nir/lower_idiv: add new llvm-based path
[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,
49 gl_shader_stage type, unsigned nin, unsigned nout)
50 {
51 struct ir3 *shader = rzalloc(NULL, struct ir3);
52
53 shader->compiler = compiler;
54 shader->type = type;
55 shader->ninputs = nin;
56 shader->inputs = ir3_alloc(shader, sizeof(shader->inputs[0]) * nin);
57
58 shader->noutputs = nout;
59 shader->outputs = ir3_alloc(shader, sizeof(shader->outputs[0]) * nout);
60
61 list_inithead(&shader->block_list);
62 list_inithead(&shader->array_list);
63
64 return shader;
65 }
66
67 void ir3_destroy(struct ir3 *shader)
68 {
69 ralloc_free(shader);
70 }
71
72 #define iassert(cond) do { \
73 if (!(cond)) { \
74 debug_assert(cond); \
75 return -1; \
76 } } while (0)
77
78 #define iassert_type(reg, full) do { \
79 if ((full)) { \
80 iassert(!((reg)->flags & IR3_REG_HALF)); \
81 } else { \
82 iassert((reg)->flags & IR3_REG_HALF); \
83 } } while (0);
84
85 static uint32_t reg(struct ir3_register *reg, struct ir3_info *info,
86 uint32_t repeat, uint32_t valid_flags)
87 {
88 reg_t val = { .dummy32 = 0 };
89
90 if (reg->flags & ~valid_flags) {
91 debug_printf("INVALID FLAGS: %x vs %x\n",
92 reg->flags, valid_flags);
93 }
94
95 if (!(reg->flags & IR3_REG_R))
96 repeat = 0;
97
98 if (reg->flags & IR3_REG_IMMED) {
99 val.iim_val = reg->iim_val;
100 } else {
101 unsigned components;
102 int16_t max;
103
104 if (reg->flags & IR3_REG_RELATIV) {
105 components = reg->size;
106 val.idummy10 = reg->array.offset;
107 max = (reg->array.offset + repeat + components - 1);
108 } else {
109 components = util_last_bit(reg->wrmask);
110 val.comp = reg->num & 0x3;
111 val.num = reg->num >> 2;
112 max = (reg->num + repeat + components - 1);
113 }
114
115 if (reg->flags & IR3_REG_CONST) {
116 info->max_const = MAX2(info->max_const, max >> 2);
117 } else if (val.num == 63) {
118 /* ignore writes to dummy register r63.x */
119 } else if (max < regid(48, 0)) {
120 if (reg->flags & IR3_REG_HALF) {
121 if (info->gpu_id >= 600) {
122 /* starting w/ a6xx, half regs conflict with full regs: */
123 info->max_reg = MAX2(info->max_reg, max >> 3);
124 } else {
125 info->max_half_reg = MAX2(info->max_half_reg, max >> 2);
126 }
127 } else {
128 info->max_reg = MAX2(info->max_reg, max >> 2);
129 }
130 }
131 }
132
133 return val.dummy32;
134 }
135
136 static int emit_cat0(struct ir3_instruction *instr, void *ptr,
137 struct ir3_info *info)
138 {
139 instr_cat0_t *cat0 = ptr;
140
141 if (info->gpu_id >= 500) {
142 cat0->a5xx.immed = instr->cat0.immed;
143 } else if (info->gpu_id >= 400) {
144 cat0->a4xx.immed = instr->cat0.immed;
145 } else {
146 cat0->a3xx.immed = instr->cat0.immed;
147 }
148 cat0->repeat = instr->repeat;
149 cat0->ss = !!(instr->flags & IR3_INSTR_SS);
150 cat0->inv = instr->cat0.inv;
151 cat0->comp = instr->cat0.comp;
152 cat0->opc = instr->opc;
153 cat0->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
154 cat0->sync = !!(instr->flags & IR3_INSTR_SY);
155 cat0->opc_cat = 0;
156
157 return 0;
158 }
159
160 static int emit_cat1(struct ir3_instruction *instr, void *ptr,
161 struct ir3_info *info)
162 {
163 struct ir3_register *dst = instr->regs[0];
164 struct ir3_register *src = instr->regs[1];
165 instr_cat1_t *cat1 = ptr;
166
167 iassert(instr->regs_count == 2);
168 iassert_type(dst, type_size(instr->cat1.dst_type) == 32);
169 if (!(src->flags & IR3_REG_IMMED))
170 iassert_type(src, type_size(instr->cat1.src_type) == 32);
171
172 if (src->flags & IR3_REG_IMMED) {
173 cat1->iim_val = src->iim_val;
174 cat1->src_im = 1;
175 } else if (src->flags & IR3_REG_RELATIV) {
176 cat1->off = reg(src, info, instr->repeat,
177 IR3_REG_R | IR3_REG_CONST | IR3_REG_HALF | IR3_REG_RELATIV);
178 cat1->src_rel = 1;
179 cat1->src_rel_c = !!(src->flags & IR3_REG_CONST);
180 } else {
181 cat1->src = reg(src, info, instr->repeat,
182 IR3_REG_R | IR3_REG_CONST | IR3_REG_HALF);
183 cat1->src_c = !!(src->flags & IR3_REG_CONST);
184 }
185
186 cat1->dst = reg(dst, info, instr->repeat,
187 IR3_REG_RELATIV | IR3_REG_EVEN |
188 IR3_REG_R | IR3_REG_POS_INF | IR3_REG_HALF);
189 cat1->repeat = instr->repeat;
190 cat1->src_r = !!(src->flags & IR3_REG_R);
191 cat1->ss = !!(instr->flags & IR3_INSTR_SS);
192 cat1->ul = !!(instr->flags & IR3_INSTR_UL);
193 cat1->dst_type = instr->cat1.dst_type;
194 cat1->dst_rel = !!(dst->flags & IR3_REG_RELATIV);
195 cat1->src_type = instr->cat1.src_type;
196 cat1->even = !!(dst->flags & IR3_REG_EVEN);
197 cat1->pos_inf = !!(dst->flags & IR3_REG_POS_INF);
198 cat1->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
199 cat1->sync = !!(instr->flags & IR3_INSTR_SY);
200 cat1->opc_cat = 1;
201
202 return 0;
203 }
204
205 static int emit_cat2(struct ir3_instruction *instr, void *ptr,
206 struct ir3_info *info)
207 {
208 struct ir3_register *dst = instr->regs[0];
209 struct ir3_register *src1 = instr->regs[1];
210 struct ir3_register *src2 = instr->regs[2];
211 instr_cat2_t *cat2 = ptr;
212 unsigned absneg = ir3_cat2_absneg(instr->opc);
213
214 iassert((instr->regs_count == 2) || (instr->regs_count == 3));
215
216 if (instr->nop) {
217 iassert(!instr->repeat);
218 iassert(instr->nop <= 3);
219
220 cat2->src1_r = instr->nop & 0x1;
221 cat2->src2_r = (instr->nop >> 1) & 0x1;
222 } else {
223 cat2->src1_r = !!(src1->flags & IR3_REG_R);
224 if (src2)
225 cat2->src2_r = !!(src2->flags & IR3_REG_R);
226 }
227
228 if (src1->flags & IR3_REG_RELATIV) {
229 iassert(src1->array.offset < (1 << 10));
230 cat2->rel1.src1 = reg(src1, info, instr->repeat,
231 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
232 IR3_REG_HALF | absneg);
233 cat2->rel1.src1_c = !!(src1->flags & IR3_REG_CONST);
234 cat2->rel1.src1_rel = 1;
235 } else if (src1->flags & IR3_REG_CONST) {
236 iassert(src1->num < (1 << 12));
237 cat2->c1.src1 = reg(src1, info, instr->repeat,
238 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF);
239 cat2->c1.src1_c = 1;
240 } else {
241 iassert(src1->num < (1 << 11));
242 cat2->src1 = reg(src1, info, instr->repeat,
243 IR3_REG_IMMED | IR3_REG_R | IR3_REG_HALF |
244 absneg);
245 }
246 cat2->src1_im = !!(src1->flags & IR3_REG_IMMED);
247 cat2->src1_neg = !!(src1->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
248 cat2->src1_abs = !!(src1->flags & (IR3_REG_FABS | IR3_REG_SABS));
249
250 if (src2) {
251 iassert((src2->flags & IR3_REG_IMMED) ||
252 !((src1->flags ^ src2->flags) & IR3_REG_HALF));
253
254 if (src2->flags & IR3_REG_RELATIV) {
255 iassert(src2->array.offset < (1 << 10));
256 cat2->rel2.src2 = reg(src2, info, instr->repeat,
257 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
258 IR3_REG_HALF | absneg);
259 cat2->rel2.src2_c = !!(src2->flags & IR3_REG_CONST);
260 cat2->rel2.src2_rel = 1;
261 } else if (src2->flags & IR3_REG_CONST) {
262 iassert(src2->num < (1 << 12));
263 cat2->c2.src2 = reg(src2, info, instr->repeat,
264 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF);
265 cat2->c2.src2_c = 1;
266 } else {
267 iassert(src2->num < (1 << 11));
268 cat2->src2 = reg(src2, info, instr->repeat,
269 IR3_REG_IMMED | IR3_REG_R | IR3_REG_HALF |
270 absneg);
271 }
272
273 cat2->src2_im = !!(src2->flags & IR3_REG_IMMED);
274 cat2->src2_neg = !!(src2->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
275 cat2->src2_abs = !!(src2->flags & (IR3_REG_FABS | IR3_REG_SABS));
276 }
277
278 cat2->dst = reg(dst, info, instr->repeat,
279 IR3_REG_R | IR3_REG_EI | IR3_REG_HALF);
280 cat2->repeat = instr->repeat;
281 cat2->sat = !!(instr->flags & IR3_INSTR_SAT);
282 cat2->ss = !!(instr->flags & IR3_INSTR_SS);
283 cat2->ul = !!(instr->flags & IR3_INSTR_UL);
284 cat2->dst_half = !!((src1->flags ^ dst->flags) & IR3_REG_HALF);
285 cat2->ei = !!(dst->flags & IR3_REG_EI);
286 cat2->cond = instr->cat2.condition;
287 cat2->full = ! (src1->flags & IR3_REG_HALF);
288 cat2->opc = instr->opc;
289 cat2->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
290 cat2->sync = !!(instr->flags & IR3_INSTR_SY);
291 cat2->opc_cat = 2;
292
293 return 0;
294 }
295
296 static int emit_cat3(struct ir3_instruction *instr, void *ptr,
297 struct ir3_info *info)
298 {
299 struct ir3_register *dst = instr->regs[0];
300 struct ir3_register *src1 = instr->regs[1];
301 struct ir3_register *src2 = instr->regs[2];
302 struct ir3_register *src3 = instr->regs[3];
303 unsigned absneg = ir3_cat3_absneg(instr->opc);
304 instr_cat3_t *cat3 = ptr;
305 uint32_t src_flags = 0;
306
307 switch (instr->opc) {
308 case OPC_MAD_F16:
309 case OPC_MAD_U16:
310 case OPC_MAD_S16:
311 case OPC_SEL_B16:
312 case OPC_SEL_S16:
313 case OPC_SEL_F16:
314 case OPC_SAD_S16:
315 case OPC_SAD_S32: // really??
316 src_flags |= IR3_REG_HALF;
317 break;
318 default:
319 break;
320 }
321
322 iassert(instr->regs_count == 4);
323 iassert(!((src1->flags ^ src_flags) & IR3_REG_HALF));
324 iassert(!((src2->flags ^ src_flags) & IR3_REG_HALF));
325 iassert(!((src3->flags ^ src_flags) & IR3_REG_HALF));
326
327 if (instr->nop) {
328 iassert(!instr->repeat);
329 iassert(instr->nop <= 3);
330
331 cat3->src1_r = instr->nop & 0x1;
332 cat3->src2_r = (instr->nop >> 1) & 0x1;
333 } else {
334 cat3->src1_r = !!(src1->flags & IR3_REG_R);
335 cat3->src2_r = !!(src2->flags & IR3_REG_R);
336 }
337
338 if (src1->flags & IR3_REG_RELATIV) {
339 iassert(src1->array.offset < (1 << 10));
340 cat3->rel1.src1 = reg(src1, info, instr->repeat,
341 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
342 IR3_REG_HALF | absneg);
343 cat3->rel1.src1_c = !!(src1->flags & IR3_REG_CONST);
344 cat3->rel1.src1_rel = 1;
345 } else if (src1->flags & IR3_REG_CONST) {
346 iassert(src1->num < (1 << 12));
347 cat3->c1.src1 = reg(src1, info, instr->repeat,
348 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF);
349 cat3->c1.src1_c = 1;
350 } else {
351 iassert(src1->num < (1 << 11));
352 cat3->src1 = reg(src1, info, instr->repeat,
353 IR3_REG_R | IR3_REG_HALF | absneg);
354 }
355
356 cat3->src1_neg = !!(src1->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
357
358 cat3->src2 = reg(src2, info, instr->repeat,
359 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF | absneg);
360 cat3->src2_c = !!(src2->flags & IR3_REG_CONST);
361 cat3->src2_neg = !!(src2->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
362
363 if (src3->flags & IR3_REG_RELATIV) {
364 iassert(src3->array.offset < (1 << 10));
365 cat3->rel2.src3 = reg(src3, info, instr->repeat,
366 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_R |
367 IR3_REG_HALF | absneg);
368 cat3->rel2.src3_c = !!(src3->flags & IR3_REG_CONST);
369 cat3->rel2.src3_rel = 1;
370 } else if (src3->flags & IR3_REG_CONST) {
371 iassert(src3->num < (1 << 12));
372 cat3->c2.src3 = reg(src3, info, instr->repeat,
373 IR3_REG_CONST | IR3_REG_R | IR3_REG_HALF);
374 cat3->c2.src3_c = 1;
375 } else {
376 iassert(src3->num < (1 << 11));
377 cat3->src3 = reg(src3, info, instr->repeat,
378 IR3_REG_R | IR3_REG_HALF | absneg);
379 }
380
381 cat3->src3_neg = !!(src3->flags & (IR3_REG_FNEG | IR3_REG_SNEG | IR3_REG_BNOT));
382 cat3->src3_r = !!(src3->flags & IR3_REG_R);
383
384 cat3->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
385 cat3->repeat = instr->repeat;
386 cat3->sat = !!(instr->flags & IR3_INSTR_SAT);
387 cat3->ss = !!(instr->flags & IR3_INSTR_SS);
388 cat3->ul = !!(instr->flags & IR3_INSTR_UL);
389 cat3->dst_half = !!((src_flags ^ dst->flags) & IR3_REG_HALF);
390 cat3->opc = instr->opc;
391 cat3->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
392 cat3->sync = !!(instr->flags & IR3_INSTR_SY);
393 cat3->opc_cat = 3;
394
395 return 0;
396 }
397
398 static int emit_cat4(struct ir3_instruction *instr, void *ptr,
399 struct ir3_info *info)
400 {
401 struct ir3_register *dst = instr->regs[0];
402 struct ir3_register *src = instr->regs[1];
403 instr_cat4_t *cat4 = ptr;
404
405 iassert(instr->regs_count == 2);
406
407 if (src->flags & IR3_REG_RELATIV) {
408 iassert(src->array.offset < (1 << 10));
409 cat4->rel.src = reg(src, info, instr->repeat,
410 IR3_REG_RELATIV | IR3_REG_CONST | IR3_REG_FNEG |
411 IR3_REG_FABS | IR3_REG_R | IR3_REG_HALF);
412 cat4->rel.src_c = !!(src->flags & IR3_REG_CONST);
413 cat4->rel.src_rel = 1;
414 } else if (src->flags & IR3_REG_CONST) {
415 iassert(src->num < (1 << 12));
416 cat4->c.src = reg(src, info, instr->repeat,
417 IR3_REG_CONST | IR3_REG_FNEG | IR3_REG_FABS |
418 IR3_REG_R | IR3_REG_HALF);
419 cat4->c.src_c = 1;
420 } else {
421 iassert(src->num < (1 << 11));
422 cat4->src = reg(src, info, instr->repeat,
423 IR3_REG_IMMED | IR3_REG_FNEG | IR3_REG_FABS |
424 IR3_REG_R | IR3_REG_HALF);
425 }
426
427 cat4->src_im = !!(src->flags & IR3_REG_IMMED);
428 cat4->src_neg = !!(src->flags & IR3_REG_FNEG);
429 cat4->src_abs = !!(src->flags & IR3_REG_FABS);
430 cat4->src_r = !!(src->flags & IR3_REG_R);
431
432 cat4->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
433 cat4->repeat = instr->repeat;
434 cat4->sat = !!(instr->flags & IR3_INSTR_SAT);
435 cat4->ss = !!(instr->flags & IR3_INSTR_SS);
436 cat4->ul = !!(instr->flags & IR3_INSTR_UL);
437 cat4->dst_half = !!((src->flags ^ dst->flags) & IR3_REG_HALF);
438 cat4->full = ! (src->flags & IR3_REG_HALF);
439 cat4->opc = instr->opc;
440 cat4->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
441 cat4->sync = !!(instr->flags & IR3_INSTR_SY);
442 cat4->opc_cat = 4;
443
444 return 0;
445 }
446
447 static int emit_cat5(struct ir3_instruction *instr, void *ptr,
448 struct ir3_info *info)
449 {
450 struct ir3_register *dst = instr->regs[0];
451 /* To simplify things when there could be zero, one, or two args other
452 * than tex/sampler idx, we use the first src reg in the ir to hold
453 * samp_tex hvec2:
454 */
455 struct ir3_register *src1;
456 struct ir3_register *src2;
457 instr_cat5_t *cat5 = ptr;
458
459 iassert((instr->regs_count == 2) ||
460 (instr->regs_count == 3) || (instr->regs_count == 4));
461
462 switch (instr->opc) {
463 case OPC_DSX:
464 case OPC_DSXPP_1:
465 case OPC_DSY:
466 case OPC_DSYPP_1:
467 case OPC_RGETPOS:
468 case OPC_RGETINFO:
469 iassert((instr->flags & IR3_INSTR_S2EN) == 0);
470 src1 = instr->regs[1];
471 src2 = instr->regs_count > 2 ? instr->regs[2] : NULL;
472 break;
473 default:
474 src1 = instr->regs[2];
475 src2 = instr->regs_count > 3 ? instr->regs[3] : NULL;
476 break;
477 }
478
479 assume(src1 || !src2);
480
481 if (src1) {
482 cat5->full = ! (src1->flags & IR3_REG_HALF);
483 cat5->src1 = reg(src1, info, instr->repeat, IR3_REG_HALF);
484 }
485
486 if (instr->flags & IR3_INSTR_S2EN) {
487 struct ir3_register *samp_tex = instr->regs[1];
488 if (src2) {
489 iassert(!((src1->flags ^ src2->flags) & IR3_REG_HALF));
490 cat5->s2en.src2 = reg(src2, info, instr->repeat, IR3_REG_HALF);
491 }
492 iassert(samp_tex->flags & IR3_REG_HALF);
493 cat5->s2en.src3 = reg(samp_tex, info, instr->repeat, IR3_REG_HALF);
494 iassert(!(instr->cat5.samp | instr->cat5.tex));
495 } else {
496 if (src2) {
497 iassert(!((src1->flags ^ src2->flags) & IR3_REG_HALF));
498 cat5->norm.src2 = reg(src2, info, instr->repeat, IR3_REG_HALF);
499 }
500 cat5->norm.samp = instr->cat5.samp;
501 cat5->norm.tex = instr->cat5.tex;
502 }
503
504 cat5->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
505 cat5->wrmask = dst->wrmask;
506 cat5->type = instr->cat5.type;
507 cat5->is_3d = !!(instr->flags & IR3_INSTR_3D);
508 cat5->is_a = !!(instr->flags & IR3_INSTR_A);
509 cat5->is_s = !!(instr->flags & IR3_INSTR_S);
510 cat5->is_s2en = !!(instr->flags & IR3_INSTR_S2EN);
511 cat5->is_o = !!(instr->flags & IR3_INSTR_O);
512 cat5->is_p = !!(instr->flags & IR3_INSTR_P);
513 cat5->opc = instr->opc;
514 cat5->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
515 cat5->sync = !!(instr->flags & IR3_INSTR_SY);
516 cat5->opc_cat = 5;
517
518 return 0;
519 }
520
521 static int emit_cat6_a6xx(struct ir3_instruction *instr, void *ptr,
522 struct ir3_info *info)
523 {
524 struct ir3_register *src1, *src2;
525 instr_cat6_a6xx_t *cat6 = ptr;
526 bool has_dest = (instr->opc == OPC_LDIB);
527
528 /* first reg should be SSBO binding point: */
529 iassert(instr->regs[1]->flags & IR3_REG_IMMED);
530
531 src1 = instr->regs[2];
532
533 if (has_dest) {
534 /* the src2 field in the instruction is actually the destination
535 * register for load instructions:
536 */
537 src2 = instr->regs[0];
538 } else {
539 src2 = instr->regs[3];
540 }
541
542 cat6->type = instr->cat6.type;
543 cat6->d = instr->cat6.d - 1;
544 cat6->typed = instr->cat6.typed;
545 cat6->type_size = instr->cat6.iim_val - 1;
546 cat6->opc = instr->opc;
547 cat6->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
548 cat6->sync = !!(instr->flags & IR3_INSTR_SY);
549 cat6->opc_cat = 6;
550
551 cat6->src1 = reg(src1, info, instr->repeat, 0);
552 cat6->src2 = reg(src2, info, instr->repeat, 0);
553 cat6->ssbo = instr->regs[1]->iim_val;
554
555 switch (instr->opc) {
556 case OPC_ATOMIC_ADD:
557 case OPC_ATOMIC_SUB:
558 case OPC_ATOMIC_XCHG:
559 case OPC_ATOMIC_INC:
560 case OPC_ATOMIC_DEC:
561 case OPC_ATOMIC_CMPXCHG:
562 case OPC_ATOMIC_MIN:
563 case OPC_ATOMIC_MAX:
564 case OPC_ATOMIC_AND:
565 case OPC_ATOMIC_OR:
566 case OPC_ATOMIC_XOR:
567 cat6->pad1 = 0x1;
568 cat6->pad2 = 0xc;
569 cat6->pad3 = 0x0;
570 cat6->pad4 = 0x3;
571 break;
572 case OPC_STIB:
573 cat6->pad1 = 0x0;
574 cat6->pad2 = 0xc;
575 cat6->pad3 = 0x0;
576 cat6->pad4 = 0x2;
577 break;
578 case OPC_LDIB:
579 cat6->pad1 = 0x1;
580 cat6->pad2 = 0xc;
581 cat6->pad3 = 0x0;
582 cat6->pad4 = 0x2;
583 break;
584 case OPC_LDC:
585 cat6->pad1 = 0x0;
586 cat6->pad2 = 0x8;
587 cat6->pad3 = 0x0;
588 cat6->pad4 = 0x2;
589 break;
590 default:
591 iassert(0);
592 }
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 cat6a->src1 = reg(src1, info, instr->repeat, IR3_REG_IMMED);
803 cat6a->src1_im = !!(src1->flags & IR3_REG_IMMED);
804
805 /* Num components */
806 cat6a->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
807 cat6a->src2_im = true;
808
809 /* Offset */
810 iassert(src3->flags & IR3_REG_IMMED);
811 cat6a->off = reg(src3, info, instr->repeat, IR3_REG_IMMED);
812 } else {
813 instr_cat6b_t *cat6b = ptr;
814
815 cat6->src_off = false;
816
817 cat6b->src1 = reg(src1, info, instr->repeat, IR3_REG_IMMED | IR3_REG_HALF);
818 cat6b->src1_im = !!(src1->flags & IR3_REG_IMMED);
819 if (src2) {
820 cat6b->src2 = reg(src2, info, instr->repeat, IR3_REG_IMMED);
821 cat6b->src2_im = !!(src2->flags & IR3_REG_IMMED);
822 }
823 }
824
825 if (instr->cat6.dst_offset || (instr->opc == OPC_STG) ||
826 (instr->opc == OPC_STL) || (instr->opc == OPC_STLW)) {
827 instr_cat6c_t *cat6c = ptr;
828 cat6->dst_off = true;
829 cat6c->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
830 cat6c->off = instr->cat6.dst_offset;
831 } else {
832 instr_cat6d_t *cat6d = ptr;
833 cat6->dst_off = false;
834 cat6d->dst = reg(dst, info, instr->repeat, IR3_REG_R | IR3_REG_HALF);
835 }
836
837 return 0;
838 }
839
840 static int emit_cat7(struct ir3_instruction *instr, void *ptr,
841 struct ir3_info *info)
842 {
843 instr_cat7_t *cat7 = ptr;
844
845 cat7->ss = !!(instr->flags & IR3_INSTR_SS);
846 cat7->w = instr->cat7.w;
847 cat7->r = instr->cat7.r;
848 cat7->l = instr->cat7.l;
849 cat7->g = instr->cat7.g;
850 cat7->opc = instr->opc;
851 cat7->jmp_tgt = !!(instr->flags & IR3_INSTR_JP);
852 cat7->sync = !!(instr->flags & IR3_INSTR_SY);
853 cat7->opc_cat = 7;
854
855 return 0;
856 }
857
858 static int (*emit[])(struct ir3_instruction *instr, void *ptr,
859 struct ir3_info *info) = {
860 emit_cat0, emit_cat1, emit_cat2, emit_cat3, emit_cat4, emit_cat5, emit_cat6,
861 emit_cat7,
862 };
863
864 void * ir3_assemble(struct ir3 *shader, struct ir3_info *info,
865 uint32_t gpu_id)
866 {
867 uint32_t *ptr, *dwords;
868
869 info->gpu_id = gpu_id;
870 info->max_reg = -1;
871 info->max_half_reg = -1;
872 info->max_const = -1;
873 info->instrs_count = 0;
874 info->sizedwords = 0;
875 info->ss = info->sy = 0;
876
877 list_for_each_entry (struct ir3_block, block, &shader->block_list, node) {
878 list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
879 info->sizedwords += 2;
880 }
881 }
882
883 /* need an integer number of instruction "groups" (sets of 16
884 * instructions on a4xx or sets of 4 instructions on a3xx),
885 * so pad out w/ NOPs if needed: (NOTE each instruction is 64bits)
886 */
887 if (gpu_id >= 400) {
888 info->sizedwords = align(info->sizedwords, 16 * 2);
889 } else {
890 info->sizedwords = align(info->sizedwords, 4 * 2);
891 }
892
893 ptr = dwords = calloc(4, info->sizedwords);
894
895 list_for_each_entry (struct ir3_block, block, &shader->block_list, node) {
896 list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
897 int ret = emit[opc_cat(instr->opc)](instr, dwords, info);
898 if (ret)
899 goto fail;
900 info->instrs_count += 1 + instr->repeat + instr->nop;
901 dwords += 2;
902
903 if (instr->flags & IR3_INSTR_SS)
904 info->ss++;
905
906 if (instr->flags & IR3_INSTR_SY)
907 info->sy++;
908 }
909 }
910
911 return ptr;
912
913 fail:
914 free(ptr);
915 return NULL;
916 }
917
918 static struct ir3_register * reg_create(struct ir3 *shader,
919 int num, int flags)
920 {
921 struct ir3_register *reg =
922 ir3_alloc(shader, sizeof(struct ir3_register));
923 reg->wrmask = 1;
924 reg->flags = flags;
925 reg->num = num;
926 if (shader->compiler->gpu_id >= 600)
927 reg->merged = true;
928 return reg;
929 }
930
931 static void insert_instr(struct ir3_block *block,
932 struct ir3_instruction *instr)
933 {
934 struct ir3 *shader = block->shader;
935 #ifdef DEBUG
936 instr->serialno = ++shader->instr_count;
937 #endif
938 list_addtail(&instr->node, &block->instr_list);
939
940 if (is_input(instr))
941 array_insert(shader, shader->baryfs, instr);
942 }
943
944 struct ir3_block * ir3_block_create(struct ir3 *shader)
945 {
946 struct ir3_block *block = ir3_alloc(shader, sizeof(*block));
947 #ifdef DEBUG
948 block->serialno = ++shader->block_count;
949 #endif
950 block->shader = shader;
951 list_inithead(&block->node);
952 list_inithead(&block->instr_list);
953 return block;
954 }
955
956 static struct ir3_instruction *instr_create(struct ir3_block *block, int nreg)
957 {
958 struct ir3_instruction *instr;
959 unsigned sz = sizeof(*instr) + (nreg * sizeof(instr->regs[0]));
960 char *ptr = ir3_alloc(block->shader, sz);
961
962 instr = (struct ir3_instruction *)ptr;
963 ptr += sizeof(*instr);
964 instr->regs = (struct ir3_register **)ptr;
965
966 #ifdef DEBUG
967 instr->regs_max = nreg;
968 #endif
969
970 return instr;
971 }
972
973 struct ir3_instruction * ir3_instr_create2(struct ir3_block *block,
974 opc_t opc, int nreg)
975 {
976 struct ir3_instruction *instr = instr_create(block, nreg);
977 instr->block = block;
978 instr->opc = opc;
979 insert_instr(block, instr);
980 return instr;
981 }
982
983 struct ir3_instruction * ir3_instr_create(struct ir3_block *block, opc_t opc)
984 {
985 /* NOTE: we could be slightly more clever, at least for non-meta,
986 * and choose # of regs based on category.
987 */
988 return ir3_instr_create2(block, opc, 4);
989 }
990
991 struct ir3_instruction * ir3_instr_clone(struct ir3_instruction *instr)
992 {
993 struct ir3_instruction *new_instr = instr_create(instr->block,
994 instr->regs_count);
995 struct ir3_register **regs;
996 unsigned i;
997
998 regs = new_instr->regs;
999 *new_instr = *instr;
1000 new_instr->regs = regs;
1001
1002 insert_instr(instr->block, new_instr);
1003
1004 /* clone registers: */
1005 new_instr->regs_count = 0;
1006 for (i = 0; i < instr->regs_count; i++) {
1007 struct ir3_register *reg = instr->regs[i];
1008 struct ir3_register *new_reg =
1009 ir3_reg_create(new_instr, reg->num, reg->flags);
1010 *new_reg = *reg;
1011 }
1012
1013 return new_instr;
1014 }
1015
1016 /* Add a false dependency to instruction, to ensure it is scheduled first: */
1017 void ir3_instr_add_dep(struct ir3_instruction *instr, struct ir3_instruction *dep)
1018 {
1019 array_insert(instr, instr->deps, dep);
1020 }
1021
1022 struct ir3_register * ir3_reg_create(struct ir3_instruction *instr,
1023 int num, int flags)
1024 {
1025 struct ir3 *shader = instr->block->shader;
1026 struct ir3_register *reg = reg_create(shader, num, flags);
1027 #ifdef DEBUG
1028 debug_assert(instr->regs_count < instr->regs_max);
1029 #endif
1030 instr->regs[instr->regs_count++] = reg;
1031 return reg;
1032 }
1033
1034 struct ir3_register * ir3_reg_clone(struct ir3 *shader,
1035 struct ir3_register *reg)
1036 {
1037 struct ir3_register *new_reg = reg_create(shader, 0, 0);
1038 *new_reg = *reg;
1039 return new_reg;
1040 }
1041
1042 void
1043 ir3_instr_set_address(struct ir3_instruction *instr,
1044 struct ir3_instruction *addr)
1045 {
1046 if (instr->address != addr) {
1047 struct ir3 *ir = instr->block->shader;
1048
1049 debug_assert(!instr->address);
1050 debug_assert(instr->block == addr->block);
1051
1052 instr->address = addr;
1053 array_insert(ir, ir->indirects, instr);
1054 }
1055 }
1056
1057 void
1058 ir3_block_clear_mark(struct ir3_block *block)
1059 {
1060 list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node)
1061 instr->flags &= ~IR3_INSTR_MARK;
1062 }
1063
1064 void
1065 ir3_clear_mark(struct ir3 *ir)
1066 {
1067 list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
1068 ir3_block_clear_mark(block);
1069 }
1070 }
1071
1072 /* note: this will destroy instr->depth, don't do it until after sched! */
1073 unsigned
1074 ir3_count_instructions(struct ir3 *ir)
1075 {
1076 unsigned cnt = 0;
1077 list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
1078 block->start_ip = cnt;
1079 block->end_ip = cnt;
1080 list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
1081 instr->ip = cnt++;
1082 block->end_ip = instr->ip;
1083 }
1084 }
1085 return cnt;
1086 }
1087
1088 struct ir3_array *
1089 ir3_lookup_array(struct ir3 *ir, unsigned id)
1090 {
1091 list_for_each_entry (struct ir3_array, arr, &ir->array_list, node)
1092 if (arr->id == id)
1093 return arr;
1094 return NULL;
1095 }