr_nr <= 2);
3910 pc->sysval_nr = p->info.file_max[TGSI_FILE_SYSTEM_VALUE] + 1;
3911
3912 p->cfg.high_temp = 4;
3913
3914 p->cfg.two_side[0].hw = 0x40;
3915 p->cfg.two_side[1].hw = 0x40;
3916 p->cfg.prim_id = 0x40;
3917
3918 p->cfg.edgeflag_in = pc->edgeflag_out = 0xff;
3919
3920 for (i = 0; i < p->info.num_properties; ++i) {
3921 unsigned *data = &p->info.properties[i].data[0];
3922
3923 switch (p->info.properties[i].name) {
3924 case TGSI_PROPERTY_GS_OUTPUT_PRIM:
3925 p->cfg.prim_type = nv50_map_gs_output_prim(data[0]);
3926 break;
3927 case TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES:
3928 p->cfg.vert_count = data[0];
3929 break;
3930 default:
3931 break;
3932 }
3933 }
3934
3935 switch (p->type) {
3936 case PIPE_SHADER_VERTEX:
3937 p->cfg.psiz = 0x40;
3938 p->cfg.clpd = 0x40;
3939 p->cfg.out_nr = pc->result_nr;
3940 break;
3941 case PIPE_SHADER_GEOMETRY:
3942 assert(p->cfg.prim_type);
3943 assert(p->cfg.vert_count);
3944
3945 p->cfg.psiz = 0x80;
3946 p->cfg.clpd = 0x80;
3947 p->cfg.prim_id = 0x80;
3948 p->cfg.out_nr = pc->result_nr;
3949 p->cfg.in_nr = pc->attr_nr;
3950
3951 p->cfg.two_side[0].hw = 0x80;
3952 p->cfg.two_side[1].hw = 0x80;
3953 break;
3954 case PIPE_SHADER_FRAGMENT:
3955 rtype[0] = rtype[1] = P_TEMP;
3956
3957 p->cfg.regs[0] = 0x01000004;
3958 p->cfg.in_nr = pc->attr_nr;
3959
3960 if (p->info.writes_z) {
3961 p->cfg.regs[2] |= 0x00000100;
3962 p->cfg.regs[3] |= 0x00000011;
3963 }
3964 if (p->info.uses_kill)
3965 p->cfg.regs[2] |= 0x00100000;
3966 break;
3967 }
3968
3969 if (pc->temp_nr) {
3970 pc->temp = MALLOC(pc->temp_nr * 4 * sizeof(struct nv50_reg));
3971 if (!pc->temp)
3972 return FALSE;
3973
3974 for (i = 0; i < pc->temp_nr * 4; ++i)
3975 ctor_reg(&pc->temp[i], P_TEMP, i / 4, -1);
3976 }
3977
3978 if (pc->attr_nr) {
3979 pc->attr = MALLOC(pc->attr_nr * 4 * sizeof(struct nv50_reg));
3980 if (!pc->attr)
3981 return FALSE;
3982
3983 for (i = 0; i < pc->attr_nr * 4; ++i)
3984 ctor_reg(&pc->attr[i], rtype[0], i / 4, -1);
3985 }
3986
3987 if (pc->result_nr) {
3988 unsigned nr = pc->result_nr * 4;
3989
3990 pc->result = MALLOC(nr * sizeof(struct nv50_reg));
3991 if (!pc->result)
3992 return FALSE;
3993
3994 for (i = 0; i < nr; ++i)
3995 ctor_reg(&pc->result[i], rtype[1], i / 4, -1);
3996 }
3997
3998 if (pc->param_nr) {
3999 int rid = 0;
4000
4001 pc->param = MALLOC(pc->param_nr * 4 * sizeof(struct nv50_reg));
4002 if (!pc->param)
4003 return FALSE;
4004
4005 for (i = 0; i < pc->param_nr; ++i)
4006 for (c = 0; c < 4; ++c, ++rid)
4007 ctor_reg(&pc->param[rid], P_CONST, i, rid);
4008 }
4009
4010 if (pc->addr_nr) {
4011 pc->addr = CALLOC(pc->addr_nr * 4, sizeof(struct nv50_reg *));
4012 if (!pc->addr)
4013 return FALSE;
4014 }
4015 for (i = 0; i < NV50_SU_MAX_ADDR; ++i)
4016 ctor_reg(&pc->r_addr[i], P_ADDR, -1, i + 1);
4017
4018 if (pc->sysval_nr) {
4019 pc->sysval = CALLOC(pc->sysval_nr, sizeof(struct nv50_reg *));
4020 if (!pc->sysval)
4021 return FALSE;
4022 /* will only ever use SYSTEM_VALUE[i].x (hopefully) */
4023 for (i = 0; i < pc->sysval_nr; ++i)
4024 ctor_reg(&pc->sysval[i], rtype[0], i, -1);
4025 }
4026
4027 return TRUE;
4028 }
4029
4030 static void
4031 nv50_program_fixup_insns(struct nv50_pc *pc)
4032 {
4033 struct nv50_program_exec *e, **bra_list;
4034 unsigned i, n, pos;
4035
4036 bra_list = CALLOC(pc->p->exec_size, sizeof(struct nv50_program_exec *));
4037
4038 /* Collect branch instructions, we need to adjust their offsets
4039 * when converting 32 bit instructions to 64 bit ones
4040 */
4041 for (n = 0, e = pc->p->exec_head; e; e = e->next)
4042 if (e->param.index >= 0 && !e->param.mask)
4043 bra_list[n++] = e;
4044
4045 /* Make sure we don't have any single 32 bit instructions. */
4046 for (e = pc->p->exec_head, pos = 0; e; e = e->next) {
4047 pos += is_long(e) ? 2 : 1;
4048
4049 if ((pos & 1) && (!e->next || is_long(e->next))) {
4050 for (i = 0; i < n; ++i)
4051 if (bra_list[i]->param.index >= pos)
4052 bra_list[i]->param.index += 1;
4053 for (i = 0; i < pc->insn_nr; ++i)
4054 if (pc->insn_pos[i] >= pos)
4055 pc->insn_pos[i] += 1;
4056 convert_to_long(pc, e);
4057 ++pos;
4058 }
4059 }
4060
4061 FREE(bra_list);
4062
4063 if (!pc->p->info.opcode_count[TGSI_OPCODE_CAL])
4064 return;
4065
4066 /* fill in CALL offsets */
4067 for (e = pc->p->exec_head; e; e = e->next) {
4068 if ((e->inst[0] & 2) && (e->inst[0] >> 28) == 0x2)
4069 e->param.index = pc->insn_pos[e->param.index];
4070 }
4071 }
4072
4073 static boolean
4074 nv50_program_tx(struct nv50_program *p)
4075 {
4076 struct tgsi_parse_context parse;
4077 struct nv50_pc *pc;
4078 boolean ret;
4079
4080 pc = CALLOC_STRUCT(nv50_pc);
4081 if (!pc)
4082 return FALSE;
4083
4084 ret = ctor_nv50_pc(pc, p);
4085 if (ret == FALSE)
4086 goto out_cleanup;
4087
4088 ret = nv50_program_tx_prep(pc);
4089 if (ret == FALSE)
4090 goto out_cleanup;
4091
4092 pc->insn_pos = MALLOC(pc->insn_nr * sizeof(unsigned));
4093
4094 tgsi_parse_init(&parse, pc->p->pipe.tokens);
4095 while (!tgsi_parse_end_of_tokens(&parse)) {
4096 const union tgsi_full_token *tok = &parse.FullToken;
4097
4098 /* previously allow32 was FALSE for first & last instruction */
4099 pc->allow32 = TRUE;
4100
4101 tgsi_parse_token(&parse);
4102
4103 switch (tok->Token.Type) {
4104 case TGSI_TOKEN_TYPE_INSTRUCTION:
4105 pc->insn_pos[pc->insn_cur] = pc->p->exec_size;
4106 ++pc->insn_cur;
4107 ret = nv50_tgsi_insn(pc, tok);
4108 if (ret == FALSE)
4109 goto out_err;
4110 break;
4111 default:
4112 break;
4113 }
4114 }
4115
4116 nv50_program_fixup_insns(pc);
4117
4118 p->param_nr = pc->param_nr * 4;
4119 p->immd_nr = pc->immd_nr * 4;
4120 p->immd = pc->immd_buf;
4121
4122 out_err:
4123 tgsi_parse_free(&parse);
4124
4125 out_cleanup:
4126 free_nv50_pc(pc);
4127 return ret;
4128 }
4129
4130 static void
4131 nv50_program_validate(struct nv50_context *nv50, struct nv50_program *p)
4132 {
4133 if (nv50_program_tx(p) == FALSE)
4134 assert(0);
4135 p->translated = TRUE;
4136 }
4137
4138 static void
4139 nv50_program_upload_data(struct nv50_context *nv50, uint32_t *map,
4140 unsigned start, unsigned count, unsigned cbuf)
4141 {
4142 struct nouveau_channel *chan = nv50->screen->base.channel;
4143 struct nouveau_grobj *tesla = nv50->screen->tesla;
4144
4145 while (count) {
4146 unsigned nr = count > 2047 ? 2047 : count;
4147
4148 BEGIN_RING(chan, tesla, NV50TCL_CB_ADDR, 1);
4149 OUT_RING (chan, (cbuf << 0) | (start << 8));
4150 BEGIN_RING(chan, tesla, NV50TCL_CB_DATA(0) | 0x40000000, nr);
4151 OUT_RINGp (chan, map, nr);
4152
4153 map += nr;
4154 start += nr;
4155 count -= nr;
4156 }
4157 }
4158
4159 static void
4160 nv50_program_validate_data(struct nv50_context *nv50, struct nv50_program *p)
4161 {
4162 struct pipe_context *pipe = &nv50->pipe;
4163 struct pipe_transfer *transfer;
4164
4165 if (!p->data[0] && p->immd_nr) {
4166 struct nouveau_resource *heap = nv50->screen->immd_heap;
4167
4168 if (nouveau_resource_alloc(heap, p->immd_nr, p, &p->data[0])) {
4169 while (heap->next && heap->size < p->immd_nr) {
4170 struct nv50_program *evict = heap->next->priv;
4171 nouveau_resource_free(&evict->data[0]);
4172 }
4173
4174 if (nouveau_resource_alloc(heap, p->immd_nr, p,
4175 &p->data[0]))
4176 assert(0);
4177 }
4178
4179 /* immediates only need to be uploaded again when freed */
4180 nv50_program_upload_data(nv50, p->immd, p->data[0]->start,
4181 p->immd_nr, NV50_CB_PMISC);
4182 }
4183
4184 assert(p->param_nr <= 16384);
4185
4186 if (p->param_nr) {
4187 unsigned cb;
4188 uint32_t *map = pipe_buffer_map(pipe,
4189 nv50->constbuf[p->type],
4190 PIPE_TRANSFER_READ,
4191 &transfer);
4192 switch (p->type) {
4193 case PIPE_SHADER_GEOMETRY: cb = NV50_CB_PGP; break;
4194 case PIPE_SHADER_FRAGMENT: cb = NV50_CB_PFP; break;
4195 default:
4196 cb = NV50_CB_PVP;
4197 assert(p->type == PIPE_SHADER_VERTEX);
4198 break;
4199 }
4200
4201 nv50_program_upload_data(nv50, map, 0, p->param_nr, cb);
4202 pipe_buffer_unmap(pipe, nv50->constbuf[p->type],
4203 transfer);
4204 }
4205 }
4206
4207 static void
4208 nv50_program_validate_code(struct nv50_context *nv50, struct nv50_program *p)
4209 {
4210 struct nouveau_channel *chan = nv50->screen->base.channel;
4211 struct nouveau_grobj *tesla = nv50->screen->tesla;
4212 struct nv50_program_exec *e;
4213 uint32_t *up, i;
4214 boolean upload = FALSE;
4215 unsigned offset;
4216 int width;
4217
4218 if (!p->bo) {
4219 nouveau_bo_new(chan->device, NOUVEAU_BO_VRAM, 0x100,
4220 p->exec_size * 4, &p->bo);
4221 upload = TRUE;
4222 }
4223
4224 if (p->data[0] && p->data[0]->start != p->data_start[0])
4225 upload = TRUE;
4226
4227 if (!upload)
4228 return;
4229
4230 up = MALLOC(p->exec_size * 4);
4231
4232 for (i = 0, e = p->exec_head; e; e = e->next) {
4233 unsigned ei, ci, bs;
4234
4235 if (e->param.index >= 0 && e->param.mask) {
4236 bs = (e->inst[1] >> 22) & 0x07;
4237 assert(bs < 2);
4238 ei = e->param.shift >> 5;
4239 ci = e->param.index;
4240 if (bs == 0)
4241 ci += p->data[bs]->start;
4242
4243 e->inst[ei] &= ~e->param.mask;
4244 e->inst[ei] |= (ci << e->param.shift);
4245 } else
4246 if (e->param.index >= 0) {
4247 /* zero mask means param is a jump/branch offset */
4248 assert(!(e->param.index & 1));
4249 /* seem to be 8 byte steps */
4250 ei = (e->param.index >> 1) + 0 /* START_ID */;
4251
4252 e->inst[0] &= 0xf0000fff;
4253 e->inst[0] |= ei << 12;
4254 }
4255
4256 up[i++] = e->inst[0];
4257 if (is_long(e))
4258 up[i++] = e->inst[1];
4259 }
4260 assert(i == p->exec_size);
4261
4262 if (p->data[0])
4263 p->data_start[0] = p->data[0]->start;
4264
4265 #ifdef NV50_PROGRAM_DUMP
4266 NOUVEAU_ERR("-------\n");
4267 for (e = p->exec_head; e; e = e->next) {
4268 NOUVEAU_ERR("0x%08x\n", e->inst[0]);
4269 if (is_long(e))
4270 NOUVEAU_ERR("0x%08x\n", e->inst[1]);
4271 }
4272 #endif
4273
4274 /* SIFC_HEIGHT/SIFC_WIDTH of 65536 do not work, and are not reported
4275 * as data error either. hw bug ? */
4276 #define SIFC_MAX_WIDTH (65536 - 256)
4277 offset = 0;
4278 width = p->exec_size * 4;
4279 while (width > 0) {
4280 nv50_upload_sifc(nv50, p->bo, offset, NOUVEAU_BO_VRAM,
4281 NV50_2D_DST_FORMAT_R8_UNORM, 65536, 1, 262144,
4282 &up[offset / 4], NV50_2D_SIFC_FORMAT_R8_UNORM,
4283 0, 0, 0, MIN2(SIFC_MAX_WIDTH, width), 1, 1);
4284 width -= SIFC_MAX_WIDTH;
4285 offset += SIFC_MAX_WIDTH;
4286 }
4287 BEGIN_RING(chan, tesla, NV50TCL_CODE_CB_FLUSH, 1);
4288 OUT_RING (chan, 0);
4289
4290 FREE(up);
4291 }
4292
4293 struct nouveau_stateobj *
4294 nv50_vertprog_validate(struct nv50_context *nv50)
4295 {
4296 struct nouveau_grobj *tesla = nv50->screen->tesla;
4297 struct nv50_program *p = nv50->vertprog;
4298 struct nouveau_stateobj *so;
4299
4300 if (!p->translated) {
4301 nv50_program_validate(nv50, p);
4302 if (!p->translated)
4303 assert(0);
4304 }
4305
4306 nv50_program_validate_data(nv50, p);
4307 nv50_program_validate_code(nv50, p);
4308
4309 if (!(nv50->dirty & NV50_NEW_VERTPROG))
4310 return NULL;
4311
4312 so = so_new(5, 7, 2);
4313 so_method(so, tesla, NV50TCL_VP_ADDRESS_HIGH, 2);
4314 so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
4315 NOUVEAU_BO_HIGH, 0, 0);
4316 so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
4317 NOUVEAU_BO_LOW, 0, 0);
4318 so_method(so, tesla, NV50TCL_VP_ATTR_EN_0, 2);
4319 so_data (so, p->cfg.attr[0]);
4320 so_data (so, p->cfg.attr[1]);
4321 so_method(so, tesla, NV50TCL_VP_REG_ALLOC_RESULT, 1);
4322 so_data (so, p->cfg.high_result);
4323 so_method(so, tesla, NV50TCL_VP_REG_ALLOC_TEMP, 1);
4324 so_data (so, p->cfg.high_temp);
4325 so_method(so, tesla, NV50TCL_VP_START_ID, 1);
4326 so_data (so, 0); /* program start offset */
4327 return so;
4328 }
4329
4330 struct nouveau_stateobj *
4331 nv50_fragprog_validate(struct nv50_context *nv50)
4332 {
4333 struct nouveau_grobj *tesla = nv50->screen->tesla;
4334 struct nv50_program *p = nv50->fragprog;
4335 struct nouveau_stateobj *so;
4336
4337 if (!p->translated) {
4338 nv50_program_validate(nv50, p);
4339 if (!p->translated)
4340 assert(0);
4341 }
4342
4343 nv50_program_validate_data(nv50, p);
4344 nv50_program_validate_code(nv50, p);
4345
4346 if (!(nv50->dirty & NV50_NEW_FRAGPROG))
4347 return NULL;
4348
4349 so = so_new(6, 7, 2);
4350 so_method(so, tesla, NV50TCL_FP_ADDRESS_HIGH, 2);
4351 so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
4352 NOUVEAU_BO_HIGH, 0, 0);
4353 so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
4354 NOUVEAU_BO_LOW, 0, 0);
4355 so_method(so, tesla, NV50TCL_FP_REG_ALLOC_TEMP, 1);
4356 so_data (so, p->cfg.high_temp);
4357 so_method(so, tesla, NV50TCL_FP_RESULT_COUNT, 1);
4358 so_data (so, p->cfg.high_result);
4359 so_method(so, tesla, NV50TCL_FP_CONTROL, 1);
4360 so_data (so, p->cfg.regs[2]);
4361 so_method(so, tesla, NV50TCL_FP_CTRL_UNK196C, 1);
4362 so_data (so, p->cfg.regs[3]);
4363 so_method(so, tesla, NV50TCL_FP_START_ID, 1);
4364 so_data (so, 0); /* program start offset */
4365 return so;
4366 }
4367
4368 struct nouveau_stateobj *
4369 nv50_geomprog_validate(struct nv50_context *nv50)
4370 {
4371 struct nouveau_grobj *tesla = nv50->screen->tesla;
4372 struct nv50_program *p = nv50->geomprog;
4373 struct nouveau_stateobj *so;
4374
4375 if (!p->translated) {
4376 nv50_program_validate(nv50, p);
4377 if (!p->translated)
4378 assert(0);
4379 }
4380
4381 nv50_program_validate_data(nv50, p);
4382 nv50_program_validate_code(nv50, p);
4383
4384 if (!(nv50->dirty & NV50_NEW_GEOMPROG))
4385 return NULL;
4386
4387 so = so_new(6, 7, 2);
4388 so_method(so, tesla, NV50TCL_GP_ADDRESS_HIGH, 2);
4389 so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
4390 NOUVEAU_BO_HIGH, 0, 0);
4391 so_reloc (so, p->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
4392 NOUVEAU_BO_LOW, 0, 0);
4393 so_method(so, tesla, NV50TCL_GP_REG_ALLOC_TEMP, 1);
4394 so_data (so, p->cfg.high_temp);
4395 so_method(so, tesla, NV50TCL_GP_REG_ALLOC_RESULT, 1);
4396 so_data (so, p->cfg.high_result);
4397 so_method(so, tesla, NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE, 1);
4398 so_data (so, p->cfg.prim_type);
4399 so_method(so, tesla, NV50TCL_GP_VERTEX_OUTPUT_COUNT, 1);
4400 so_data (so, p->cfg.vert_count);
4401 so_method(so, tesla, NV50TCL_GP_START_ID, 1);
4402 so_data (so, 0);
4403 return so;
4404 }
4405
4406 static uint32_t
4407 nv50_pntc_replace(struct nv50_context *nv50, uint32_t pntc[8], unsigned base)
4408 {
4409 struct nv50_program *vp;
4410 struct nv50_program *fp = nv50->fragprog;
4411 unsigned i, c, m = base;
4412 uint32_t origin = 0x00000010;
4413
4414 vp = nv50->geomprog ? nv50->geomprog : nv50->vertprog;
4415
4416 /* XXX: this might not work correctly in all cases yet - we'll
4417 * just assume that an FP generic input that is not written in
4418 * the VP is PointCoord.
4419 */
4420 memset(pntc, 0, 8 * sizeof(uint32_t));
4421
4422 for (i = 0; i < fp->cfg.in_nr; i++) {
4423 unsigned j, n = popcnt4(fp->cfg.in[i].mask);
4424
4425 if (fp->cfg.in[i].sn != TGSI_SEMANTIC_GENERIC) {
4426 m += n;
4427 continue;
4428 }
4429
4430 for (j = 0; j < vp->cfg.out_nr; ++j)
4431 if (vp->cfg.out[j].sn == fp->cfg.in[i].sn &&
4432 vp->cfg.out[j].si == fp->cfg.in[i].si)
4433 break;
4434
4435 if (j < vp->info.num_outputs) {
4436 ubyte enable =
4437 (nv50->rasterizer->pipe.sprite_coord_enable >> vp->cfg.out[j].si) & 1;
4438
4439 if (enable == 0) {
4440 m += n;
4441 continue;
4442 }
4443 }
4444
4445 /* this is either PointCoord or replaced by sprite coords */
4446 for (c = 0; c < 4; c++) {
4447 if (!(fp->cfg.in[i].mask & (1 << c)))
4448 continue;
4449 pntc[m / 8] |= (c + 1) << ((m % 8) * 4);
4450 ++m;
4451 }
4452 }
4453 return (nv50->rasterizer->pipe.sprite_coord_mode == PIPE_SPRITE_COORD_LOWER_LEFT ? 0 : origin);
4454 }
4455
4456 static int
4457 nv50_vec4_map(uint32_t *map32, int mid, uint8_t zval, uint32_t lin[4],
4458 struct nv50_sreg4 *fpi, struct nv50_sreg4 *vpo)
4459 {
4460 int c;
4461 uint8_t mv = vpo->mask, mf = fpi->mask, oid = vpo->hw;
4462 uint8_t *map = (uint8_t *)map32;
4463
4464 for (c = 0; c < 4; ++c) {
4465 if (mf & 1) {
4466 if (fpi->linear == TRUE)
4467 lin[mid / 32] |= 1 << (mid % 32);
4468 if (mv & 1)
4469 map[mid] = oid;
4470 else
4471 map[mid] = (c == 3) ? (zval + 1) : zval;
4472 ++mid;
4473 }
4474
4475 oid += mv & 1;
4476 mf >>= 1;
4477 mv >>= 1;
4478 }
4479
4480 return mid;
4481 }
4482
4483 struct nouveau_stateobj *
4484 nv50_fp_linkage_validate(struct nv50_context *nv50)
4485 {
4486 struct nouveau_grobj *tesla = nv50->screen->tesla;
4487 struct nv50_program *vp = nv50->vertprog;
4488 struct nv50_program *fp = nv50->fragprog;
4489 struct nouveau_stateobj *so;
4490 struct nv50_sreg4 dummy;
4491 int i, n, c, m = 0;
4492 uint32_t map[16], lin[4], reg[6], pcrd[8];
4493 uint8_t zval = 0x40;
4494
4495 if (nv50->geomprog) {
4496 vp = nv50->geomprog;
4497 zval = 0x80;
4498 }
4499 memset(map, 0, sizeof(map));
4500 memset(lin, 0, sizeof(lin));
4501
4502 reg[1] = 0x00000004; /* low and high clip distance map ids */
4503 reg[2] = 0x00000000; /* layer index map id (disabled, GP only) */
4504 reg[3] = 0x00000000; /* point size map id & enable */
4505 reg[5] = 0x00000000; /* primitive ID map slot */
4506 reg[0] = fp->cfg.regs[0]; /* colour semantic reg */
4507 reg[4] = fp->cfg.regs[1]; /* interpolant info */
4508
4509 dummy.linear = FALSE;
4510 dummy.mask = 0xf; /* map all components of HPOS */
4511 m = nv50_vec4_map(map, m, zval, lin, &dummy, &vp->cfg.out[0]);
4512
4513 dummy.mask = 0x0;
4514
4515 if (vp->cfg.clpd < 0x40) {
4516 for (c = 0; c < vp->cfg.clpd_nr; ++c) {
4517 map[m / 4] |= (vp->cfg.clpd + c) << ((m % 4) * 8);
4518 ++m;
4519 }
4520 reg[1] = (m << 8);
4521 }
4522
4523 reg[0] |= m << 8; /* adjust BFC0 id */
4524
4525 /* if light_twoside is active, it seems FFC0_ID == BFC0_ID is bad */
4526 if (nv50->rasterizer->pipe.light_twoside) {
4527 struct nv50_sreg4 *vpo = &vp->cfg.two_side[0];
4528 struct nv50_sreg4 *fpi = &fp->cfg.two_side[0];
4529
4530 m = nv50_vec4_map(map, m, zval, lin, &fpi[0], &vpo[0]);
4531 m = nv50_vec4_map(map, m, zval, lin, &fpi[1], &vpo[1]);
4532 }
4533
4534 reg[0] += m - 4; /* adjust FFC0 id */
4535 reg[4] |= m << 8; /* set mid where 'normal' FP inputs start */
4536
4537 for (i = 0; i < fp->cfg.in_nr; i++) {
4538 /* maybe even remove these from cfg.io */
4539 if (fp->cfg.in[i].sn == TGSI_SEMANTIC_POSITION ||
4540 fp->cfg.in[i].sn == TGSI_SEMANTIC_FACE)
4541 continue;
4542
4543 for (n = 0; n < vp->cfg.out_nr; ++n)
4544 if (vp->cfg.out[n].sn == fp->cfg.in[i].sn &&
4545 vp->cfg.out[n].si == fp->cfg.in[i].si)
4546 break;
4547
4548 m = nv50_vec4_map(map, m, zval, lin, &fp->cfg.in[i],
4549 (n < vp->cfg.out_nr) ?
4550 &vp->cfg.out[n] : &dummy);
4551 }
4552 /* PrimitiveID either is replaced by the system value, or
4553 * written by the geometry shader into an output register
4554 */
4555 if (fp->cfg.prim_id < 0x40) {
4556 map[m / 4] |= vp->cfg.prim_id << ((m % 4) * 8);
4557 reg[5] = m++;
4558 }
4559
4560 if (nv50->rasterizer->pipe.point_size_per_vertex) {
4561 map[m / 4] |= vp->cfg.psiz << ((m % 4) * 8);
4562 reg[3] = (m++ << 4) | 1;
4563 }
4564
4565 /* now fill the stateobj (at most 28 so_data) */
4566 so = so_new(10, 54, 0);
4567
4568 n = (m + 3) / 4;
4569 assert(m <= 64);
4570 if (vp->type == PIPE_SHADER_GEOMETRY) {
4571 so_method(so, tesla, NV50TCL_GP_RESULT_MAP_SIZE, 1);
4572 so_data (so, m);
4573 so_method(so, tesla, NV50TCL_GP_RESULT_MAP(0), n);
4574 so_datap (so, map, n);
4575 } else {
4576 so_method(so, tesla, NV50TCL_VP_GP_BUILTIN_ATTR_EN, 1);
4577 so_data (so, vp->cfg.regs[0]);
4578
4579 so_method(so, tesla, NV50TCL_MAP_SEMANTIC_4, 1);
4580 so_data (so, reg[5]);
4581
4582 so_method(so, tesla, NV50TCL_VP_RESULT_MAP_SIZE, 1);
4583 so_data (so, m);
4584 so_method(so, tesla, NV50TCL_VP_RESULT_MAP(0), n);
4585 so_datap (so, map, n);
4586 }
4587
4588 so_method(so, tesla, NV50TCL_MAP_SEMANTIC_0, 4);
4589 so_datap (so, reg, 4);
4590
4591 so_method(so, tesla, NV50TCL_FP_INTERPOLANT_CTRL, 1);
4592 so_data (so, reg[4]);
4593
4594 so_method(so, tesla, NV50TCL_NOPERSPECTIVE_BITMAP(0), 4);
4595 so_datap (so, lin, 4);
4596
4597 if (nv50->rasterizer->pipe.sprite_coord_enable) {
4598 so_method(so, tesla, NV50TCL_POINT_SPRITE_CTRL, 1);
4599 so_data (so,
4600 nv50_pntc_replace(nv50, pcrd, (reg[4] >> 8) & 0xff));
4601
4602 so_method(so, tesla, NV50TCL_POINT_COORD_REPLACE_MAP(0), 8);
4603 so_datap (so, pcrd, 8);
4604 }
4605
4606 so_method(so, tesla, NV50TCL_GP_ENABLE, 1);
4607 so_data (so, (vp->type == PIPE_SHADER_GEOMETRY) ? 1 : 0);
4608
4609 return so;
4610 }
4611
4612 static int
4613 construct_vp_gp_mapping(uint32_t *map32, int m,
4614 struct nv50_program *vp, struct nv50_program *gp)
4615 {
4616 uint8_t *map = (uint8_t *)map32;
4617 int i, j, c;
4618
4619 for (i = 0; i < gp->cfg.in_nr; ++i) {
4620 uint8_t oid = 0, mv = 0, mg = gp->cfg.in[i].mask;
4621
4622 for (j = 0; j < vp->cfg.out_nr; ++j) {
4623 if (vp->cfg.out[j].sn == gp->cfg.in[i].sn &&
4624 vp->cfg.out[j].si == gp->cfg.in[i].si) {
4625 mv = vp->cfg.out[j].mask;
4626 oid = vp->cfg.out[j].hw;
4627 break;
4628 }
4629 }
4630
4631 for (c = 0; c < 4; ++c, mv >>= 1, mg >>= 1) {
4632 if (mg & mv & 1)
4633 map[m++] = oid;
4634 else
4635 if (mg & 1)
4636 map[m++] = (c == 3) ? 0x41 : 0x40;
4637 oid += mv & 1;
4638 }
4639 }
4640 return m;
4641 }
4642
4643 struct nouveau_stateobj *
4644 nv50_gp_linkage_validate(struct nv50_context *nv50)
4645 {
4646 struct nouveau_grobj *tesla = nv50->screen->tesla;
4647 struct nouveau_stateobj *so;
4648 struct nv50_program *vp = nv50->vertprog;
4649 struct nv50_program *gp = nv50->geomprog;
4650 uint32_t map[16];
4651 int m = 0;
4652
4653 if (!gp)
4654 return NULL;
4655 memset(map, 0, sizeof(map));
4656
4657 m = construct_vp_gp_mapping(map, m, vp, gp);
4658
4659 so = so_new(3, 24 - 3, 0);
4660
4661 so_method(so, tesla, NV50TCL_VP_GP_BUILTIN_ATTR_EN, 1);
4662 so_data (so, vp->cfg.regs[0] | gp->cfg.regs[0]);
4663
4664 assert(m <= 32);
4665 so_method(so, tesla, NV50TCL_VP_RESULT_MAP_SIZE, 1);
4666 so_data (so, m);
4667
4668 m = (m + 3) / 4;
4669 so_method(so, tesla, NV50TCL_VP_RESULT_MAP(0), m);
4670 so_datap (so, map, m);
4671
4672 return so;
4673 }
4674
4675 void
4676 nv50_program_destroy(struct nv50_context *nv50, struct nv50_program *p)
4677 {
4678 while (p->exec_head) {
4679 struct nv50_program_exec *e = p->exec_head;
4680
4681 p->exec_head = e->next;
4682 FREE(e);
4683 }
4684 p->exec_tail = NULL;
4685 p->exec_size = 0;
4686
4687 nouveau_bo_ref(NULL, &p->bo);
4688
4689 FREE(p->immd);
4690 nouveau_resource_free(&p->data[0]);
4691
4692 p->translated = 0;
4693 }