projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ir3: LDC also has a destination
[mesa.git] / src / freedreno / ir3 / ir3.c
1 /*
2 * Copyright (c) 2012 Rob Clark <robdclark@gmail.com>
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 (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 #include "ir3.h"
25
26 #include <stdlib.h>
27 #include <stdio.h>
28 #include <string.h>
29 #include <assert.h>
30 #include <stdbool.h>
31 #include <errno.h>
32
33 #include "util/bitscan.h"
34 #include "util/ralloc.h"
35 #include "util/u_math.h"
36
37 #include "instr-a3xx.h"
38 #include "ir3_compiler.h"
39
40 /* simple allocator to carve allocations out of an up-front allocated heap,
41 * so that we can free everything easily in one shot.
42 */
43 void * ir3_alloc(struct ir3 *shader, int sz)
44 {
45 return rzalloc_size(shader, sz); /* TODO: don't use rzalloc */
46 }
47
48 struct ir3 * ir3_create(struct ir3_compiler *compiler, gl_shader_stage type)
49 {
50 struct ir3 *shader = rzalloc(NULL, struct ir3);
51
52 shader->compiler = compiler;
53 shader->type = type;
54
55 list_inithead(&shader->block_list);
56 list_inithead(&shader->array_list);
57
58 return shader;
59 }
60
61 void ir3_destroy(struct ir3 *shader)
62 {
63 ralloc_free(shader);
64 }
65
66 #define iassert(cond) do { \
67 if (!(cond)) { \
68 debug_assert(cond); \
69 return -1; \
70 } } while (0)
71
72 #define iassert_type(reg, full) do { \
73 if ((full)) { \
74 iassert(!((reg)->flags & IR3_REG_HALF)); \
75 } else { \
76 iassert((reg)->flags & IR3_REG_HALF); \
77 } } while (0);
78
79 static uint32_t reg(struct ir3_register *reg, struct ir3_info *info,
80 uint32_t repeat, uint32_t valid_flags)
81 {
82 reg_t val = { .dummy32 = 0 };
83
84 if (reg->flags & ~valid_flags) {
85 debug_printf("INVALID FLAGS: %x vs %x\n",
86 reg->flags, valid_flags);
87 }
88
89 if (!(reg->flags & IR3_REG_R))
90 repeat = 0;
91
92 if (reg->flags & IR3_REG_IMMED) {
93 val.iim_val = reg->iim_val;
94 } else {
95 unsigned components;
96 int16_t max;
97
98 if (reg->flags & IR3_REG_RELATIV) {
99 components = reg->size;
100 val.idummy10 = reg->array.offset;
101 max = (reg->array.offset + repeat + components - 1);
102 } else {
103 components = util_last_bit(reg->wrmask);
104 val.comp = reg->num & 0x3;
105 val.num = reg->num >> 2;
106 max = (reg->num + repeat + components - 1);
107 }
108
109 if (reg->flags & IR3_REG_CONST) {
110 info->max_const = MAX2(info->max_const, max >> 2);
111 } else if (val.num == 63) {
112 /* ignore writes to dummy register r63.x */
113 } else if (max < regid(48, 0)) {
114 if (reg->flags & IR3_REG_HALF) {
115 if (info->gpu_id >= 600) {
116 /* starting w/ a6xx, half regs conflict with full regs: */
117 info->max_reg = MAX2(info->max_reg, max >> 3);
118 } else {
119 info->max_half_reg = MAX2(info->max_half_reg, max >> 2);
120 }
121 } else {
122 info->max_reg = MAX2(info->max_reg, max >> 2);
123 }
124 }
125 }
126
127 return val.dummy32;
128 }
129
130 static int emit_cat0(struct ir3_instruction *instr, void *ptr,
131 struct ir3_info *info)
132 {
133 instr_cat0_t *cat0 = ptr;
134
135 if (info->gpu_id >= 500) {
136 cat0->a5xx.immed = instr->cat0.immed;
137 } else if (info->gpu_id >= 400) {
138 cat0->a4xx.immed = instr->cat0.immed;
139 } else {
140 cat0->a3xx.immed = instr->cat0.immed;
141 }
142 cat0->repeat = instr->repeat;
143 cat0->ss = !!(instr->flags & IR3_INSTR_SS);
144 cat0->inv = instr->cat0.inv;
145 cat0->comp = instr->cat0.comp;
146 cat0->opc = instr->opc;
147 cat0->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
148 cat0->sync = !!(instr->flags & IR3_INSTR_SY);
149 cat0->opc_cat = 0;
150
151 switch (instr->opc) {
152 case OPC_IF:
153 case OPC_ELSE:
154 case OPC_ENDIF:
155 cat0->dummy4 = 16;
156 break;
157 default:
158 break;
159 }
160
161 return 0;
162 }
163
164 static int emit_cat1(struct ir3_instruction *instr, void *ptr,
165 struct ir3_info *info)
166 {
167 struct ir3_register *dst = instr->regs[0];
168 struct ir3_register *src = instr->regs[1];
169 instr_cat1_t *cat1 = ptr;
170
171 iassert(instr->regs_count == 2);
172 iassert_type(dst, type_size(instr->cat1.dst_type) == 32);
173 if (!(src->flags & IR3_REG_IMMED))
174 iassert_type(src, type_size(instr->cat1.src_type) == 32);
175
176 if (src->flags & IR3_REG_IMMED) {
177 cat1->iim_val = src->iim_val;
178 cat1->src_im = 1;
179 } else if (src->flags & IR3_REG_RELATIV) {
180 cat1->off = reg(src, info, instr->repeat,
181 IR3_REG_R | IR3_REG_CONST | IR3_REG_HALF | IR3_REG_RELATIV);
182 cat1->src_rel = 1;
183 cat1->src_rel_c = !!(src->flags & IR3_REG_CONST);
184 } else {
185 cat1->src = reg(src, info, instr->repeat,
186 IR3_REG_R | IR3_REG_CONST | IR3_REG_HALF);
187 cat1->src_c = !!(src->flags & IR3_REG_CONST);
188 }
189
190 cat1->dst = reg(dst, info, instr->repeat,
191 IR3_REG_RELATIV | IR3_REG_EVEN |
192 IR3_REG_R | IR3_REG_POS_INF | IR3_REG_HALF);
193 cat1->repeat = instr->repeat;
194 cat1->src_r = !!(src->flags & IR3_REG_R);
195 cat1->ss = !!(instr->flags & IR3_INSTR_SS);
196 cat1->ul = !!(instr->flags & IR3_INSTR_UL);
197 cat1->dst_type = instr->cat1.dst_type;
198 cat1->dst_rel = !!(dst->flags & IR3_REG_RELATIV);
199 cat1->src_type = instr->cat1.src_type;
200 cat1->even = !!(dst->flags & IR3_REG_EVEN);
201 cat1->pos_inf = !!(dst->flags & IR3_REG_POS_INF);
202 cat1->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
203 cat1->sync = !!(instr->flags & IR3_INSTR_SY);
204 cat1->opc_cat = 1;
205
206 return 0;
207 }
208
209 static int emit_cat2(struct ir3_instruction *instr, void *ptr,
210 struct ir3_info *info)
211 {
212 struct ir3_register *dst = instr->regs[0];
213 struct ir3_register *src1 = instr->regs[1];
214 struct ir3_register *src2 = instr->regs[2];
215 instr_cat2_t *cat2 = ptr;
216 unsigned absneg = ir3_cat2_absneg(instr->opc);
217
218 iassert((instr->regs_count == 2) || (instr->regs_count == 3));
219
220 if (instr->nop) {
221 iassert(!instr->repeat);
222 iassert(instr->nop <= 3);
223
224 cat2->src1_r = instr->nop & 0x1;
225 cat2->src2_r = (instr->nop >> 1) & 0x1;
226 } else {
227 cat2->src1_r = !!(src1->flags & IR3_REG_R);
228 if (src2)
229 cat2->src2_r = !!(src2->flags & IR3_REG_R);
230 }
231
232 if (src1->flags & IR3_REG_RELATIV) {
233 iassert(src1->array.offset < (1 << 10));
234 cat2->rel1.src1 = reg(src1, info, instr->repeat,
235 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
236 IR3_REG_HALF | absneg);
237 cat2->rel1.src1_c = !!(src1->flags & IR3_REG_CONST);
238 cat2->rel1.src1_rel = 1;
239 } else if (src1->flags & IR3_REG_CONST) {
240 iassert(src1->num < (1 << 12));
241 cat2->c1.src1 = reg(src1, info, instr->repeat,
242 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF |
243 absneg);
244 cat2->c1.src1_c = 1;
245 } else {
246 iassert(src1->num < (1 << 11));
247 cat2->src1 = reg(src1, info, instr->repeat,
248 IR3_REG_IMMED | IR3_REG_R | IR3_REG_HALF |
249 absneg);
250 }
251 cat2->src1_im = !!(src1->flags & IR3_REG_IMMED);
252 cat2->src1_neg = !!(src1->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
253 cat2->src1_abs = !!(src1->flags & (IR3_REG_FABS | IR3_REG_SABS));
254
255 if (src2) {
256 iassert((src2->flags & IR3_REG_IMMED) ||
257 !((src1->flags ^ src2->flags) & IR3_REG_HALF));
258
259 if (src2->flags & IR3_REG_RELATIV) {
260 iassert(src2->array.offset < (1 << 10));
261 cat2->rel2.src2 = reg(src2, info, instr->repeat,
262 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
263 IR3_REG_HALF | absneg);
264 cat2->rel2.src2_c = !!(src2->flags & IR3_REG_CONST);
265 cat2->rel2.src2_rel = 1;
266 } else if (src2->flags & IR3_REG_CONST) {
267 iassert(src2->num < (1 << 12));
268 cat2->c2.src2 = reg(src2, info, instr->repeat,
269 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF |
270 absneg);
271 cat2->c2.src2_c = 1;
272 } else {
273 iassert(src2->num < (1 << 11));
274 cat2->src2 = reg(src2, info, instr->repeat,
275 IR3_REG_IMMED | IR3_REG_R | IR3_REG_HALF |
276 absneg);
277 }
278
279 cat2->src2_im = !!(src2->flags & IR3_REG_IMMED);
280 cat2->src2_neg = !!(src2->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
281 cat2->src2_abs = !!(src2->flags & (IR3_REG_FABS | IR3_REG_SABS));
282 }
283
284 cat2->dst = reg(dst, info, instr->repeat,
285 IR3_REG_R | IR3_REG_EI | IR3_REG_HALF);
286 cat2->repeat = instr->repeat;
287 cat2->sat = !!(instr->flags & IR3_INSTR_SAT);
288 cat2->ss = !!(instr->flags & IR3_INSTR_SS);
289 cat2->ul = !!(instr->flags & IR3_INSTR_UL);
290 cat2->dst_half = !!((src1->flags ^ dst->flags) & IR3_REG_HALF);
291 cat2->ei = !!(dst->flags & IR3_REG_EI);
292 cat2->cond = instr->cat2.condition;
293 cat2->full = ! (src1->flags & IR3_REG_HALF);
294 cat2->opc = instr->opc;
295 cat2->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
296 cat2->sync = !!(instr->flags & IR3_INSTR_SY);
297 cat2->opc_cat = 2;
298
299 return 0;
300 }
301
302 static int emit_cat3(struct ir3_instruction *instr, void *ptr,
303 struct ir3_info *info)
304 {
305 struct ir3_register *dst = instr->regs[0];
306 struct ir3_register *src1 = instr->regs[1];
307 struct ir3_register *src2 = instr->regs[2];
308 struct ir3_register *src3 = instr->regs[3];
309 unsigned absneg = ir3_cat3_absneg(instr->opc);
310 instr_cat3_t *cat3 = ptr;
311 uint32_t src_flags = 0;
312
313 switch (instr->opc) {
314 case OPC_MAD_F16:
315 case OPC_MAD_U16:
316 case OPC_MAD_S16:
317 case OPC_SEL_B16:
318 case OPC_SEL_S16:
319 case OPC_SEL_F16:
320 case OPC_SAD_S16:
321 case OPC_SAD_S32: // really??
322 src_flags |= IR3_REG_HALF;
323 break;
324 default:
325 break;
326 }
327
328 iassert(instr->regs_count == 4);
329 iassert(!((src1->flags ^ src_flags) & IR3_REG_HALF));
330 iassert(!((src2->flags ^ src_flags) & IR3_REG_HALF));
331 iassert(!((src3->flags ^ src_flags) & IR3_REG_HALF));
332
333 if (instr->nop) {
334 iassert(!instr->repeat);
335 iassert(instr->nop <= 3);
336
337 cat3->src1_r = instr->nop & 0x1;
338 cat3->src2_r = (instr->nop >> 1) & 0x1;
339 } else {
340 cat3->src1_r = !!(src1->flags & IR3_REG_R);
341 cat3->src2_r = !!(src2->flags & IR3_REG_R);
342 }
343
344 if (src1->flags & IR3_REG_RELATIV) {
345 iassert(src1->array.offset < (1 << 10));
346 cat3->rel1.src1 = reg(src1, info, instr->repeat,
347 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
348 IR3_REG_HALF | absneg);
349 cat3->rel1.src1_c = !!(src1->flags & IR3_REG_CONST);
350 cat3->rel1.src1_rel = 1;
351 } else if (src1->flags & IR3_REG_CONST) {
352 iassert(src1->num < (1 << 12));
353 cat3->c1.src1 = reg(src1, info, instr->repeat,
354 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF | absneg);
355 cat3->c1.src1_c = 1;
356 } else {
357 iassert(src1->num < (1 << 11));
358 cat3->src1 = reg(src1, info, instr->repeat,
359 IR3_REG_R | IR3_REG_HALF | absneg);
360 }
361
362 cat3->src1_neg = !!(src1->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
363
364 cat3->src2 = reg(src2, info, instr->repeat,
365 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF | absneg);
366 cat3->src2_c = !!(src2->flags & IR3_REG_CONST);
367 cat3->src2_neg = !!(src2->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
368
369 if (src3->flags & IR3_REG_RELATIV) {
370 iassert(src3->array.offset < (1 << 10));
371 cat3->rel2.src3 = reg(src3, info, instr->repeat,
372 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
373 IR3_REG_HALF | absneg);
374 cat3->rel2.src3_c = !!(src3->flags & IR3_REG_CONST);
375 cat3->rel2.src3_rel = 1;
376 } else if (src3->flags & IR3_REG_CONST) {
377 iassert(src3->num < (1 << 12));
378 cat3->c2.src3 = reg(src3, info, instr->repeat,
379 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF | absneg);
380 cat3->c2.src3_c = 1;
381 } else {
382 iassert(src3->num < (1 << 11));
383 cat3->src3 = reg(src3, info, instr->repeat,
384 IR3_REG_R | IR3_REG_HALF | absneg);
385 }
386
387 cat3->src3_neg = !!(src3->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
388 cat3->src3_r = !!(src3->flags & IR3_REG_R);
389
390 cat3->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
391 cat3->repeat = instr->repeat;
392 cat3->sat = !!(instr->flags & IR3_INSTR_SAT);
393 cat3->ss = !!(instr->flags & IR3_INSTR_SS);
394 cat3->ul = !!(instr->flags & IR3_INSTR_UL);
395 cat3->dst_half = !!((src_flags ^ dst->flags) & IR3_REG_HALF);
396 cat3->opc = instr->opc;
397 cat3->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
398 cat3->sync = !!(instr->flags & IR3_INSTR_SY);
399 cat3->opc_cat = 3;
400
401 return 0;
402 }
403
404 static int emit_cat4(struct ir3_instruction *instr, void *ptr,
405 struct ir3_info *info)
406 {
407 struct ir3_register *dst = instr->regs[0];
408 struct ir3_register *src = instr->regs[1];
409 instr_cat4_t *cat4 = ptr;
410
411 iassert(instr->regs_count == 2);
412
413 if (src->flags & IR3_REG_RELATIV) {
414 iassert(src->array.offset < (1 << 10));
415 cat4->rel.src = reg(src, info, instr->repeat,
416 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_FNEG |
417 IR3_REG_FABS | IR3_REG_R | IR3_REG_HALF);
418 cat4->rel.src_c = !!(src->flags & IR3_REG_CONST);
419 cat4->rel.src_rel = 1;
420 } else if (src->flags & IR3_REG_CONST) {
421 iassert(src->num < (1 << 12));
422 cat4->c.src = reg(src, info, instr->repeat,
423 IR3_REG_CONST | IR3_REG_FNEG | IR3_REG_FABS |
424 IR3_REG_R | IR3_REG_HALF);
425 cat4->c.src_c = 1;
426 } else {
427 iassert(src->num < (1 << 11));
428 cat4->src = reg(src, info, instr->repeat,
429 IR3_REG_IMMED | IR3_REG_FNEG | IR3_REG_FABS |
430 IR3_REG_R | IR3_REG_HALF);
431 }
432
433 cat4->src_im = !!(src->flags & IR3_REG_IMMED);
434 cat4->src_neg = !!(src->flags & IR3_REG_FNEG);
435 cat4->src_abs = !!(src->flags & IR3_REG_FABS);
436 cat4->src_r = !!(src->flags & IR3_REG_R);
437
438 cat4->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
439 cat4->repeat = instr->repeat;
440 cat4->sat = !!(instr->flags & IR3_INSTR_SAT);
441 cat4->ss = !!(instr->flags & IR3_INSTR_SS);
442 cat4->ul = !!(instr->flags & IR3_INSTR_UL);
443 cat4->dst_half = !!((src->flags ^ dst->flags) & IR3_REG_HALF);
444 cat4->full = ! (src->flags & IR3_REG_HALF);
445 cat4->opc = instr->opc;
446 cat4->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
447 cat4->sync = !!(instr->flags & IR3_INSTR_SY);
448 cat4->opc_cat = 4;
449
450 return 0;
451 }
452
453 static int emit_cat5(struct ir3_instruction *instr, void *ptr,
454 struct ir3_info *info)
455 {
456 struct ir3_register *dst = instr->regs[0];
457 /* To simplify things when there could be zero, one, or two args other
458 * than tex/sampler idx, we use the first src reg in the ir to hold
459 * samp_tex hvec2:
460 */
461 struct ir3_register *src1;
462 struct ir3_register *src2;
463 instr_cat5_t *cat5 = ptr;
464
465 iassert((instr->regs_count == 1) ||
466 (instr->regs_count == 2) ||
467 (instr->regs_count == 3) ||
468 (instr->regs_count == 4));
469
470 if (instr->flags & IR3_INSTR_S2EN) {
471 src1 = instr->regs[2];
472 src2 = instr->regs_count > 3 ? instr->regs[3] : NULL;
473 } else {
474 src1 = instr->regs_count > 1 ? instr->regs[1] : NULL;
475 src2 = instr->regs_count > 2 ? instr->regs[2] : NULL;
476 }
477
478 assume(src1 || !src2);
479
480 if (src1) {
481 cat5->full = ! (src1->flags & IR3_REG_HALF);
482 cat5->src1 = reg(src1, info, instr->repeat, IR3_REG_HALF);
483 }
484
485 if (src2) {
486 iassert(!((src1->flags ^ src2->flags) & IR3_REG_HALF));
487 cat5->src2 = reg(src2, info, instr->repeat, IR3_REG_HALF);
488 }
489
490 if (instr->flags & IR3_INSTR_S2EN) {
491 struct ir3_register *samp_tex = instr->regs[1];
492 iassert(samp_tex->flags & IR3_REG_HALF);
493 cat5->s2en_bindless.src3 = reg(samp_tex, info, instr->repeat, IR3_REG_HALF);
494 /* TODO: This should probably be CAT5_UNIFORM, at least on a6xx, as
495 * this is what the blob does and it is presumably faster, but first
496 * we should confirm it is actually nonuniform and figure out when the
497 * whole descriptor mode mechanism was introduced.
498 */
499 cat5->s2en_bindless.desc_mode = CAT5_NONUNIFORM;
500 iassert(!(instr->cat5.samp | instr->cat5.tex));
501 } else {
502 cat5->norm.samp = instr->cat5.samp;
503 cat5->norm.tex = instr->cat5.tex;
504 }
505
506 cat5->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
507 cat5->wrmask = dst->wrmask;
508 cat5->type = instr->cat5.type;
509 cat5->is_3d = !!(instr->flags & IR3_INSTR_3D);
510 cat5->is_a = !!(instr->flags & IR3_INSTR_A);
511 cat5->is_s = !!(instr->flags & IR3_INSTR_S);
512 cat5->is_s2en_bindless = !!(instr->flags & IR3_INSTR_S2EN);
513 cat5->is_o = !!(instr->flags & IR3_INSTR_O);
514 cat5->is_p = !!(instr->flags & IR3_INSTR_P);
515 cat5->opc = instr->opc;
516 cat5->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
517 cat5->sync = !!(instr->flags & IR3_INSTR_SY);
518 cat5->opc_cat = 5;
519
520 return 0;
521 }
522
523 static int emit_cat6_a6xx(struct ir3_instruction *instr, void *ptr,
524 struct ir3_info *info)
525 {
526 struct ir3_register *src1, *src2;
527 instr_cat6_a6xx_t *cat6 = ptr;
528 bool has_dest = (instr->opc == OPC_LDIB || instr->opc == OPC_LDC);
529
530 /* first reg should be SSBO binding point: */
531 iassert(instr->regs[1]->flags & IR3_REG_IMMED);
532
533 src1 = instr->regs[2];
534
535 if (has_dest) {
536 /* the src2 field in the instruction is actually the destination
537 * register for load instructions:
538 */
539 src2 = instr->regs[0];
540 } else {
541 src2 = instr->regs[3];
542 }
543
544 cat6->type = instr->cat6.type;
545 cat6->d = instr->cat6.d - 1;
546 cat6->typed = instr->cat6.typed;
547 cat6->type_size = instr->cat6.iim_val - 1;
548 cat6->opc = instr->opc;
549 cat6->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
550 cat6->sync = !!(instr->flags & IR3_INSTR_SY);
551 cat6->opc_cat = 6;
552
553 cat6->src1 = reg(src1, info, instr->repeat, 0);
554 cat6->src2 = reg(src2, info, instr->repeat, 0);
555 cat6->ssbo = instr->regs[1]->iim_val;
556
557 switch (instr->opc) {
558 case OPC_ATOMIC_ADD:
559 case OPC_ATOMIC_SUB:
560 case OPC_ATOMIC_XCHG:
561 case OPC_ATOMIC_INC:
562 case OPC_ATOMIC_DEC:
563 case OPC_ATOMIC_CMPXCHG:
564 case OPC_ATOMIC_MIN:
565 case OPC_ATOMIC_MAX:
566 case OPC_ATOMIC_AND:
567 case OPC_ATOMIC_OR:
568 case OPC_ATOMIC_XOR:
569 cat6->pad1 = 0x1;
570 cat6->pad3 = 0xc;
571 cat6->pad5 = 0x3;
572 break;
573 case OPC_STIB:
574 cat6->pad1 = 0x0;
575 cat6->pad3 = 0xc;
576 cat6->pad5 = 0x2;
577 break;
578 case OPC_LDIB:
579 cat6->pad1 = 0x1;
580 cat6->pad3 = 0xc;
581 cat6->pad5 = 0x2;
582 break;
583 case OPC_LDC:
584 cat6->pad1 = 0x0;
585 cat6->pad3 = 0x8;
586 cat6->pad5 = 0x2;
587 break;
588 default:
589 iassert(0);
590 }
591 cat6->pad2 = 0x0;
592 cat6->pad4 = 0x0;
593
594 return 0;
595 }
596
597 static int emit_cat6(struct ir3_instruction *instr, void *ptr,
598 struct ir3_info *info)
599 {
600 struct ir3_register *dst, *src1, *src2;
601 instr_cat6_t *cat6 = ptr;
602
603 /* In a6xx we start using a new instruction encoding for some of
604 * these instructions:
605 */
606 if (info->gpu_id >= 600) {
607 switch (instr->opc) {
608 case OPC_ATOMIC_ADD:
609 case OPC_ATOMIC_SUB:
610 case OPC_ATOMIC_XCHG:
611 case OPC_ATOMIC_INC:
612 case OPC_ATOMIC_DEC:
613 case OPC_ATOMIC_CMPXCHG:
614 case OPC_ATOMIC_MIN:
615 case OPC_ATOMIC_MAX:
616 case OPC_ATOMIC_AND:
617 case OPC_ATOMIC_OR:
618 case OPC_ATOMIC_XOR:
619 /* The shared variants of these still use the old encoding: */
620 if (!(instr->flags & IR3_INSTR_G))
621 break;
622 /* fallthrough */
623 case OPC_STIB:
624 case OPC_LDIB:
625 case OPC_LDC:
626 return emit_cat6_a6xx(instr, ptr, info);
627 default:
628 break;
629 }
630 }
631
632 bool type_full = type_size(instr->cat6.type) == 32;
633
634 cat6->type = instr->cat6.type;
635 cat6->opc = instr->opc;
636 cat6->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
637 cat6->sync = !!(instr->flags & IR3_INSTR_SY);
638 cat6->g = !!(instr->flags & IR3_INSTR_G);
639 cat6->opc_cat = 6;
640
641 switch (instr->opc) {
642 case OPC_RESINFO:
643 case OPC_RESFMT:
644 iassert_type(instr->regs[0], type_full); /* dst */
645 iassert_type(instr->regs[1], type_full); /* src1 */
646 break;
647 case OPC_L2G:
648 case OPC_G2L:
649 iassert_type(instr->regs[0], true); /* dst */
650 iassert_type(instr->regs[1], true); /* src1 */
651 break;
652 case OPC_STG:
653 case OPC_STL:
654 case OPC_STP:
655 case OPC_STLW:
656 case OPC_STIB:
657 /* no dst, so regs[0] is dummy */
658 iassert_type(instr->regs[1], true); /* dst */
659 iassert_type(instr->regs[2], type_full); /* src1 */
660 iassert_type(instr->regs[3], true); /* src2 */
661 break;
662 default:
663 iassert_type(instr->regs[0], type_full); /* dst */
664 iassert_type(instr->regs[1], true); /* src1 */
665 if (instr->regs_count > 2)
666 iassert_type(instr->regs[2], true); /* src1 */
667 break;
668 }
669
670 /* the "dst" for a store instruction is (from the perspective
671 * of data flow in the shader, ie. register use/def, etc) in
672 * fact a register that is read by the instruction, rather
673 * than written:
674 */
675 if (is_store(instr)) {
676 iassert(instr->regs_count >= 3);
677
678 dst = instr->regs[1];
679 src1 = instr->regs[2];
680 src2 = (instr->regs_count >= 4) ? instr->regs[3] : NULL;
681 } else {
682 iassert(instr->regs_count >= 2);
683
684 dst = instr->regs[0];
685 src1 = instr->regs[1];
686 src2 = (instr->regs_count >= 3) ? instr->regs[2] : NULL;
687 }
688
689 /* TODO we need a more comprehensive list about which instructions
690 * can be encoded which way. Or possibly use IR3_INSTR_0 flag to
691 * indicate to use the src_off encoding even if offset is zero
692 * (but then what to do about dst_off?)
693 */
694 if (is_atomic(instr->opc)) {
695 instr_cat6ldgb_t *ldgb = ptr;
696
697 /* maybe these two bits both determine the instruction encoding? */
698 cat6->src_off = false;
699
700 ldgb->d = instr->cat6.d - 1;
701 ldgb->typed = instr->cat6.typed;
702 ldgb->type_size = instr->cat6.iim_val - 1;
703
704 ldgb->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
705
706 if (ldgb->g) {
707 struct ir3_register *src3 = instr->regs[3];
708 struct ir3_register *src4 = instr->regs[4];
709
710 /* first src is src_ssbo: */
711 iassert(src1->flags & IR3_REG_IMMED);
712 ldgb->src_ssbo = src1->uim_val;
713
714 ldgb->src1 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
715 ldgb->src1_im = !!(src2->flags & IR3_REG_IMMED);
716 ldgb->src2 = reg(src3, info, instr->repeat, IR3_REG_IMMED);
717 ldgb->src2_im = !!(src3->flags & IR3_REG_IMMED);
718
719 ldgb->src3 = reg(src4, info, instr->repeat, 0);
720 ldgb->pad0 = 0x1;
721 ldgb->pad3 = 0x1;
722 } else {
723 ldgb->src1 = reg(src1, info, instr->repeat, IR3_REG_IMMED);
724 ldgb->src1_im = !!(src1->flags & IR3_REG_IMMED);
725 ldgb->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
726 ldgb->src2_im = !!(src2->flags & IR3_REG_IMMED);
727 ldgb->pad0 = 0x1;
728 ldgb->pad3 = 0x0;
729 }
730
731 return 0;
732 } else if (instr->opc == OPC_LDGB) {
733 struct ir3_register *src3 = instr->regs[3];
734 instr_cat6ldgb_t *ldgb = ptr;
735
736 /* maybe these two bits both determine the instruction encoding? */
737 cat6->src_off = false;
738
739 ldgb->d = instr->cat6.d - 1;
740 ldgb->typed = instr->cat6.typed;
741 ldgb->type_size = instr->cat6.iim_val - 1;
742
743 ldgb->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
744
745 /* first src is src_ssbo: */
746 iassert(src1->flags & IR3_REG_IMMED);
747 ldgb->src_ssbo = src1->uim_val;
748
749 /* then next two are src1/src2: */
750 ldgb->src1 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
751 ldgb->src1_im = !!(src2->flags & IR3_REG_IMMED);
752 ldgb->src2 = reg(src3, info, instr->repeat, IR3_REG_IMMED);
753 ldgb->src2_im = !!(src3->flags & IR3_REG_IMMED);
754
755 ldgb->pad0 = 0x0;
756 ldgb->pad3 = 0x1;
757
758 return 0;
759 } else if (instr->opc == OPC_RESINFO) {
760 instr_cat6ldgb_t *ldgb = ptr;
761
762 ldgb->d = instr->cat6.d - 1;
763
764 ldgb->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
765
766 /* first src is src_ssbo: */
767 iassert(src1->flags & IR3_REG_IMMED);
768 ldgb->src_ssbo = src1->uim_val;
769
770 return 0;
771 } else if ((instr->opc == OPC_STGB) || (instr->opc == OPC_STIB)) {
772 struct ir3_register *src3 = instr->regs[4];
773 instr_cat6stgb_t *stgb = ptr;
774
775 /* maybe these two bits both determine the instruction encoding? */
776 cat6->src_off = true;
777 stgb->pad3 = 0x2;
778
779 stgb->d = instr->cat6.d - 1;
780 stgb->typed = instr->cat6.typed;
781 stgb->type_size = instr->cat6.iim_val - 1;
782
783 /* first src is dst_ssbo: */
784 iassert(dst->flags & IR3_REG_IMMED);
785 stgb->dst_ssbo = dst->uim_val;
786
787 /* then src1/src2/src3: */
788 stgb->src1 = reg(src1, info, instr->repeat, 0);
789 stgb->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
790 stgb->src2_im = !!(src2->flags & IR3_REG_IMMED);
791 stgb->src3 = reg(src3, info, instr->repeat, IR3_REG_IMMED);
792 stgb->src3_im = !!(src3->flags & IR3_REG_IMMED);
793
794 return 0;
795 } else if (instr->cat6.src_offset || (instr->opc == OPC_LDG) ||
796 (instr->opc == OPC_LDL) || (instr->opc == OPC_LDLW)) {
797 struct ir3_register *src3 = instr->regs[3];
798 instr_cat6a_t *cat6a = ptr;
799
800 cat6->src_off = true;
801
802 if (instr->opc == OPC_LDG) {
803 /* For LDG src1 can not be immediate, so src1_imm is redundant and
804 * instead used to signal whether (when true) 'off' is a 32 bit
805 * register or an immediate offset.
806 */
807 cat6a->src1 = reg(src1, info, instr->repeat, 0);
808 cat6a->src1_im = !(src3->flags & IR3_REG_IMMED);
809 cat6a->off = reg(src3, info, instr->repeat, IR3_REG_IMMED);
810 } else {
811 cat6a->src1 = reg(src1, info, instr->repeat, IR3_REG_IMMED);
812 cat6a->src1_im = !!(src1->flags & IR3_REG_IMMED);
813 cat6a->off = reg(src3, info, instr->repeat, IR3_REG_IMMED);
814 iassert(src3->flags & IR3_REG_IMMED);
815 }
816
817 /* Num components */
818 cat6a->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
819 cat6a->src2_im = true;
820 } else {
821 instr_cat6b_t *cat6b = ptr;
822
823 cat6->src_off = false;
824
825 cat6b->src1 = reg(src1, info, instr->repeat, IR3_REG_IMMED | IR3_REG_HALF);
826 cat6b->src1_im = !!(src1->flags & IR3_REG_IMMED);
827 if (src2) {
828 cat6b->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
829 cat6b->src2_im = !!(src2->flags & IR3_REG_IMMED);
830 }
831 }
832
833 if (instr->cat6.dst_offset || (instr->opc == OPC_STG) ||
834 (instr->opc == OPC_STL) || (instr->opc == OPC_STLW)) {
835 instr_cat6c_t *cat6c = ptr;
836 cat6->dst_off = true;
837 cat6c->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
838
839 if (instr->flags & IR3_INSTR_G) {
840 struct ir3_register *src3 = instr->regs[4];
841 cat6c->off = reg(src3, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
842 if (src3->flags & IR3_REG_IMMED) {
843 /* Immediate offsets are in bytes... */
844 cat6->g = false;
845 cat6c->off *= 4;
846 }
847 } else {
848 cat6c->off = instr->cat6.dst_offset;
849 }
850 } else {
851 instr_cat6d_t *cat6d = ptr;
852 cat6->dst_off = false;
853 cat6d->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
854 }
855
856 return 0;
857 }
858
859 static int emit_cat7(struct ir3_instruction *instr, void *ptr,
860 struct ir3_info *info)
861 {
862 instr_cat7_t *cat7 = ptr;
863
864 cat7->ss = !!(instr->flags & IR3_INSTR_SS);
865 cat7->w = instr->cat7.w;
866 cat7->r = instr->cat7.r;
867 cat7->l = instr->cat7.l;
868 cat7->g = instr->cat7.g;
869 cat7->opc = instr->opc;
870 cat7->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
871 cat7->sync = !!(instr->flags & IR3_INSTR_SY);
872 cat7->opc_cat = 7;
873
874 return 0;
875 }
876
877 static int (*emit[])(struct ir3_instruction *instr, void *ptr,
878 struct ir3_info *info) = {
879 emit_cat0, emit_cat1, emit_cat2, emit_cat3, emit_cat4, emit_cat5, emit_cat6,
880 emit_cat7,
881 };
882
883 void * ir3_assemble(struct ir3 *shader, struct ir3_info *info,
884 uint32_t gpu_id)
885 {
886 uint32_t *ptr, *dwords;
887
888 info->gpu_id = gpu_id;
889 info->max_reg = -1;
890 info->max_half_reg = -1;
891 info->max_const = -1;
892 info->instrs_count = 0;
893 info->sizedwords = 0;
894 info->ss = info->sy = 0;
895
896 foreach_block (block, &shader->block_list) {
897 foreach_instr (instr, &block->instr_list) {
898 info->sizedwords += 2;
899 }
900 }
901
902 /* need an integer number of instruction "groups" (sets of 16
903 * instructions on a4xx or sets of 4 instructions on a3xx),
904 * so pad out w/ NOPs if needed: (NOTE each instruction is 64bits)
905 */
906 if (gpu_id >= 400) {
907 info->sizedwords = align(info->sizedwords, 16 * 2);
908 } else {
909 info->sizedwords = align(info->sizedwords, 4 * 2);
910 }
911
912 ptr = dwords = calloc(4, info->sizedwords);
913
914 foreach_block (block, &shader->block_list) {
915 unsigned sfu_delay = 0;
916
917 foreach_instr (instr, &block->instr_list) {
918 int ret = emit[opc_cat(instr->opc)](instr, dwords, info);
919 if (ret)
920 goto fail;
921
922 if ((instr->opc == OPC_BARY_F) && (instr->regs[0]->flags & IR3_REG_EI))
923 info->last_baryf = info->instrs_count;
924
925 info->instrs_count += 1 + instr->repeat + instr->nop;
926 info->nops_count += instr->nop;
927 if (instr->opc == OPC_NOP)
928 info->nops_count += 1 + instr->repeat;
929 dwords += 2;
930
931 if (instr->flags & IR3_INSTR_SS) {
932 info->ss++;
933 info->sstall += sfu_delay;
934 }
935
936 if (instr->flags & IR3_INSTR_SY)
937 info->sy++;
938
939 if (is_sfu(instr)) {
940 sfu_delay = 10;
941 } else if (sfu_delay > 0) {
942 sfu_delay--;
943 }
944 }
945 }
946
947 return ptr;
948
949 fail:
950 free(ptr);
951 return NULL;
952 }
953
954 static struct ir3_register * reg_create(struct ir3 *shader,
955 int num, int flags)
956 {
957 struct ir3_register *reg =
958 ir3_alloc(shader, sizeof(struct ir3_register));
959 reg->wrmask = 1;
960 reg->flags = flags;
961 reg->num = num;
962 if (shader->compiler->gpu_id >= 600)
963 reg->merged = true;
964 return reg;
965 }
966
967 static void insert_instr(struct ir3_block *block,
968 struct ir3_instruction *instr)
969 {
970 struct ir3 *shader = block->shader;
971 #ifdef DEBUG
972 instr->serialno = ++shader->instr_count;
973 #endif
974 list_addtail(&instr->node, &block->instr_list);
975
976 if (is_input(instr))
977 array_insert(shader, shader->baryfs, instr);
978 }
979
980 struct ir3_block * ir3_block_create(struct ir3 *shader)
981 {
982 struct ir3_block *block = ir3_alloc(shader, sizeof(*block));
983 #ifdef DEBUG
984 block->serialno = ++shader->block_count;
985 #endif
986 block->shader = shader;
987 list_inithead(&block->node);
988 list_inithead(&block->instr_list);
989 return block;
990 }
991
992 static struct ir3_instruction *instr_create(struct ir3_block *block, int nreg)
993 {
994 struct ir3_instruction *instr;
995 unsigned sz = sizeof(*instr) + (nreg * sizeof(instr->regs[0]));
996 char *ptr = ir3_alloc(block->shader, sz);
997
998 instr = (struct ir3_instruction *)ptr;
999 ptr += sizeof(*instr);
1000 instr->regs = (struct ir3_register **)ptr;
1001
1002 #ifdef DEBUG
1003 instr->regs_max = nreg;
1004 #endif
1005
1006 return instr;
1007 }
1008
1009 struct ir3_instruction * ir3_instr_create2(struct ir3_block *block,
1010 opc_t opc, int nreg)
1011 {
1012 struct ir3_instruction *instr = instr_create(block, nreg);
1013 instr->block = block;
1014 instr->opc = opc;
1015 insert_instr(block, instr);
1016 return instr;
1017 }
1018
1019 struct ir3_instruction * ir3_instr_create(struct ir3_block *block, opc_t opc)
1020 {
1021 /* NOTE: we could be slightly more clever, at least for non-meta,
1022 * and choose # of regs based on category.
1023 */
1024 return ir3_instr_create2(block, opc, 4);
1025 }
1026
1027 struct ir3_instruction * ir3_instr_clone(struct ir3_instruction *instr)
1028 {
1029 struct ir3_instruction *new_instr = instr_create(instr->block,
1030 instr->regs_count);
1031 struct ir3_register **regs;
1032 unsigned i;
1033
1034 regs = new_instr->regs;
1035 *new_instr = *instr;
1036 new_instr->regs = regs;
1037
1038 insert_instr(instr->block, new_instr);
1039
1040 /* clone registers: */
1041 new_instr->regs_count = 0;
1042 for (i = 0; i < instr->regs_count; i++) {
1043 struct ir3_register *reg = instr->regs[i];
1044 struct ir3_register *new_reg =
1045 ir3_reg_create(new_instr, reg->num, reg->flags);
1046 *new_reg = *reg;
1047 }
1048
1049 return new_instr;
1050 }
1051
1052 /* Add a false dependency to instruction, to ensure it is scheduled first: */
1053 void ir3_instr_add_dep(struct ir3_instruction *instr, struct ir3_instruction *dep)
1054 {
1055 array_insert(instr, instr->deps, dep);
1056 }
1057
1058 struct ir3_register * ir3_reg_create(struct ir3_instruction *instr,
1059 int num, int flags)
1060 {
1061 struct ir3 *shader = instr->block->shader;
1062 struct ir3_register *reg = reg_create(shader, num, flags);
1063 #ifdef DEBUG
1064 debug_assert(instr->regs_count < instr->regs_max);
1065 #endif
1066 instr->regs[instr->regs_count++] = reg;
1067 return reg;
1068 }
1069
1070 struct ir3_register * ir3_reg_clone(struct ir3 *shader,
1071 struct ir3_register *reg)
1072 {
1073 struct ir3_register *new_reg = reg_create(shader, 0, 0);
1074 *new_reg = *reg;
1075 return new_reg;
1076 }
1077
1078 void
1079 ir3_instr_set_address(struct ir3_instruction *instr,
1080 struct ir3_instruction *addr)
1081 {
1082 if (instr->address != addr) {
1083 struct ir3 *ir = instr->block->shader;
1084
1085 debug_assert(!instr->address);
1086 debug_assert(instr->block == addr->block);
1087
1088 instr->address = addr;
1089 debug_assert(reg_num(addr->regs[0]) == REG_A0);
1090 unsigned comp = reg_comp(addr->regs[0]);
1091 if (comp == 0) {
1092 array_insert(ir, ir->a0_users, instr);
1093 } else {
1094 debug_assert(comp == 1);
1095 array_insert(ir, ir->a1_users, instr);
1096 }
1097 }
1098 }
1099
1100 void
1101 ir3_block_clear_mark(struct ir3_block *block)
1102 {
1103 foreach_instr (instr, &block->instr_list)
1104 instr->flags &= ~IR3_INSTR_MARK;
1105 }
1106
1107 void
1108 ir3_clear_mark(struct ir3 *ir)
1109 {
1110 foreach_block (block, &ir->block_list) {
1111 ir3_block_clear_mark(block);
1112 }
1113 }
1114
1115 /* note: this will destroy instr->depth, don't do it until after sched! */
1116 unsigned
1117 ir3_count_instructions(struct ir3 *ir)
1118 {
1119 unsigned cnt = 1;
1120 foreach_block (block, &ir->block_list) {
1121 block->start_ip = cnt;
1122 block->end_ip = cnt;
1123 foreach_instr (instr, &block->instr_list) {
1124 instr->ip = cnt++;
1125 block->end_ip = instr->ip;
1126 }
1127 }
1128 return cnt;
1129 }
1130
1131 struct ir3_array *
1132 ir3_lookup_array(struct ir3 *ir, unsigned id)
1133 {
1134 foreach_array (arr, &ir->array_list)
1135 if (arr->id == id)
1136 return arr;
1137 return NULL;
1138 }
1139
1140 void
1141 ir3_find_ssa_uses(struct ir3 *ir, void *mem_ctx)
1142 {
1143 /* We could do this in a single pass if we can assume instructions
1144 * are always sorted. Which currently might not always be true.
1145 * (In particular after ir3_group pass, but maybe other places.)
1146 */
1147 foreach_block (block, &ir->block_list)
1148 foreach_instr (instr, &block->instr_list)
1149 instr->uses = NULL;
1150
1151 foreach_block (block, &ir->block_list) {
1152 foreach_instr (instr, &block->instr_list) {
1153 struct ir3_instruction *src;
1154
1155 foreach_ssa_src (src, instr) {
1156 if (!src->uses)
1157 src->uses = _mesa_pointer_set_create(mem_ctx);
1158 _mesa_set_add(src->uses, instr);
1159 }
1160 }
1161 }
1162 }