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