freedreno: fix FD_MESA_DEBUG=inorder
[mesa.git] / src / util / register_allocate.c
1 /*
2 * Copyright © 2010 Intel Corporation
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Eric Anholt <eric@anholt.net>
25 *
26 */
27
28 /** @file register_allocate.c
29 *
30 * Graph-coloring register allocator.
31 *
32 * The basic idea of graph coloring is to make a node in a graph for
33 * every thing that needs a register (color) number assigned, and make
34 * edges in the graph between nodes that interfere (can't be allocated
35 * to the same register at the same time).
36 *
37 * During the "simplify" process, any any node with fewer edges than
38 * there are registers means that that edge can get assigned a
39 * register regardless of what its neighbors choose, so that node is
40 * pushed on a stack and removed (with its edges) from the graph.
41 * That likely causes other nodes to become trivially colorable as well.
42 *
43 * Then during the "select" process, nodes are popped off of that
44 * stack, their edges restored, and assigned a color different from
45 * their neighbors. Because they were pushed on the stack only when
46 * they were trivially colorable, any color chosen won't interfere
47 * with the registers to be popped later.
48 *
49 * The downside to most graph coloring is that real hardware often has
50 * limitations, like registers that need to be allocated to a node in
51 * pairs, or aligned on some boundary. This implementation follows
52 * the paper "Retargetable Graph-Coloring Register Allocation for
53 * Irregular Architectures" by Johan Runeson and Sven-Olof Nyström.
54 *
55 * In this system, there are register classes each containing various
56 * registers, and registers may interfere with other registers. For
57 * example, one might have a class of base registers, and a class of
58 * aligned register pairs that would each interfere with their pair of
59 * the base registers. Each node has a register class it needs to be
60 * assigned to. Define p(B) to be the size of register class B, and
61 * q(B,C) to be the number of registers in B that the worst choice
62 * register in C could conflict with. Then, this system replaces the
63 * basic graph coloring test of "fewer edges from this node than there
64 * are registers" with "For this node of class B, the sum of q(B,C)
65 * for each neighbor node of class C is less than pB".
66 *
67 * A nice feature of the pq test is that q(B,C) can be computed once
68 * up front and stored in a 2-dimensional array, so that the cost of
69 * coloring a node is constant with the number of registers. We do
70 * this during ra_set_finalize().
71 */
72
73 #include <stdbool.h>
74
75 #include "ralloc.h"
76 #include "main/imports.h"
77 #include "main/macros.h"
78 #include "util/bitset.h"
79 #include "register_allocate.h"
80
81 #define NO_REG ~0U
82
83 struct ra_reg {
84 BITSET_WORD *conflicts;
85 unsigned int *conflict_list;
86 unsigned int conflict_list_size;
87 unsigned int num_conflicts;
88 };
89
90 struct ra_regs {
91 struct ra_reg *regs;
92 unsigned int count;
93
94 struct ra_class **classes;
95 unsigned int class_count;
96
97 bool round_robin;
98 };
99
100 struct ra_class {
101 /**
102 * Bitset indicating which registers belong to this class.
103 *
104 * (If bit N is set, then register N belongs to this class.)
105 */
106 BITSET_WORD *regs;
107
108 /**
109 * p(B) in Runeson/Nyström paper.
110 *
111 * This is "how many regs are in the set."
112 */
113 unsigned int p;
114
115 /**
116 * q(B,C) (indexed by C, B is this register class) in
117 * Runeson/Nyström paper. This is "how many registers of B could
118 * the worst choice register from C conflict with".
119 */
120 unsigned int *q;
121 };
122
123 struct ra_node {
124 /** @{
125 *
126 * List of which nodes this node interferes with. This should be
127 * symmetric with the other node.
128 */
129 BITSET_WORD *adjacency;
130 unsigned int *adjacency_list;
131 unsigned int adjacency_list_size;
132 unsigned int adjacency_count;
133 /** @} */
134
135 unsigned int class;
136
137 /* Client-assigned register, if assigned, or NO_REG. */
138 unsigned int forced_reg;
139
140 /* Register, if assigned, or NO_REG. */
141 unsigned int reg;
142
143 /**
144 * The q total, as defined in the Runeson/Nyström paper, for all the
145 * interfering nodes not in the stack.
146 */
147 unsigned int q_total;
148
149 /* For an implementation that needs register spilling, this is the
150 * approximate cost of spilling this node.
151 */
152 float spill_cost;
153
154 /* Temporary data for the algorithm to scratch around in */
155 struct {
156 /**
157 * Temporary version of q_total which we decrement as things are placed
158 * into the stack.
159 */
160 unsigned int q_total;
161 } tmp;
162 };
163
164 struct ra_graph {
165 struct ra_regs *regs;
166 /**
167 * the variables that need register allocation.
168 */
169 struct ra_node *nodes;
170 unsigned int count; /**< count of nodes. */
171
172 unsigned int alloc; /**< count of nodes allocated. */
173
174 unsigned int (*select_reg_callback)(struct ra_graph *g, BITSET_WORD *regs,
175 void *data);
176 void *select_reg_callback_data;
177
178 /* Temporary data for the algorithm to scratch around in */
179 struct {
180 unsigned int *stack;
181 unsigned int stack_count;
182
183 /** Bit-set indicating, for each register, if it's in the stack */
184 BITSET_WORD *in_stack;
185
186 /** Bit-set indicating, for each register, if it pre-assigned */
187 BITSET_WORD *reg_assigned;
188
189 /** Bit-set indicating, for each register, the value of the pq test */
190 BITSET_WORD *pq_test;
191
192 /** For each BITSET_WORD, the minimum q value or ~0 if unknown */
193 unsigned int *min_q_total;
194
195 /*
196 * * For each BITSET_WORD, the node with the minimum q_total if
197 * min_q_total[i] != ~0.
198 */
199 unsigned int *min_q_node;
200
201 /**
202 * Tracks the start of the set of optimistically-colored registers in the
203 * stack.
204 */
205 unsigned int stack_optimistic_start;
206 } tmp;
207 };
208
209 /**
210 * Creates a set of registers for the allocator.
211 *
212 * mem_ctx is a ralloc context for the allocator. The reg set may be freed
213 * using ralloc_free().
214 */
215 struct ra_regs *
216 ra_alloc_reg_set(void *mem_ctx, unsigned int count, bool need_conflict_lists)
217 {
218 unsigned int i;
219 struct ra_regs *regs;
220
221 regs = rzalloc(mem_ctx, struct ra_regs);
222 regs->count = count;
223 regs->regs = rzalloc_array(regs, struct ra_reg, count);
224
225 for (i = 0; i < count; i++) {
226 regs->regs[i].conflicts = rzalloc_array(regs->regs, BITSET_WORD,
227 BITSET_WORDS(count));
228 BITSET_SET(regs->regs[i].conflicts, i);
229
230 if (need_conflict_lists) {
231 regs->regs[i].conflict_list = ralloc_array(regs->regs,
232 unsigned int, 4);
233 regs->regs[i].conflict_list_size = 4;
234 regs->regs[i].conflict_list[0] = i;
235 } else {
236 regs->regs[i].conflict_list = NULL;
237 regs->regs[i].conflict_list_size = 0;
238 }
239 regs->regs[i].num_conflicts = 1;
240 }
241
242 return regs;
243 }
244
245 /**
246 * The register allocator by default prefers to allocate low register numbers,
247 * since it was written for hardware (gen4/5 Intel) that is limited in its
248 * multithreadedness by the number of registers used in a given shader.
249 *
250 * However, for hardware without that restriction, densely packed register
251 * allocation can put serious constraints on instruction scheduling. This
252 * function tells the allocator to rotate around the registers if possible as
253 * it allocates the nodes.
254 */
255 void
256 ra_set_allocate_round_robin(struct ra_regs *regs)
257 {
258 regs->round_robin = true;
259 }
260
261 static void
262 ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2)
263 {
264 struct ra_reg *reg1 = &regs->regs[r1];
265
266 if (reg1->conflict_list) {
267 if (reg1->conflict_list_size == reg1->num_conflicts) {
268 reg1->conflict_list_size *= 2;
269 reg1->conflict_list = reralloc(regs->regs, reg1->conflict_list,
270 unsigned int, reg1->conflict_list_size);
271 }
272 reg1->conflict_list[reg1->num_conflicts++] = r2;
273 }
274 BITSET_SET(reg1->conflicts, r2);
275 }
276
277 void
278 ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2)
279 {
280 if (!BITSET_TEST(regs->regs[r1].conflicts, r2)) {
281 ra_add_conflict_list(regs, r1, r2);
282 ra_add_conflict_list(regs, r2, r1);
283 }
284 }
285
286 /**
287 * Adds a conflict between base_reg and reg, and also between reg and
288 * anything that base_reg conflicts with.
289 *
290 * This can simplify code for setting up multiple register classes
291 * which are aggregates of some base hardware registers, compared to
292 * explicitly using ra_add_reg_conflict.
293 */
294 void
295 ra_add_transitive_reg_conflict(struct ra_regs *regs,
296 unsigned int base_reg, unsigned int reg)
297 {
298 unsigned int i;
299
300 ra_add_reg_conflict(regs, reg, base_reg);
301
302 for (i = 0; i < regs->regs[base_reg].num_conflicts; i++) {
303 ra_add_reg_conflict(regs, reg, regs->regs[base_reg].conflict_list[i]);
304 }
305 }
306
307 /**
308 * Set up conflicts between base_reg and it's two half registers reg0 and
309 * reg1, but take care to not add conflicts between reg0 and reg1.
310 *
311 * This is useful for architectures where full size registers are aliased by
312 * two half size registers (eg 32 bit float and 16 bit float registers).
313 */
314 void
315 ra_add_transitive_reg_pair_conflict(struct ra_regs *regs,
316 unsigned int base_reg, unsigned int reg0, unsigned int reg1)
317 {
318 unsigned int i;
319
320 ra_add_reg_conflict(regs, reg0, base_reg);
321 ra_add_reg_conflict(regs, reg1, base_reg);
322
323 for (i = 0; i < regs->regs[base_reg].num_conflicts; i++) {
324 unsigned int conflict = regs->regs[base_reg].conflict_list[i];
325 if (conflict != reg1)
326 ra_add_reg_conflict(regs, reg0, regs->regs[base_reg].conflict_list[i]);
327 if (conflict != reg0)
328 ra_add_reg_conflict(regs, reg1, regs->regs[base_reg].conflict_list[i]);
329 }
330 }
331
332 /**
333 * Makes every conflict on the given register transitive. In other words,
334 * every register that conflicts with r will now conflict with every other
335 * register conflicting with r.
336 *
337 * This can simplify code for setting up multiple register classes
338 * which are aggregates of some base hardware registers, compared to
339 * explicitly using ra_add_reg_conflict.
340 */
341 void
342 ra_make_reg_conflicts_transitive(struct ra_regs *regs, unsigned int r)
343 {
344 struct ra_reg *reg = &regs->regs[r];
345 int c;
346
347 BITSET_FOREACH_SET(c, reg->conflicts, regs->count) {
348 struct ra_reg *other = &regs->regs[c];
349 unsigned i;
350 for (i = 0; i < BITSET_WORDS(regs->count); i++)
351 other->conflicts[i] |= reg->conflicts[i];
352 }
353 }
354
355 unsigned int
356 ra_alloc_reg_class(struct ra_regs *regs)
357 {
358 struct ra_class *class;
359
360 regs->classes = reralloc(regs->regs, regs->classes, struct ra_class *,
361 regs->class_count + 1);
362
363 class = rzalloc(regs, struct ra_class);
364 regs->classes[regs->class_count] = class;
365
366 class->regs = rzalloc_array(class, BITSET_WORD, BITSET_WORDS(regs->count));
367
368 return regs->class_count++;
369 }
370
371 void
372 ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r)
373 {
374 struct ra_class *class = regs->classes[c];
375
376 BITSET_SET(class->regs, r);
377 class->p++;
378 }
379
380 /**
381 * Returns true if the register belongs to the given class.
382 */
383 static bool
384 reg_belongs_to_class(unsigned int r, struct ra_class *c)
385 {
386 return BITSET_TEST(c->regs, r);
387 }
388
389 /**
390 * Must be called after all conflicts and register classes have been
391 * set up and before the register set is used for allocation.
392 * To avoid costly q value computation, use the q_values paramater
393 * to pass precomputed q values to this function.
394 */
395 void
396 ra_set_finalize(struct ra_regs *regs, unsigned int **q_values)
397 {
398 unsigned int b, c;
399
400 for (b = 0; b < regs->class_count; b++) {
401 regs->classes[b]->q = ralloc_array(regs, unsigned int, regs->class_count);
402 }
403
404 if (q_values) {
405 for (b = 0; b < regs->class_count; b++) {
406 for (c = 0; c < regs->class_count; c++) {
407 regs->classes[b]->q[c] = q_values[b][c];
408 }
409 }
410 } else {
411 /* Compute, for each class B and C, how many regs of B an
412 * allocation to C could conflict with.
413 */
414 for (b = 0; b < regs->class_count; b++) {
415 for (c = 0; c < regs->class_count; c++) {
416 unsigned int rc;
417 int max_conflicts = 0;
418
419 for (rc = 0; rc < regs->count; rc++) {
420 int conflicts = 0;
421 unsigned int i;
422
423 if (!reg_belongs_to_class(rc, regs->classes[c]))
424 continue;
425
426 for (i = 0; i < regs->regs[rc].num_conflicts; i++) {
427 unsigned int rb = regs->regs[rc].conflict_list[i];
428 if (reg_belongs_to_class(rb, regs->classes[b]))
429 conflicts++;
430 }
431 max_conflicts = MAX2(max_conflicts, conflicts);
432 }
433 regs->classes[b]->q[c] = max_conflicts;
434 }
435 }
436 }
437
438 for (b = 0; b < regs->count; b++) {
439 ralloc_free(regs->regs[b].conflict_list);
440 regs->regs[b].conflict_list = NULL;
441 }
442 }
443
444 static void
445 ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2)
446 {
447 BITSET_SET(g->nodes[n1].adjacency, n2);
448
449 assert(n1 != n2);
450
451 int n1_class = g->nodes[n1].class;
452 int n2_class = g->nodes[n2].class;
453 g->nodes[n1].q_total += g->regs->classes[n1_class]->q[n2_class];
454
455 if (g->nodes[n1].adjacency_count >=
456 g->nodes[n1].adjacency_list_size) {
457 g->nodes[n1].adjacency_list_size *= 2;
458 g->nodes[n1].adjacency_list = reralloc(g, g->nodes[n1].adjacency_list,
459 unsigned int,
460 g->nodes[n1].adjacency_list_size);
461 }
462
463 g->nodes[n1].adjacency_list[g->nodes[n1].adjacency_count] = n2;
464 g->nodes[n1].adjacency_count++;
465 }
466
467 static void
468 ra_node_remove_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2)
469 {
470 BITSET_CLEAR(g->nodes[n1].adjacency, n2);
471
472 assert(n1 != n2);
473
474 int n1_class = g->nodes[n1].class;
475 int n2_class = g->nodes[n2].class;
476 g->nodes[n1].q_total -= g->regs->classes[n1_class]->q[n2_class];
477
478 unsigned int i;
479 for (i = 0; i < g->nodes[n1].adjacency_count; i++) {
480 if (g->nodes[n1].adjacency_list[i] == n2) {
481 memmove(&g->nodes[n1].adjacency_list[i],
482 &g->nodes[n1].adjacency_list[i + 1],
483 (g->nodes[n1].adjacency_count - i - 1) *
484 sizeof(g->nodes[n1].adjacency_list[0]));
485 break;
486 }
487 }
488 assert(i < g->nodes[n1].adjacency_count);
489 g->nodes[n1].adjacency_count--;
490 }
491
492 static void
493 ra_realloc_interference_graph(struct ra_graph *g, unsigned int alloc)
494 {
495 if (alloc <= g->alloc)
496 return;
497
498 /* If we always have a whole number of BITSET_WORDs, it makes it much
499 * easier to memset the top of the growing bitsets.
500 */
501 assert(g->alloc % BITSET_WORDBITS == 0);
502 alloc = ALIGN(alloc, BITSET_WORDBITS);
503
504 g->nodes = reralloc(g, g->nodes, struct ra_node, alloc);
505
506 unsigned g_bitset_count = BITSET_WORDS(g->alloc);
507 unsigned bitset_count = BITSET_WORDS(alloc);
508 /* For nodes already in the graph, we just have to grow the adjacency set */
509 for (unsigned i = 0; i < g->alloc; i++) {
510 assert(g->nodes[i].adjacency != NULL);
511 g->nodes[i].adjacency = rerzalloc(g, g->nodes[i].adjacency, BITSET_WORD,
512 g_bitset_count, bitset_count);
513 }
514
515 /* For new nodes, we have to fully initialize them */
516 for (unsigned i = g->alloc; i < alloc; i++) {
517 memset(&g->nodes[i], 0, sizeof(g->nodes[i]));
518 g->nodes[i].adjacency = rzalloc_array(g, BITSET_WORD, bitset_count);
519 g->nodes[i].adjacency_list_size = 4;
520 g->nodes[i].adjacency_list =
521 ralloc_array(g, unsigned int, g->nodes[i].adjacency_list_size);
522 g->nodes[i].adjacency_count = 0;
523 g->nodes[i].q_total = 0;
524
525 g->nodes[i].forced_reg = NO_REG;
526 g->nodes[i].reg = NO_REG;
527 }
528
529 /* These are scratch values and don't need to be zeroed. We'll clear them
530 * as part of ra_select() setup.
531 */
532 g->tmp.stack = reralloc(g, g->tmp.stack, unsigned int, alloc);
533 g->tmp.in_stack = reralloc(g, g->tmp.in_stack, BITSET_WORD, bitset_count);
534
535 g->tmp.reg_assigned = reralloc(g, g->tmp.reg_assigned, BITSET_WORD,
536 bitset_count);
537 g->tmp.pq_test = reralloc(g, g->tmp.pq_test, BITSET_WORD, bitset_count);
538 g->tmp.min_q_total = reralloc(g, g->tmp.min_q_total, unsigned int,
539 bitset_count);
540 g->tmp.min_q_node = reralloc(g, g->tmp.min_q_node, unsigned int,
541 bitset_count);
542
543 g->alloc = alloc;
544 }
545
546 struct ra_graph *
547 ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count)
548 {
549 struct ra_graph *g;
550
551 g = rzalloc(NULL, struct ra_graph);
552 g->regs = regs;
553 g->count = count;
554 ra_realloc_interference_graph(g, count);
555
556 return g;
557 }
558
559 void
560 ra_resize_interference_graph(struct ra_graph *g, unsigned int count)
561 {
562 g->count = count;
563 if (count > g->alloc)
564 ra_realloc_interference_graph(g, g->alloc * 2);
565 }
566
567 void ra_set_select_reg_callback(struct ra_graph *g,
568 unsigned int (*callback)(struct ra_graph *g,
569 BITSET_WORD *regs,
570 void *data),
571 void *data)
572 {
573 g->select_reg_callback = callback;
574 g->select_reg_callback_data = data;
575 }
576
577 void
578 ra_set_node_class(struct ra_graph *g,
579 unsigned int n, unsigned int class)
580 {
581 g->nodes[n].class = class;
582 }
583
584 unsigned int
585 ra_get_node_class(struct ra_graph *g,
586 unsigned int n)
587 {
588 return g->nodes[n].class;
589 }
590
591 unsigned int
592 ra_add_node(struct ra_graph *g, unsigned int class)
593 {
594 unsigned int n = g->count;
595 ra_resize_interference_graph(g, g->count + 1);
596
597 ra_set_node_class(g, n, class);
598
599 return n;
600 }
601
602 void
603 ra_add_node_interference(struct ra_graph *g,
604 unsigned int n1, unsigned int n2)
605 {
606 assert(n1 < g->count && n2 < g->count);
607 if (n1 != n2 && !BITSET_TEST(g->nodes[n1].adjacency, n2)) {
608 ra_add_node_adjacency(g, n1, n2);
609 ra_add_node_adjacency(g, n2, n1);
610 }
611 }
612
613 void
614 ra_reset_node_interference(struct ra_graph *g, unsigned int n)
615 {
616 for (unsigned int i = 0; i < g->nodes[n].adjacency_count; i++)
617 ra_node_remove_adjacency(g, g->nodes[n].adjacency_list[i], n);
618
619 memset(g->nodes[n].adjacency, 0,
620 BITSET_WORDS(g->count) * sizeof(BITSET_WORD));
621 g->nodes[n].adjacency_count = 0;
622 }
623
624 static void
625 update_pq_info(struct ra_graph *g, unsigned int n)
626 {
627 int i = n / BITSET_WORDBITS;
628 int n_class = g->nodes[n].class;
629 if (g->nodes[n].tmp.q_total < g->regs->classes[n_class]->p) {
630 BITSET_SET(g->tmp.pq_test, n);
631 } else if (g->tmp.min_q_total[i] != UINT_MAX) {
632 /* Only update min_q_total and min_q_node if min_q_total != UINT_MAX so
633 * that we don't update while we have stale data and accidentally mark
634 * it as non-stale. Also, in order to remain consistent with the old
635 * naive implementation of the algorithm, we do a lexicographical sort
636 * to ensure that we always choose the node with the highest node index.
637 */
638 if (g->nodes[n].tmp.q_total < g->tmp.min_q_total[i] ||
639 (g->nodes[n].tmp.q_total == g->tmp.min_q_total[i] &&
640 n > g->tmp.min_q_node[i])) {
641 g->tmp.min_q_total[i] = g->nodes[n].tmp.q_total;
642 g->tmp.min_q_node[i] = n;
643 }
644 }
645 }
646
647 static void
648 add_node_to_stack(struct ra_graph *g, unsigned int n)
649 {
650 unsigned int i;
651 int n_class = g->nodes[n].class;
652
653 assert(!BITSET_TEST(g->tmp.in_stack, n));
654
655 for (i = 0; i < g->nodes[n].adjacency_count; i++) {
656 unsigned int n2 = g->nodes[n].adjacency_list[i];
657 unsigned int n2_class = g->nodes[n2].class;
658
659 if (!BITSET_TEST(g->tmp.in_stack, n2) &&
660 !BITSET_TEST(g->tmp.reg_assigned, n2)) {
661 assert(g->nodes[n2].tmp.q_total >= g->regs->classes[n2_class]->q[n_class]);
662 g->nodes[n2].tmp.q_total -= g->regs->classes[n2_class]->q[n_class];
663 update_pq_info(g, n2);
664 }
665 }
666
667 g->tmp.stack[g->tmp.stack_count] = n;
668 g->tmp.stack_count++;
669 BITSET_SET(g->tmp.in_stack, n);
670
671 /* Flag the min_q_total for n's block as dirty so it gets recalculated */
672 g->tmp.min_q_total[n / BITSET_WORDBITS] = UINT_MAX;
673 }
674
675 /**
676 * Simplifies the interference graph by pushing all
677 * trivially-colorable nodes into a stack of nodes to be colored,
678 * removing them from the graph, and rinsing and repeating.
679 *
680 * If we encounter a case where we can't push any nodes on the stack, then
681 * we optimistically choose a node and push it on the stack. We heuristically
682 * push the node with the lowest total q value, since it has the fewest
683 * neighbors and therefore is most likely to be allocated.
684 */
685 static void
686 ra_simplify(struct ra_graph *g)
687 {
688 bool progress = true;
689 unsigned int stack_optimistic_start = UINT_MAX;
690
691 /* Figure out the high bit and bit mask for the first iteration of a loop
692 * over BITSET_WORDs.
693 */
694 const unsigned int top_word_high_bit = (g->count - 1) % BITSET_WORDBITS;
695
696 /* Do a quick pre-pass to set things up */
697 g->tmp.stack_count = 0;
698 for (int i = BITSET_WORDS(g->count) - 1, high_bit = top_word_high_bit;
699 i >= 0; i--, high_bit = BITSET_WORDBITS - 1) {
700 g->tmp.in_stack[i] = 0;
701 g->tmp.reg_assigned[i] = 0;
702 g->tmp.pq_test[i] = 0;
703 g->tmp.min_q_total[i] = UINT_MAX;
704 g->tmp.min_q_node[i] = UINT_MAX;
705 for (int j = high_bit; j >= 0; j--) {
706 unsigned int n = i * BITSET_WORDBITS + j;
707 g->nodes[n].reg = g->nodes[n].forced_reg;
708 g->nodes[n].tmp.q_total = g->nodes[n].q_total;
709 if (g->nodes[n].reg != NO_REG)
710 g->tmp.reg_assigned[i] |= BITSET_BIT(j);
711 update_pq_info(g, n);
712 }
713 }
714
715 while (progress) {
716 unsigned int min_q_total = UINT_MAX;
717 unsigned int min_q_node = UINT_MAX;
718
719 progress = false;
720
721 for (int i = BITSET_WORDS(g->count) - 1, high_bit = top_word_high_bit;
722 i >= 0; i--, high_bit = BITSET_WORDBITS - 1) {
723 BITSET_WORD mask = ~(BITSET_WORD)0 >> (31 - high_bit);
724
725 BITSET_WORD skip = g->tmp.in_stack[i] | g->tmp.reg_assigned[i];
726 if (skip == mask)
727 continue;
728
729 BITSET_WORD pq = g->tmp.pq_test[i] & ~skip;
730 if (pq) {
731 /* In this case, we have stuff we can immediately take off the
732 * stack. This also means that we're guaranteed to make progress
733 * and we don't need to bother updating lowest_q_total because we
734 * know we're going to loop again before attempting to do anything
735 * optimistic.
736 */
737 for (int j = high_bit; j >= 0; j--) {
738 if (pq & BITSET_BIT(j)) {
739 unsigned int n = i * BITSET_WORDBITS + j;
740 assert(n < g->count);
741 add_node_to_stack(g, n);
742 /* add_node_to_stack() may update pq_test for this word so
743 * we need to update our local copy.
744 */
745 pq = g->tmp.pq_test[i] & ~skip;
746 progress = true;
747 }
748 }
749 } else if (!progress) {
750 if (g->tmp.min_q_total[i] == UINT_MAX) {
751 /* The min_q_total and min_q_node are dirty because we added
752 * one of these nodes to the stack. It needs to be
753 * recalculated.
754 */
755 for (int j = high_bit; j >= 0; j--) {
756 if (skip & BITSET_BIT(j))
757 continue;
758
759 unsigned int n = i * BITSET_WORDBITS + j;
760 assert(n < g->count);
761 if (g->nodes[n].tmp.q_total < g->tmp.min_q_total[i]) {
762 g->tmp.min_q_total[i] = g->nodes[n].tmp.q_total;
763 g->tmp.min_q_node[i] = n;
764 }
765 }
766 }
767 if (g->tmp.min_q_total[i] < min_q_total) {
768 min_q_node = g->tmp.min_q_node[i];
769 min_q_total = g->tmp.min_q_total[i];
770 }
771 }
772 }
773
774 if (!progress && min_q_total != UINT_MAX) {
775 if (stack_optimistic_start == UINT_MAX)
776 stack_optimistic_start = g->tmp.stack_count;
777
778 add_node_to_stack(g, min_q_node);
779 progress = true;
780 }
781 }
782
783 g->tmp.stack_optimistic_start = stack_optimistic_start;
784 }
785
786 static bool
787 ra_any_neighbors_conflict(struct ra_graph *g, unsigned int n, unsigned int r)
788 {
789 unsigned int i;
790
791 for (i = 0; i < g->nodes[n].adjacency_count; i++) {
792 unsigned int n2 = g->nodes[n].adjacency_list[i];
793
794 if (!BITSET_TEST(g->tmp.in_stack, n2) &&
795 BITSET_TEST(g->regs->regs[r].conflicts, g->nodes[n2].reg)) {
796 return true;
797 }
798 }
799
800 return false;
801 }
802
803 /* Computes a bitfield of what regs are available for a given register
804 * selection.
805 *
806 * This lets drivers implement a more complicated policy than our simple first
807 * or round robin policies (which don't require knowing the whole bitset)
808 */
809 static bool
810 ra_compute_available_regs(struct ra_graph *g, unsigned int n, BITSET_WORD *regs)
811 {
812 struct ra_class *c = g->regs->classes[g->nodes[n].class];
813
814 /* Populate with the set of regs that are in the node's class. */
815 memcpy(regs, c->regs, BITSET_WORDS(g->regs->count) * sizeof(BITSET_WORD));
816
817 /* Remove any regs that conflict with nodes that we're adjacent to and have
818 * already colored.
819 */
820 for (int i = 0; i < g->nodes[n].adjacency_count; i++) {
821 unsigned int n2 = g->nodes[n].adjacency_list[i];
822 unsigned int r = g->nodes[n2].reg;
823
824 if (!BITSET_TEST(g->tmp.in_stack, n2)) {
825 for (int j = 0; j < BITSET_WORDS(g->regs->count); j++)
826 regs[j] &= ~g->regs->regs[r].conflicts[j];
827 }
828 }
829
830 for (int i = 0; i < BITSET_WORDS(g->regs->count); i++) {
831 if (regs[i])
832 return true;
833 }
834
835 return false;
836 }
837
838 /**
839 * Pops nodes from the stack back into the graph, coloring them with
840 * registers as they go.
841 *
842 * If all nodes were trivially colorable, then this must succeed. If
843 * not (optimistic coloring), then it may return false;
844 */
845 static bool
846 ra_select(struct ra_graph *g)
847 {
848 int start_search_reg = 0;
849 BITSET_WORD *select_regs = NULL;
850
851 if (g->select_reg_callback)
852 select_regs = malloc(BITSET_WORDS(g->regs->count) * sizeof(BITSET_WORD));
853
854 while (g->tmp.stack_count != 0) {
855 unsigned int ri;
856 unsigned int r = -1;
857 int n = g->tmp.stack[g->tmp.stack_count - 1];
858 struct ra_class *c = g->regs->classes[g->nodes[n].class];
859
860 /* set this to false even if we return here so that
861 * ra_get_best_spill_node() considers this node later.
862 */
863 BITSET_CLEAR(g->tmp.in_stack, n);
864
865 if (g->select_reg_callback) {
866 if (!ra_compute_available_regs(g, n, select_regs)) {
867 free(select_regs);
868 return false;
869 }
870
871 r = g->select_reg_callback(g, select_regs, g->select_reg_callback_data);
872 } else {
873 /* Find the lowest-numbered reg which is not used by a member
874 * of the graph adjacent to us.
875 */
876 for (ri = 0; ri < g->regs->count; ri++) {
877 r = (start_search_reg + ri) % g->regs->count;
878 if (!reg_belongs_to_class(r, c))
879 continue;
880
881 if (!ra_any_neighbors_conflict(g, n, r))
882 break;
883 }
884
885 if (ri >= g->regs->count)
886 return false;
887 }
888
889 g->nodes[n].reg = r;
890 g->tmp.stack_count--;
891
892 /* Rotate the starting point except for any nodes above the lowest
893 * optimistically colorable node. The likelihood that we will succeed
894 * at allocating optimistically colorable nodes is highly dependent on
895 * the way that the previous nodes popped off the stack are laid out.
896 * The round-robin strategy increases the fragmentation of the register
897 * file and decreases the number of nearby nodes assigned to the same
898 * color, what increases the likelihood of spilling with respect to the
899 * dense packing strategy.
900 */
901 if (g->regs->round_robin &&
902 g->tmp.stack_count - 1 <= g->tmp.stack_optimistic_start)
903 start_search_reg = r + 1;
904 }
905
906 free(select_regs);
907
908 return true;
909 }
910
911 bool
912 ra_allocate(struct ra_graph *g)
913 {
914 ra_simplify(g);
915 return ra_select(g);
916 }
917
918 unsigned int
919 ra_get_node_reg(struct ra_graph *g, unsigned int n)
920 {
921 if (g->nodes[n].forced_reg != NO_REG)
922 return g->nodes[n].forced_reg;
923 else
924 return g->nodes[n].reg;
925 }
926
927 /**
928 * Forces a node to a specific register. This can be used to avoid
929 * creating a register class containing one node when handling data
930 * that must live in a fixed location and is known to not conflict
931 * with other forced register assignment (as is common with shader
932 * input data). These nodes do not end up in the stack during
933 * ra_simplify(), and thus at ra_select() time it is as if they were
934 * the first popped off the stack and assigned their fixed locations.
935 * Nodes that use this function do not need to be assigned a register
936 * class.
937 *
938 * Must be called before ra_simplify().
939 */
940 void
941 ra_set_node_reg(struct ra_graph *g, unsigned int n, unsigned int reg)
942 {
943 g->nodes[n].forced_reg = reg;
944 }
945
946 static float
947 ra_get_spill_benefit(struct ra_graph *g, unsigned int n)
948 {
949 unsigned int j;
950 float benefit = 0;
951 int n_class = g->nodes[n].class;
952
953 /* Define the benefit of eliminating an interference between n, n2
954 * through spilling as q(C, B) / p(C). This is similar to the
955 * "count number of edges" approach of traditional graph coloring,
956 * but takes classes into account.
957 */
958 for (j = 0; j < g->nodes[n].adjacency_count; j++) {
959 unsigned int n2 = g->nodes[n].adjacency_list[j];
960 unsigned int n2_class = g->nodes[n2].class;
961 benefit += ((float)g->regs->classes[n_class]->q[n2_class] /
962 g->regs->classes[n_class]->p);
963 }
964
965 return benefit;
966 }
967
968 /**
969 * Returns a node number to be spilled according to the cost/benefit using
970 * the pq test, or -1 if there are no spillable nodes.
971 */
972 int
973 ra_get_best_spill_node(struct ra_graph *g)
974 {
975 unsigned int best_node = -1;
976 float best_benefit = 0.0;
977 unsigned int n;
978
979 /* Consider any nodes that we colored successfully or the node we failed to
980 * color for spilling. When we failed to color a node in ra_select(), we
981 * only considered these nodes, so spilling any other ones would not result
982 * in us making progress.
983 */
984 for (n = 0; n < g->count; n++) {
985 float cost = g->nodes[n].spill_cost;
986 float benefit;
987
988 if (cost <= 0.0f)
989 continue;
990
991 if (BITSET_TEST(g->tmp.in_stack, n))
992 continue;
993
994 benefit = ra_get_spill_benefit(g, n);
995
996 if (benefit / cost > best_benefit) {
997 best_benefit = benefit / cost;
998 best_node = n;
999 }
1000 }
1001
1002 return best_node;
1003 }
1004
1005 /**
1006 * Only nodes with a spill cost set (cost != 0.0) will be considered
1007 * for register spilling.
1008 */
1009 void
1010 ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost)
1011 {
1012 g->nodes[n].spill_cost = cost;
1013 }