projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: tell LLVM not to remove s_barrier instructions
[mesa.git] / src / gallium / drivers / r600 / sb / sb_gcm.cpp
1 /*
2 * Copyright 2013 Vadim Girlin <vadimgirlin@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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 * Vadim Girlin
25 */
26
27 #define GCM_DEBUG 0
28
29 #if GCM_DEBUG
30 #define GCM_DUMP(a) do { a } while(0);
31 #else
32 #define GCM_DUMP(a)
33 #endif
34
35 #include <map>
36
37 #include "sb_bc.h"
38 #include "sb_shader.h"
39 #include "sb_pass.h"
40 #include "eg_sq.h" // V_SQ_CF_INDEX_NONE
41
42 namespace r600_sb {
43
44 int gcm::run() {
45
46 GCM_DUMP( sblog << "==== GCM ==== \n"; sh.dump_ir(); );
47
48 collect_instructions(sh.root, true);
49
50 init_def_count(uses, pending);
51
52 for (node_iterator N, I = pending.begin(), E = pending.end();
53 I != E; I = N) {
54 N = I;
55 ++N;
56 node *o = *I;
57
58 GCM_DUMP(
59 sblog << "pending : ";
60 dump::dump_op(o);
61 sblog << "\n";
62 );
63
64 if (td_is_ready(o)) {
65
66 GCM_DUMP(
67 sblog << " ready: ";
68 dump::dump_op(o);
69 sblog << "\n";
70 );
71 pending.remove_node(o);
72 ready.push_back(o);
73 } else {
74 }
75 }
76
77 sched_early(sh.root);
78
79 if (!pending.empty()) {
80 sblog << "##### gcm_sched_early_pass: unscheduled ops:\n";
81 dump::dump_op(pending.front());
82 }
83
84 assert(pending.empty());
85
86 GCM_DUMP( sh.dump_ir(); );
87
88 GCM_DUMP( sblog << "\n\n ############## gcm late\n\n"; );
89
90 collect_instructions(sh.root, false);
91
92 init_use_count(uses, pending);
93
94 sched_late(sh.root);
95 if (!pending.empty()) {
96 sblog << "##### gcm_sched_late_pass: unscheduled ops:\n";
97 dump::dump_op(pending.front());
98 }
99
100 assert(ucs_level == 0);
101 assert(pending.empty());
102
103 return 0;
104 }
105
106
107 void gcm::collect_instructions(container_node *c, bool early_pass) {
108 if (c->is_bb()) {
109
110 if (early_pass) {
111 for (node_iterator I = c->begin(), E = c->end(); I != E; ++I) {
112 node *n = *I;
113 if (n->flags & NF_DONT_MOVE) {
114 op_info &o = op_map[n];
115 o.top_bb = o.bottom_bb = static_cast<bb_node*>(c);
116 }
117 }
118 }
119
120 pending.append_from(c);
121 return;
122 }
123
124 for (node_iterator I = c->begin(), E = c->end(); I != E; ++I) {
125 if (I->is_container()) {
126 collect_instructions(static_cast<container_node*>(*I), early_pass);
127 }
128 }
129 }
130
131 void gcm::sched_early(container_node *n) {
132
133 region_node *r =
134 (n->type == NT_REGION) ? static_cast<region_node*>(n) : NULL;
135
136 if (r && r->loop_phi) {
137 sched_early(r->loop_phi);
138 }
139
140 for (node_iterator I = n->begin(), E = n->end(); I != E; ++I) {
141 if (I->type == NT_OP) {
142 node *op = *I;
143 if (op->subtype == NST_PHI) {
144 td_release_uses(op->dst);
145 }
146 } else if (I->is_container()) {
147 if (I->subtype == NST_BB) {
148 bb_node* bb = static_cast<bb_node*>(*I);
149 td_sched_bb(bb);
150 } else {
151 sched_early(static_cast<container_node*>(*I));
152 }
153 }
154 }
155
156 if (r && r->phi) {
157 sched_early(r->phi);
158 }
159 }
160
161 void gcm::td_schedule(bb_node *bb, node *n) {
162 GCM_DUMP(
163 sblog << "scheduling : ";
164 dump::dump_op(n);
165 sblog << "\n";
166 );
167 td_release_uses(n->dst);
168
169 bb->push_back(n);
170
171 op_map[n].top_bb = bb;
172
173 }
174
175 void gcm::td_sched_bb(bb_node* bb) {
176 GCM_DUMP(
177 sblog << "td scheduling BB_" << bb->id << "\n";
178 );
179
180 while (!ready.empty()) {
181 for (sq_iterator N, I = ready.begin(), E = ready.end(); I != E;
182 I = N) {
183 N = I; ++N;
184 td_schedule(bb, *I);
185 ready.erase(I);
186 }
187 }
188 }
189
190 bool gcm::td_is_ready(node* n) {
191 return uses[n] == 0;
192 }
193
194 void gcm::td_release_val(value *v) {
195
196 GCM_DUMP(
197 sblog << "td checking uses: ";
198 dump::dump_val(v);
199 sblog << "\n";
200 );
201
202 for (uselist::iterator I = v->uses.begin(), E = v->uses.end(); I != E; ++I) {
203 node *op = *I;
204 if (op->parent != &pending) {
205 continue;
206 }
207
208 GCM_DUMP(
209 sblog << "td used in ";
210 dump::dump_op(op);
211 sblog << "\n";
212 );
213
214 assert(uses[op] > 0);
215 if (--uses[op] == 0) {
216 GCM_DUMP(
217 sblog << "td released : ";
218 dump::dump_op(op);
219 sblog << "\n";
220 );
221
222 pending.remove_node(op);
223 ready.push_back(op);
224 }
225 }
226
227 }
228
229 void gcm::td_release_uses(vvec& v) {
230 for (vvec::iterator I = v.begin(), E = v.end(); I != E; ++I) {
231 value *v = *I;
232 if (!v)
233 continue;
234
235 if (v->is_rel())
236 td_release_uses(v->mdef);
237 else
238 td_release_val(v);
239 }
240 }
241
242 void gcm::sched_late(container_node *n) {
243
244 bool stack_pushed = false;
245
246 if (n->is_depart()) {
247 depart_node *d = static_cast<depart_node*>(n);
248 push_uc_stack();
249 stack_pushed = true;
250 bu_release_phi_defs(d->target->phi, d->dep_id);
251 } else if (n->is_repeat()) {
252 repeat_node *r = static_cast<repeat_node*>(n);
253 assert(r->target->loop_phi);
254 push_uc_stack();
255 stack_pushed = true;
256 bu_release_phi_defs(r->target->loop_phi, r->rep_id);
257 }
258
259 for (node_riterator I = n->rbegin(), E = n->rend(); I != E; ++I) {
260 if (I->is_container()) {
261 if (I->subtype == NST_BB) {
262 bb_node* bb = static_cast<bb_node*>(*I);
263 bu_sched_bb(bb);
264 } else {
265 sched_late(static_cast<container_node*>(*I));
266 }
267 }
268 }
269
270 if (n->type == NT_IF) {
271 if_node *f = static_cast<if_node*>(n);
272 if (f->cond)
273 pending_defs.push_back(f->cond);
274 } else if (n->type == NT_REGION) {
275 region_node *r = static_cast<region_node*>(n);
276 if (r->loop_phi)
277 bu_release_phi_defs(r->loop_phi, 0);
278 }
279
280 if (stack_pushed)
281 pop_uc_stack();
282
283 }
284
285 void gcm::bu_sched_bb(bb_node* bb) {
286 GCM_DUMP(
287 sblog << "bu scheduling BB_" << bb->id << "\n";
288 );
289
290 bu_bb = bb;
291
292 if (!pending_nodes.empty()) {
293 GCM_DUMP(
294 sblog << "pending nodes:\n";
295 );
296
297 // TODO consider sorting the exports by array_base,
298 // possibly it can improve performance
299
300 for (node_list::iterator I = pending_nodes.begin(),
301 E = pending_nodes.end(); I != E; ++I) {
302 bu_release_op(*I);
303 }
304 pending_nodes.clear();
305 GCM_DUMP(
306 sblog << "pending nodes processed...\n";
307 );
308 }
309
310
311 if (!pending_defs.empty()) {
312 for (vvec::iterator I = pending_defs.begin(), E = pending_defs.end();
313 I != E; ++I) {
314 bu_release_val(*I);
315 }
316 pending_defs.clear();
317 }
318
319 for (sched_queue::iterator N, I = ready_above.begin(), E = ready_above.end();
320 I != E; I = N) {
321 N = I;
322 ++N;
323 node *n = *I;
324 if (op_map[n].bottom_bb == bb) {
325 add_ready(*I);
326 ready_above.erase(I);
327 }
328 }
329
330 unsigned cnt_ready[SQ_NUM];
331
332 container_node *clause = NULL;
333 unsigned last_inst_type = ~0;
334 unsigned last_count = 0;
335
336 bool s = true;
337 while (s) {
338 node *n;
339
340 s = false;
341
342 unsigned ready_mask = 0;
343
344 for (unsigned sq = SQ_CF; sq < SQ_NUM; ++sq) {
345 if (!bu_ready[sq].empty() || !bu_ready_next[sq].empty())
346 ready_mask |= (1 << sq);
347 }
348
349 if (!ready_mask) {
350 for (unsigned sq = SQ_CF; sq < SQ_NUM; ++sq) {
351 if (!bu_ready_early[sq].empty()) {
352 node *n = bu_ready_early[sq].front();
353 bu_ready_early[sq].pop_front();
354 bu_ready[sq].push_back(n);
355 break;
356 }
357 }
358 }
359
360 for (unsigned sq = SQ_CF; sq < SQ_NUM; ++sq) {
361
362 if (sq == SQ_CF && pending_exec_mask_update) {
363 pending_exec_mask_update = false;
364 sq = SQ_ALU;
365 --sq;
366 continue;
367 }
368
369 if (!bu_ready_next[sq].empty())
370 bu_ready[sq].splice(bu_ready[sq].end(), bu_ready_next[sq]);
371
372 cnt_ready[sq] = bu_ready[sq].size();
373
374 if ((sq == SQ_TEX || sq == SQ_VTX) && live_count <= rp_threshold &&
375 cnt_ready[sq] < ctx.max_fetch/2 &&
376 !bu_ready_next[SQ_ALU].empty()) {
377 sq = SQ_ALU;
378 --sq;
379 continue;
380 }
381
382 while (!bu_ready[sq].empty()) {
383
384 if (last_inst_type != sq) {
385 clause = NULL;
386 last_count = 0;
387 last_inst_type = sq;
388 }
389
390 // simple heuristic to limit register pressure,
391 if (sq == SQ_ALU && live_count > rp_threshold &&
392 (!bu_ready[SQ_TEX].empty() ||
393 !bu_ready[SQ_VTX].empty() ||
394 !bu_ready_next[SQ_TEX].empty() ||
395 !bu_ready_next[SQ_VTX].empty())) {
396 GCM_DUMP( sblog << "switching to fetch (regpressure)\n"; );
397 break;
398 }
399
400 n = bu_ready[sq].front();
401
402 // real count (e.g. SAMPLE_G will be expanded to 3 instructions,
403 // 2 SET_GRAD_ + 1 SAMPLE_G
404 unsigned ncnt = 1;
405 if (n->is_fetch_inst() && n->src.size() == 12) {
406 ncnt = 3;
407 }
408
409 bool sampler_indexing = false;
410 if (n->is_fetch_inst() &&
411 static_cast<fetch_node *>(n)->bc.sampler_index_mode != V_SQ_CF_INDEX_NONE)
412 {
413 sampler_indexing = true; // Give sampler indexed ops get their own clause
414 ncnt = sh.get_ctx().is_cayman() ? 2 : 3; // MOVA + SET_CF_IDX0/1
415 }
416
417 if ((sq == SQ_TEX || sq == SQ_VTX) &&
418 ((last_count >= ctx.max_fetch/2 &&
419 check_alu_ready_count(24)) ||
420 last_count + ncnt > ctx.max_fetch))
421 break;
422 else if (sq == SQ_CF && last_count > 4 &&
423 check_alu_ready_count(24))
424 break;
425
426 bu_ready[sq].pop_front();
427
428 if (sq != SQ_CF) {
429 if (!clause || sampler_indexing) {
430 clause = sh.create_clause(sq == SQ_ALU ?
431 NST_ALU_CLAUSE :
432 sq == SQ_TEX ? NST_TEX_CLAUSE :
433 NST_VTX_CLAUSE);
434 bb->push_front(clause);
435 }
436 } else {
437 clause = bb;
438 }
439
440 bu_schedule(clause, n);
441 s = true;
442 last_count += ncnt;
443 }
444 }
445 }
446
447 bu_bb = NULL;
448
449 GCM_DUMP(
450 sblog << "bu finished scheduling BB_" << bb->id << "\n";
451 );
452 }
453
454 void gcm::bu_release_defs(vvec& v, bool src) {
455 for (vvec::reverse_iterator I = v.rbegin(), E = v.rend(); I != E; ++I) {
456 value *v = *I;
457 if (!v || v->is_readonly())
458 continue;
459
460 if (v->is_rel()) {
461 if (!v->rel->is_readonly())
462 bu_release_val(v->rel);
463 bu_release_defs(v->muse, true);
464 } else if (src)
465 bu_release_val(v);
466 else {
467 if (live.remove_val(v)) {
468 --live_count;
469 }
470 }
471 }
472 }
473
474 void gcm::push_uc_stack() {
475 GCM_DUMP(
476 sblog << "pushing use count stack prev_level " << ucs_level
477 << " new level " << (ucs_level + 1) << "\n";
478 );
479 ++ucs_level;
480 if (ucs_level == nuc_stk.size()) {
481 nuc_stk.resize(ucs_level + 1);
482 }
483 else {
484 nuc_stk[ucs_level].clear();
485 }
486 }
487
488 bool gcm::bu_is_ready(node* n) {
489 nuc_map &cm = nuc_stk[ucs_level];
490 nuc_map::iterator F = cm.find(n);
491 unsigned uc = (F == cm.end() ? 0 : F->second);
492 return uc == uses[n];
493 }
494
495 void gcm::bu_schedule(container_node* c, node* n) {
496 GCM_DUMP(
497 sblog << "bu scheduling : ";
498 dump::dump_op(n);
499 sblog << "\n";
500 );
501
502 assert(op_map[n].bottom_bb == bu_bb);
503
504 bu_release_defs(n->src, true);
505 bu_release_defs(n->dst, false);
506
507 c->push_front(n);
508 }
509
510 void gcm::dump_uc_stack() {
511 sblog << "##### uc_stk start ####\n";
512 for (unsigned l = 0; l <= ucs_level; ++l) {
513 nuc_map &m = nuc_stk[l];
514
515 sblog << "nuc_stk[" << l << "] : @" << &m << "\n";
516
517 for (nuc_map::iterator I = m.begin(), E = m.end(); I != E; ++I) {
518 sblog << " uc " << I->second << " for ";
519 dump::dump_op(I->first);
520 sblog << "\n";
521 }
522 }
523 sblog << "##### uc_stk end ####\n";
524 }
525
526 void gcm::pop_uc_stack() {
527 nuc_map &pm = nuc_stk[ucs_level];
528 --ucs_level;
529 nuc_map &cm = nuc_stk[ucs_level];
530
531 GCM_DUMP(
532 sblog << "merging use stack from level " << (ucs_level+1)
533 << " to " << ucs_level << "\n";
534 );
535
536 for (nuc_map::iterator N, I = pm.begin(), E = pm.end(); I != E; ++I) {
537 node *n = I->first;
538
539 GCM_DUMP(
540 sblog << " " << cm[n] << " += " << I->second << " for ";
541 dump::dump_op(n);
542 sblog << "\n";
543 );
544
545 unsigned uc = cm[n] += I->second;
546
547 if (n->parent == &pending && uc == uses[n]) {
548 cm.erase(n);
549 pending_nodes.push_back(n);
550 GCM_DUMP(
551 sblog << "pushed pending_node due to stack pop ";
552 dump::dump_op(n);
553 sblog << "\n";
554 );
555 }
556 }
557 }
558
559 void gcm::bu_find_best_bb(node *n, op_info &oi) {
560
561 GCM_DUMP(
562 sblog << " find best bb : ";
563 dump::dump_op(n);
564 sblog << "\n";
565 );
566
567 if (oi.bottom_bb)
568 return;
569
570 // don't hoist generated copies
571 if (n->flags & NF_DONT_HOIST) {
572 oi.bottom_bb = bu_bb;
573 return;
574 }
575
576 bb_node* best_bb = bu_bb;
577 bb_node* top_bb = oi.top_bb;
578 assert(oi.top_bb && !oi.bottom_bb);
579
580 node *c = best_bb;
581
582 // FIXME top_bb may be located inside the loop so we'll never enter it
583 // in the loop below, and the instruction will be incorrectly placed at the
584 // beginning of the shader.
585 // For now just check if top_bb's loop_level is higher than of
586 // current bb and abort the search for better bb in such case,
587 // but this problem may require more complete (and more expensive) fix
588 if (top_bb->loop_level <= best_bb->loop_level) {
589 while (c && c != top_bb) {
590
591 if (c->prev) {
592 c = c->prev;
593 } else {
594 c = c->parent;
595 if (!c)
596 break;
597 continue;
598 }
599
600 if (c->subtype == NST_BB) {
601 bb_node *bb = static_cast<bb_node*>(c);
602 if (bb->loop_level < best_bb->loop_level)
603 best_bb = bb;
604 }
605 }
606 }
607
608 oi.bottom_bb = best_bb;
609 }
610
611 void gcm::add_ready(node *n) {
612 sched_queue_id sq = sh.get_queue_id(n);
613 if (n->flags & NF_SCHEDULE_EARLY)
614 bu_ready_early[sq].push_back(n);
615 else if (sq == SQ_ALU && n->is_copy_mov())
616 bu_ready[sq].push_front(n);
617 else if (n->is_alu_inst()) {
618 alu_node *a = static_cast<alu_node*>(n);
619 if (a->bc.op_ptr->flags & AF_PRED && a->dst[2]) {
620 // PRED_SET instruction that updates exec mask
621 pending_exec_mask_update = true;
622 }
623 bu_ready_next[sq].push_back(n);
624 } else
625 bu_ready_next[sq].push_back(n);
626 }
627
628 void gcm::bu_release_op(node * n) {
629 op_info &oi = op_map[n];
630
631 GCM_DUMP(
632 sblog << " bu release op ";
633 dump::dump_op(n);
634 );
635
636 nuc_stk[ucs_level].erase(n);
637 pending.remove_node(n);
638
639 bu_find_best_bb(n, oi);
640
641 if (oi.bottom_bb == bu_bb) {
642 GCM_DUMP( sblog << " ready\n";);
643 add_ready(n);
644 } else {
645 GCM_DUMP( sblog << " ready_above\n";);
646 ready_above.push_back(n);
647 }
648 }
649
650 void gcm::bu_release_phi_defs(container_node* p, unsigned op)
651 {
652 for (node_riterator I = p->rbegin(), E = p->rend(); I != E; ++I) {
653 node *o = *I;
654 value *v = o->src[op];
655 if (v && !v->is_readonly())
656 pending_defs.push_back(o->src[op]);
657
658 }
659 }
660
661 unsigned gcm::get_uc_vec(vvec &vv) {
662 unsigned c = 0;
663 for (vvec::iterator I = vv.begin(), E = vv.end(); I != E; ++I) {
664 value *v = *I;
665 if (!v)
666 continue;
667
668 if (v->is_rel())
669 c += get_uc_vec(v->mdef);
670 else
671 c += v->use_count();
672 }
673 return c;
674 }
675
676 void gcm::init_use_count(nuc_map& m, container_node &s) {
677 m.clear();
678 for (node_iterator I = s.begin(), E = s.end(); I != E; ++I) {
679 node *n = *I;
680 unsigned uc = get_uc_vec(n->dst);
681 GCM_DUMP(
682 sblog << "uc " << uc << " ";
683 dump::dump_op(n);
684 sblog << "\n";
685 );
686 if (!uc) {
687 pending_nodes.push_back(n);
688 GCM_DUMP(
689 sblog << "pushed pending_node in init ";
690 dump::dump_op(n);
691 sblog << "\n";
692 );
693
694 } else
695 m[n] = uc;
696 }
697 }
698
699 void gcm::bu_release_val(value* v) {
700 node *n = v->any_def();
701
702 if (n && n->parent == &pending) {
703 nuc_map &m = nuc_stk[ucs_level];
704 unsigned uc = ++m[n];
705 unsigned uc2 = uses[n];
706
707 if (live.add_val(v)) {
708 ++live_count;
709 GCM_DUMP ( sblog << "live_count: " << live_count << "\n"; );
710 }
711
712 GCM_DUMP(
713 sblog << "release val ";
714 dump::dump_val(v);
715 sblog << " for node ";
716 dump::dump_op(n);
717 sblog << " new uc=" << uc << ", total " << uc2 << "\n";
718 );
719
720 if (uc == uc2)
721 bu_release_op(n);
722 }
723
724 }
725
726 void gcm::init_def_count(nuc_map& m, container_node& s) {
727 m.clear();
728 for (node_iterator I = s.begin(), E = s.end(); I != E; ++I) {
729 node *n = *I;
730 unsigned dc = get_dc_vec(n->src, true) + get_dc_vec(n->dst, false);
731 m[n] = dc;
732
733 GCM_DUMP(
734 sblog << "dc " << dc << " ";
735 dump::dump_op(n);
736 sblog << "\n";
737 );
738 }
739 }
740
741 unsigned gcm::get_dc_vec(vvec& vv, bool src) {
742 unsigned c = 0;
743 for (vvec::iterator I = vv.begin(), E = vv.end(); I != E; ++I) {
744 value *v = *I;
745 if (!v || v->is_readonly())
746 continue;
747
748 if (v->is_rel()) {
749 c += v->rel->def != NULL;
750 c += get_dc_vec(v->muse, true);
751 }
752 else if (src) {
753 c += v->def != NULL;
754 c += v->adef != NULL;
755 }
756 }
757 return c;
758 }
759
760 unsigned gcm::real_alu_count(sched_queue& q, unsigned max) {
761 sq_iterator I(q.begin()), E(q.end());
762 unsigned c = 0;
763
764 while (I != E && c < max) {
765 node *n = *I;
766 if (n->is_alu_inst()) {
767 if (!n->is_copy_mov() || !n->src[0]->is_any_gpr())
768 ++c;
769 } else if (n->is_alu_packed()) {
770 c += static_cast<container_node*>(n)->count();
771 }
772 ++I;
773 }
774
775 return c;
776 }
777
778 bool gcm::check_alu_ready_count(unsigned threshold) {
779 unsigned r = real_alu_count(bu_ready[SQ_ALU], threshold);
780 if (r >= threshold)
781 return true;
782 r += real_alu_count(bu_ready_next[SQ_ALU], threshold - r);
783 return r >= threshold;
784 }
785
786 } // namespace r600_sb