Added few more stubs so that control reaches to DestroyDevice().
[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 #include <stdlib.h>
75
76 #include "blob.h"
77 #include "ralloc.h"
78 #include "main/macros.h"
79 #include "util/bitset.h"
80 #include "util/u_dynarray.h"
81 #include "u_math.h"
82 #include "register_allocate.h"
83
84 struct ra_reg {
85 BITSET_WORD *conflicts;
86 struct util_dynarray conflict_list;
87 };
88
89 struct ra_regs {
90 struct ra_reg *regs;
91 unsigned int count;
92
93 struct ra_class **classes;
94 unsigned int class_count;
95
96 bool round_robin;
97 };
98
99 struct ra_class {
100 /**
101 * Bitset indicating which registers belong to this class.
102 *
103 * (If bit N is set, then register N belongs to this class.)
104 */
105 BITSET_WORD *regs;
106
107 /**
108 * p(B) in Runeson/Nyström paper.
109 *
110 * This is "how many regs are in the set."
111 */
112 unsigned int p;
113
114 /**
115 * q(B,C) (indexed by C, B is this register class) in
116 * Runeson/Nyström paper. This is "how many registers of B could
117 * the worst choice register from C conflict with".
118 */
119 unsigned int *q;
120 };
121
122 struct ra_node {
123 /** @{
124 *
125 * List of which nodes this node interferes with. This should be
126 * symmetric with the other node.
127 */
128 BITSET_WORD *adjacency;
129
130 struct util_dynarray adjacency_list;
131 /** @} */
132
133 unsigned int class;
134
135 /* Client-assigned register, if assigned, or NO_REG. */
136 unsigned int forced_reg;
137
138 /* Register, if assigned, or NO_REG. */
139 unsigned int reg;
140
141 /**
142 * The q total, as defined in the Runeson/Nyström paper, for all the
143 * interfering nodes not in the stack.
144 */
145 unsigned int q_total;
146
147 /* For an implementation that needs register spilling, this is the
148 * approximate cost of spilling this node.
149 */
150 float spill_cost;
151
152 /* Temporary data for the algorithm to scratch around in */
153 struct {
154 /**
155 * Temporary version of q_total which we decrement as things are placed
156 * into the stack.
157 */
158 unsigned int q_total;
159 } tmp;
160 };
161
162 struct ra_graph {
163 struct ra_regs *regs;
164 /**
165 * the variables that need register allocation.
166 */
167 struct ra_node *nodes;
168 unsigned int count; /**< count of nodes. */
169
170 unsigned int alloc; /**< count of nodes allocated. */
171
172 ra_select_reg_callback select_reg_callback;
173 void *select_reg_callback_data;
174
175 /* Temporary data for the algorithm to scratch around in */
176 struct {
177 unsigned int *stack;
178 unsigned int stack_count;
179
180 /** Bit-set indicating, for each register, if it's in the stack */
181 BITSET_WORD *in_stack;
182
183 /** Bit-set indicating, for each register, if it pre-assigned */
184 BITSET_WORD *reg_assigned;
185
186 /** Bit-set indicating, for each register, the value of the pq test */
187 BITSET_WORD *pq_test;
188
189 /** For each BITSET_WORD, the minimum q value or ~0 if unknown */
190 unsigned int *min_q_total;
191
192 /*
193 * * For each BITSET_WORD, the node with the minimum q_total if
194 * min_q_total[i] != ~0.
195 */
196 unsigned int *min_q_node;
197
198 /**
199 * Tracks the start of the set of optimistically-colored registers in the
200 * stack.
201 */
202 unsigned int stack_optimistic_start;
203 } tmp;
204 };
205
206 /**
207 * Creates a set of registers for the allocator.
208 *
209 * mem_ctx is a ralloc context for the allocator. The reg set may be freed
210 * using ralloc_free().
211 */
212 struct ra_regs *
213 ra_alloc_reg_set(void *mem_ctx, unsigned int count, bool need_conflict_lists)
214 {
215 unsigned int i;
216 struct ra_regs *regs;
217
218 regs = rzalloc(mem_ctx, struct ra_regs);
219 regs->count = count;
220 regs->regs = rzalloc_array(regs, struct ra_reg, count);
221
222 for (i = 0; i < count; i++) {
223 regs->regs[i].conflicts = rzalloc_array(regs->regs, BITSET_WORD,
224 BITSET_WORDS(count));
225 BITSET_SET(regs->regs[i].conflicts, i);
226
227 util_dynarray_init(&regs->regs[i].conflict_list,
228 need_conflict_lists ? regs->regs : NULL);
229 if (need_conflict_lists)
230 util_dynarray_append(&regs->regs[i].conflict_list, unsigned int, i);
231 }
232
233 return regs;
234 }
235
236 /**
237 * The register allocator by default prefers to allocate low register numbers,
238 * since it was written for hardware (gen4/5 Intel) that is limited in its
239 * multithreadedness by the number of registers used in a given shader.
240 *
241 * However, for hardware without that restriction, densely packed register
242 * allocation can put serious constraints on instruction scheduling. This
243 * function tells the allocator to rotate around the registers if possible as
244 * it allocates the nodes.
245 */
246 void
247 ra_set_allocate_round_robin(struct ra_regs *regs)
248 {
249 regs->round_robin = true;
250 }
251
252 static void
253 ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2)
254 {
255 struct ra_reg *reg1 = &regs->regs[r1];
256
257 if (reg1->conflict_list.mem_ctx) {
258 util_dynarray_append(&reg1->conflict_list, unsigned int, r2);
259 }
260 BITSET_SET(reg1->conflicts, r2);
261 }
262
263 void
264 ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2)
265 {
266 if (!BITSET_TEST(regs->regs[r1].conflicts, r2)) {
267 ra_add_conflict_list(regs, r1, r2);
268 ra_add_conflict_list(regs, r2, r1);
269 }
270 }
271
272 /**
273 * Adds a conflict between base_reg and reg, and also between reg and
274 * anything that base_reg conflicts with.
275 *
276 * This can simplify code for setting up multiple register classes
277 * which are aggregates of some base hardware registers, compared to
278 * explicitly using ra_add_reg_conflict.
279 */
280 void
281 ra_add_transitive_reg_conflict(struct ra_regs *regs,
282 unsigned int base_reg, unsigned int reg)
283 {
284 ra_add_reg_conflict(regs, reg, base_reg);
285
286 util_dynarray_foreach(&regs->regs[base_reg].conflict_list, unsigned int,
287 r2p) {
288 ra_add_reg_conflict(regs, reg, *r2p);
289 }
290 }
291
292 /**
293 * Set up conflicts between base_reg and it's two half registers reg0 and
294 * reg1, but take care to not add conflicts between reg0 and reg1.
295 *
296 * This is useful for architectures where full size registers are aliased by
297 * two half size registers (eg 32 bit float and 16 bit float registers).
298 */
299 void
300 ra_add_transitive_reg_pair_conflict(struct ra_regs *regs,
301 unsigned int base_reg, unsigned int reg0, unsigned int reg1)
302 {
303 ra_add_reg_conflict(regs, reg0, base_reg);
304 ra_add_reg_conflict(regs, reg1, base_reg);
305
306 util_dynarray_foreach(&regs->regs[base_reg].conflict_list, unsigned int, i) {
307 unsigned int conflict = *i;
308 if (conflict != reg1)
309 ra_add_reg_conflict(regs, reg0, conflict);
310 if (conflict != reg0)
311 ra_add_reg_conflict(regs, reg1, conflict);
312 }
313 }
314
315 /**
316 * Makes every conflict on the given register transitive. In other words,
317 * every register that conflicts with r will now conflict with every other
318 * register conflicting with r.
319 *
320 * This can simplify code for setting up multiple register classes
321 * which are aggregates of some base hardware registers, compared to
322 * explicitly using ra_add_reg_conflict.
323 */
324 void
325 ra_make_reg_conflicts_transitive(struct ra_regs *regs, unsigned int r)
326 {
327 struct ra_reg *reg = &regs->regs[r];
328 int c;
329
330 BITSET_FOREACH_SET(c, reg->conflicts, regs->count) {
331 struct ra_reg *other = &regs->regs[c];
332 unsigned i;
333 for (i = 0; i < BITSET_WORDS(regs->count); i++)
334 other->conflicts[i] |= reg->conflicts[i];
335 }
336 }
337
338 unsigned int
339 ra_alloc_reg_class(struct ra_regs *regs)
340 {
341 struct ra_class *class;
342
343 regs->classes = reralloc(regs->regs, regs->classes, struct ra_class *,
344 regs->class_count + 1);
345
346 class = rzalloc(regs, struct ra_class);
347 regs->classes[regs->class_count] = class;
348
349 class->regs = rzalloc_array(class, BITSET_WORD, BITSET_WORDS(regs->count));
350
351 return regs->class_count++;
352 }
353
354 void
355 ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r)
356 {
357 struct ra_class *class = regs->classes[c];
358
359 assert(r < regs->count);
360
361 BITSET_SET(class->regs, r);
362 class->p++;
363 }
364
365 /**
366 * Returns true if the register belongs to the given class.
367 */
368 static bool
369 reg_belongs_to_class(unsigned int r, struct ra_class *c)
370 {
371 return BITSET_TEST(c->regs, r);
372 }
373
374 /**
375 * Must be called after all conflicts and register classes have been
376 * set up and before the register set is used for allocation.
377 * To avoid costly q value computation, use the q_values paramater
378 * to pass precomputed q values to this function.
379 */
380 void
381 ra_set_finalize(struct ra_regs *regs, unsigned int **q_values)
382 {
383 unsigned int b, c;
384
385 for (b = 0; b < regs->class_count; b++) {
386 regs->classes[b]->q = ralloc_array(regs, unsigned int, regs->class_count);
387 }
388
389 if (q_values) {
390 for (b = 0; b < regs->class_count; b++) {
391 for (c = 0; c < regs->class_count; c++) {
392 regs->classes[b]->q[c] = q_values[b][c];
393 }
394 }
395 } else {
396 /* Compute, for each class B and C, how many regs of B an
397 * allocation to C could conflict with.
398 */
399 for (b = 0; b < regs->class_count; b++) {
400 for (c = 0; c < regs->class_count; c++) {
401 unsigned int rc;
402 int max_conflicts = 0;
403
404 BITSET_FOREACH_SET(rc, regs->classes[c]->regs, regs->count) {
405 int conflicts = 0;
406
407 util_dynarray_foreach(&regs->regs[rc].conflict_list,
408 unsigned int, rbp) {
409 unsigned int rb = *rbp;
410 if (reg_belongs_to_class(rb, regs->classes[b]))
411 conflicts++;
412 }
413 max_conflicts = MAX2(max_conflicts, conflicts);
414 }
415 regs->classes[b]->q[c] = max_conflicts;
416 }
417 }
418 }
419
420 for (b = 0; b < regs->count; b++) {
421 util_dynarray_fini(&regs->regs[b].conflict_list);
422 }
423 }
424
425 void
426 ra_set_serialize(const struct ra_regs *regs, struct blob *blob)
427 {
428 blob_write_uint32(blob, regs->count);
429 blob_write_uint32(blob, regs->class_count);
430
431 for (unsigned int r = 0; r < regs->count; r++) {
432 struct ra_reg *reg = &regs->regs[r];
433 blob_write_bytes(blob, reg->conflicts, BITSET_WORDS(regs->count) *
434 sizeof(BITSET_WORD));
435 assert(util_dynarray_num_elements(&reg->conflict_list, unsigned int) == 0);
436 }
437
438 for (unsigned int c = 0; c < regs->class_count; c++) {
439 struct ra_class *class = regs->classes[c];
440 blob_write_bytes(blob, class->regs, BITSET_WORDS(regs->count) *
441 sizeof(BITSET_WORD));
442 blob_write_uint32(blob, class->p);
443 blob_write_bytes(blob, class->q, regs->class_count * sizeof(*class->q));
444 }
445
446 blob_write_uint32(blob, regs->round_robin);
447 }
448
449 struct ra_regs *
450 ra_set_deserialize(void *mem_ctx, struct blob_reader *blob)
451 {
452 unsigned int reg_count = blob_read_uint32(blob);
453 unsigned int class_count = blob_read_uint32(blob);
454
455 struct ra_regs *regs = ra_alloc_reg_set(mem_ctx, reg_count, false);
456 assert(regs->count == reg_count);
457
458 for (unsigned int r = 0; r < reg_count; r++) {
459 struct ra_reg *reg = &regs->regs[r];
460 blob_copy_bytes(blob, reg->conflicts, BITSET_WORDS(reg_count) *
461 sizeof(BITSET_WORD));
462 }
463
464 assert(regs->classes == NULL);
465 regs->classes = ralloc_array(regs->regs, struct ra_class *, class_count);
466 regs->class_count = class_count;
467
468 for (unsigned int c = 0; c < class_count; c++) {
469 struct ra_class *class = rzalloc(regs, struct ra_class);
470 regs->classes[c] = class;
471
472 class->regs = ralloc_array(class, BITSET_WORD, BITSET_WORDS(reg_count));
473 blob_copy_bytes(blob, class->regs, BITSET_WORDS(reg_count) *
474 sizeof(BITSET_WORD));
475
476 class->p = blob_read_uint32(blob);
477
478 class->q = ralloc_array(regs->classes[c], unsigned int, class_count);
479 blob_copy_bytes(blob, class->q, class_count * sizeof(*class->q));
480 }
481
482 regs->round_robin = blob_read_uint32(blob);
483
484 return regs;
485 }
486
487 static void
488 ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2)
489 {
490 BITSET_SET(g->nodes[n1].adjacency, n2);
491
492 assert(n1 != n2);
493
494 int n1_class = g->nodes[n1].class;
495 int n2_class = g->nodes[n2].class;
496 g->nodes[n1].q_total += g->regs->classes[n1_class]->q[n2_class];
497
498 util_dynarray_append(&g->nodes[n1].adjacency_list, unsigned int, n2);
499 }
500
501 static void
502 ra_node_remove_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2)
503 {
504 BITSET_CLEAR(g->nodes[n1].adjacency, n2);
505
506 assert(n1 != n2);
507
508 int n1_class = g->nodes[n1].class;
509 int n2_class = g->nodes[n2].class;
510 g->nodes[n1].q_total -= g->regs->classes[n1_class]->q[n2_class];
511
512 util_dynarray_delete_unordered(&g->nodes[n1].adjacency_list, unsigned int,
513 n2);
514 }
515
516 static void
517 ra_realloc_interference_graph(struct ra_graph *g, unsigned int alloc)
518 {
519 if (alloc <= g->alloc)
520 return;
521
522 /* If we always have a whole number of BITSET_WORDs, it makes it much
523 * easier to memset the top of the growing bitsets.
524 */
525 assert(g->alloc % BITSET_WORDBITS == 0);
526 alloc = align64(alloc, BITSET_WORDBITS);
527
528 g->nodes = reralloc(g, g->nodes, struct ra_node, alloc);
529
530 unsigned g_bitset_count = BITSET_WORDS(g->alloc);
531 unsigned bitset_count = BITSET_WORDS(alloc);
532 /* For nodes already in the graph, we just have to grow the adjacency set */
533 for (unsigned i = 0; i < g->alloc; i++) {
534 assert(g->nodes[i].adjacency != NULL);
535 g->nodes[i].adjacency = rerzalloc(g, g->nodes[i].adjacency, BITSET_WORD,
536 g_bitset_count, bitset_count);
537 }
538
539 /* For new nodes, we have to fully initialize them */
540 for (unsigned i = g->alloc; i < alloc; i++) {
541 memset(&g->nodes[i], 0, sizeof(g->nodes[i]));
542 g->nodes[i].adjacency = rzalloc_array(g, BITSET_WORD, bitset_count);
543 util_dynarray_init(&g->nodes[i].adjacency_list, g);
544 g->nodes[i].q_total = 0;
545
546 g->nodes[i].forced_reg = NO_REG;
547 g->nodes[i].reg = NO_REG;
548 }
549
550 /* These are scratch values and don't need to be zeroed. We'll clear them
551 * as part of ra_select() setup.
552 */
553 g->tmp.stack = reralloc(g, g->tmp.stack, unsigned int, alloc);
554 g->tmp.in_stack = reralloc(g, g->tmp.in_stack, BITSET_WORD, bitset_count);
555
556 g->tmp.reg_assigned = reralloc(g, g->tmp.reg_assigned, BITSET_WORD,
557 bitset_count);
558 g->tmp.pq_test = reralloc(g, g->tmp.pq_test, BITSET_WORD, bitset_count);
559 g->tmp.min_q_total = reralloc(g, g->tmp.min_q_total, unsigned int,
560 bitset_count);
561 g->tmp.min_q_node = reralloc(g, g->tmp.min_q_node, unsigned int,
562 bitset_count);
563
564 g->alloc = alloc;
565 }
566
567 struct ra_graph *
568 ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count)
569 {
570 struct ra_graph *g;
571
572 g = rzalloc(NULL, struct ra_graph);
573 g->regs = regs;
574 g->count = count;
575 ra_realloc_interference_graph(g, count);
576
577 return g;
578 }
579
580 void
581 ra_resize_interference_graph(struct ra_graph *g, unsigned int count)
582 {
583 g->count = count;
584 if (count > g->alloc)
585 ra_realloc_interference_graph(g, g->alloc * 2);
586 }
587
588 void ra_set_select_reg_callback(struct ra_graph *g,
589 ra_select_reg_callback callback,
590 void *data)
591 {
592 g->select_reg_callback = callback;
593 g->select_reg_callback_data = data;
594 }
595
596 void
597 ra_set_node_class(struct ra_graph *g,
598 unsigned int n, unsigned int class)
599 {
600 g->nodes[n].class = class;
601 }
602
603 unsigned int
604 ra_get_node_class(struct ra_graph *g,
605 unsigned int n)
606 {
607 return g->nodes[n].class;
608 }
609
610 unsigned int
611 ra_add_node(struct ra_graph *g, unsigned int class)
612 {
613 unsigned int n = g->count;
614 ra_resize_interference_graph(g, g->count + 1);
615
616 ra_set_node_class(g, n, class);
617
618 return n;
619 }
620
621 void
622 ra_add_node_interference(struct ra_graph *g,
623 unsigned int n1, unsigned int n2)
624 {
625 assert(n1 < g->count && n2 < g->count);
626 if (n1 != n2 && !BITSET_TEST(g->nodes[n1].adjacency, n2)) {
627 ra_add_node_adjacency(g, n1, n2);
628 ra_add_node_adjacency(g, n2, n1);
629 }
630 }
631
632 void
633 ra_reset_node_interference(struct ra_graph *g, unsigned int n)
634 {
635 util_dynarray_foreach(&g->nodes[n].adjacency_list, unsigned int, n2p) {
636 ra_node_remove_adjacency(g, *n2p, n);
637 }
638
639 memset(g->nodes[n].adjacency, 0,
640 BITSET_WORDS(g->count) * sizeof(BITSET_WORD));
641 util_dynarray_clear(&g->nodes[n].adjacency_list);
642 }
643
644 static void
645 update_pq_info(struct ra_graph *g, unsigned int n)
646 {
647 int i = n / BITSET_WORDBITS;
648 int n_class = g->nodes[n].class;
649 if (g->nodes[n].tmp.q_total < g->regs->classes[n_class]->p) {
650 BITSET_SET(g->tmp.pq_test, n);
651 } else if (g->tmp.min_q_total[i] != UINT_MAX) {
652 /* Only update min_q_total and min_q_node if min_q_total != UINT_MAX so
653 * that we don't update while we have stale data and accidentally mark
654 * it as non-stale. Also, in order to remain consistent with the old
655 * naive implementation of the algorithm, we do a lexicographical sort
656 * to ensure that we always choose the node with the highest node index.
657 */
658 if (g->nodes[n].tmp.q_total < g->tmp.min_q_total[i] ||
659 (g->nodes[n].tmp.q_total == g->tmp.min_q_total[i] &&
660 n > g->tmp.min_q_node[i])) {
661 g->tmp.min_q_total[i] = g->nodes[n].tmp.q_total;
662 g->tmp.min_q_node[i] = n;
663 }
664 }
665 }
666
667 static void
668 add_node_to_stack(struct ra_graph *g, unsigned int n)
669 {
670 int n_class = g->nodes[n].class;
671
672 assert(!BITSET_TEST(g->tmp.in_stack, n));
673
674 util_dynarray_foreach(&g->nodes[n].adjacency_list, unsigned int, n2p) {
675 unsigned int n2 = *n2p;
676 unsigned int n2_class = g->nodes[n2].class;
677
678 if (!BITSET_TEST(g->tmp.in_stack, n2) &&
679 !BITSET_TEST(g->tmp.reg_assigned, n2)) {
680 assert(g->nodes[n2].tmp.q_total >= g->regs->classes[n2_class]->q[n_class]);
681 g->nodes[n2].tmp.q_total -= g->regs->classes[n2_class]->q[n_class];
682 update_pq_info(g, n2);
683 }
684 }
685
686 g->tmp.stack[g->tmp.stack_count] = n;
687 g->tmp.stack_count++;
688 BITSET_SET(g->tmp.in_stack, n);
689
690 /* Flag the min_q_total for n's block as dirty so it gets recalculated */
691 g->tmp.min_q_total[n / BITSET_WORDBITS] = UINT_MAX;
692 }
693
694 /**
695 * Simplifies the interference graph by pushing all
696 * trivially-colorable nodes into a stack of nodes to be colored,
697 * removing them from the graph, and rinsing and repeating.
698 *
699 * If we encounter a case where we can't push any nodes on the stack, then
700 * we optimistically choose a node and push it on the stack. We heuristically
701 * push the node with the lowest total q value, since it has the fewest
702 * neighbors and therefore is most likely to be allocated.
703 */
704 static void
705 ra_simplify(struct ra_graph *g)
706 {
707 bool progress = true;
708 unsigned int stack_optimistic_start = UINT_MAX;
709
710 /* Figure out the high bit and bit mask for the first iteration of a loop
711 * over BITSET_WORDs.
712 */
713 const unsigned int top_word_high_bit = (g->count - 1) % BITSET_WORDBITS;
714
715 /* Do a quick pre-pass to set things up */
716 g->tmp.stack_count = 0;
717 for (int i = BITSET_WORDS(g->count) - 1, high_bit = top_word_high_bit;
718 i >= 0; i--, high_bit = BITSET_WORDBITS - 1) {
719 g->tmp.in_stack[i] = 0;
720 g->tmp.reg_assigned[i] = 0;
721 g->tmp.pq_test[i] = 0;
722 g->tmp.min_q_total[i] = UINT_MAX;
723 g->tmp.min_q_node[i] = UINT_MAX;
724 for (int j = high_bit; j >= 0; j--) {
725 unsigned int n = i * BITSET_WORDBITS + j;
726 g->nodes[n].reg = g->nodes[n].forced_reg;
727 g->nodes[n].tmp.q_total = g->nodes[n].q_total;
728 if (g->nodes[n].reg != NO_REG)
729 g->tmp.reg_assigned[i] |= BITSET_BIT(j);
730 update_pq_info(g, n);
731 }
732 }
733
734 while (progress) {
735 unsigned int min_q_total = UINT_MAX;
736 unsigned int min_q_node = UINT_MAX;
737
738 progress = false;
739
740 for (int i = BITSET_WORDS(g->count) - 1, high_bit = top_word_high_bit;
741 i >= 0; i--, high_bit = BITSET_WORDBITS - 1) {
742 BITSET_WORD mask = ~(BITSET_WORD)0 >> (31 - high_bit);
743
744 BITSET_WORD skip = g->tmp.in_stack[i] | g->tmp.reg_assigned[i];
745 if (skip == mask)
746 continue;
747
748 BITSET_WORD pq = g->tmp.pq_test[i] & ~skip;
749 if (pq) {
750 /* In this case, we have stuff we can immediately take off the
751 * stack. This also means that we're guaranteed to make progress
752 * and we don't need to bother updating lowest_q_total because we
753 * know we're going to loop again before attempting to do anything
754 * optimistic.
755 */
756 for (int j = high_bit; j >= 0; j--) {
757 if (pq & BITSET_BIT(j)) {
758 unsigned int n = i * BITSET_WORDBITS + j;
759 assert(n < g->count);
760 add_node_to_stack(g, n);
761 /* add_node_to_stack() may update pq_test for this word so
762 * we need to update our local copy.
763 */
764 pq = g->tmp.pq_test[i] & ~skip;
765 progress = true;
766 }
767 }
768 } else if (!progress) {
769 if (g->tmp.min_q_total[i] == UINT_MAX) {
770 /* The min_q_total and min_q_node are dirty because we added
771 * one of these nodes to the stack. It needs to be
772 * recalculated.
773 */
774 for (int j = high_bit; j >= 0; j--) {
775 if (skip & BITSET_BIT(j))
776 continue;
777
778 unsigned int n = i * BITSET_WORDBITS + j;
779 assert(n < g->count);
780 if (g->nodes[n].tmp.q_total < g->tmp.min_q_total[i]) {
781 g->tmp.min_q_total[i] = g->nodes[n].tmp.q_total;
782 g->tmp.min_q_node[i] = n;
783 }
784 }
785 }
786 if (g->tmp.min_q_total[i] < min_q_total) {
787 min_q_node = g->tmp.min_q_node[i];
788 min_q_total = g->tmp.min_q_total[i];
789 }
790 }
791 }
792
793 if (!progress && min_q_total != UINT_MAX) {
794 if (stack_optimistic_start == UINT_MAX)
795 stack_optimistic_start = g->tmp.stack_count;
796
797 add_node_to_stack(g, min_q_node);
798 progress = true;
799 }
800 }
801
802 g->tmp.stack_optimistic_start = stack_optimistic_start;
803 }
804
805 static bool
806 ra_any_neighbors_conflict(struct ra_graph *g, unsigned int n, unsigned int r)
807 {
808 util_dynarray_foreach(&g->nodes[n].adjacency_list, unsigned int, n2p) {
809 unsigned int n2 = *n2p;
810
811 if (!BITSET_TEST(g->tmp.in_stack, n2) &&
812 BITSET_TEST(g->regs->regs[r].conflicts, g->nodes[n2].reg)) {
813 return true;
814 }
815 }
816
817 return false;
818 }
819
820 /* Computes a bitfield of what regs are available for a given register
821 * selection.
822 *
823 * This lets drivers implement a more complicated policy than our simple first
824 * or round robin policies (which don't require knowing the whole bitset)
825 */
826 static bool
827 ra_compute_available_regs(struct ra_graph *g, unsigned int n, BITSET_WORD *regs)
828 {
829 struct ra_class *c = g->regs->classes[g->nodes[n].class];
830
831 /* Populate with the set of regs that are in the node's class. */
832 memcpy(regs, c->regs, BITSET_WORDS(g->regs->count) * sizeof(BITSET_WORD));
833
834 /* Remove any regs that conflict with nodes that we're adjacent to and have
835 * already colored.
836 */
837 util_dynarray_foreach(&g->nodes[n].adjacency_list, unsigned int, n2p) {
838 unsigned int n2 = *n2p;
839 unsigned int r = g->nodes[n2].reg;
840
841 if (!BITSET_TEST(g->tmp.in_stack, n2)) {
842 for (int j = 0; j < BITSET_WORDS(g->regs->count); j++)
843 regs[j] &= ~g->regs->regs[r].conflicts[j];
844 }
845 }
846
847 for (int i = 0; i < BITSET_WORDS(g->regs->count); i++) {
848 if (regs[i])
849 return true;
850 }
851
852 return false;
853 }
854
855 /**
856 * Pops nodes from the stack back into the graph, coloring them with
857 * registers as they go.
858 *
859 * If all nodes were trivially colorable, then this must succeed. If
860 * not (optimistic coloring), then it may return false;
861 */
862 static bool
863 ra_select(struct ra_graph *g)
864 {
865 int start_search_reg = 0;
866 BITSET_WORD *select_regs = NULL;
867
868 if (g->select_reg_callback)
869 select_regs = malloc(BITSET_WORDS(g->regs->count) * sizeof(BITSET_WORD));
870
871 while (g->tmp.stack_count != 0) {
872 unsigned int ri;
873 unsigned int r = -1;
874 int n = g->tmp.stack[g->tmp.stack_count - 1];
875 struct ra_class *c = g->regs->classes[g->nodes[n].class];
876
877 /* set this to false even if we return here so that
878 * ra_get_best_spill_node() considers this node later.
879 */
880 BITSET_CLEAR(g->tmp.in_stack, n);
881
882 if (g->select_reg_callback) {
883 if (!ra_compute_available_regs(g, n, select_regs)) {
884 free(select_regs);
885 return false;
886 }
887
888 r = g->select_reg_callback(n, select_regs, g->select_reg_callback_data);
889 assert(r < g->regs->count);
890 } else {
891 /* Find the lowest-numbered reg which is not used by a member
892 * of the graph adjacent to us.
893 */
894 for (ri = 0; ri < g->regs->count; ri++) {
895 r = (start_search_reg + ri) % g->regs->count;
896 if (!reg_belongs_to_class(r, c))
897 continue;
898
899 if (!ra_any_neighbors_conflict(g, n, r))
900 break;
901 }
902
903 if (ri >= g->regs->count)
904 return false;
905 }
906
907 g->nodes[n].reg = r;
908 g->tmp.stack_count--;
909
910 /* Rotate the starting point except for any nodes above the lowest
911 * optimistically colorable node. The likelihood that we will succeed
912 * at allocating optimistically colorable nodes is highly dependent on
913 * the way that the previous nodes popped off the stack are laid out.
914 * The round-robin strategy increases the fragmentation of the register
915 * file and decreases the number of nearby nodes assigned to the same
916 * color, what increases the likelihood of spilling with respect to the
917 * dense packing strategy.
918 */
919 if (g->regs->round_robin &&
920 g->tmp.stack_count - 1 <= g->tmp.stack_optimistic_start)
921 start_search_reg = r + 1;
922 }
923
924 free(select_regs);
925
926 return true;
927 }
928
929 bool
930 ra_allocate(struct ra_graph *g)
931 {
932 ra_simplify(g);
933 return ra_select(g);
934 }
935
936 unsigned int
937 ra_get_node_reg(struct ra_graph *g, unsigned int n)
938 {
939 if (g->nodes[n].forced_reg != NO_REG)
940 return g->nodes[n].forced_reg;
941 else
942 return g->nodes[n].reg;
943 }
944
945 /**
946 * Forces a node to a specific register. This can be used to avoid
947 * creating a register class containing one node when handling data
948 * that must live in a fixed location and is known to not conflict
949 * with other forced register assignment (as is common with shader
950 * input data). These nodes do not end up in the stack during
951 * ra_simplify(), and thus at ra_select() time it is as if they were
952 * the first popped off the stack and assigned their fixed locations.
953 * Nodes that use this function do not need to be assigned a register
954 * class.
955 *
956 * Must be called before ra_simplify().
957 */
958 void
959 ra_set_node_reg(struct ra_graph *g, unsigned int n, unsigned int reg)
960 {
961 g->nodes[n].forced_reg = reg;
962 }
963
964 static float
965 ra_get_spill_benefit(struct ra_graph *g, unsigned int n)
966 {
967 float benefit = 0;
968 int n_class = g->nodes[n].class;
969
970 /* Define the benefit of eliminating an interference between n, n2
971 * through spilling as q(C, B) / p(C). This is similar to the
972 * "count number of edges" approach of traditional graph coloring,
973 * but takes classes into account.
974 */
975 util_dynarray_foreach(&g->nodes[n].adjacency_list, unsigned int, n2p) {
976 unsigned int n2 = *n2p;
977 unsigned int n2_class = g->nodes[n2].class;
978 benefit += ((float)g->regs->classes[n_class]->q[n2_class] /
979 g->regs->classes[n_class]->p);
980 }
981
982 return benefit;
983 }
984
985 /**
986 * Returns a node number to be spilled according to the cost/benefit using
987 * the pq test, or -1 if there are no spillable nodes.
988 */
989 int
990 ra_get_best_spill_node(struct ra_graph *g)
991 {
992 unsigned int best_node = -1;
993 float best_benefit = 0.0;
994 unsigned int n;
995
996 /* Consider any nodes that we colored successfully or the node we failed to
997 * color for spilling. When we failed to color a node in ra_select(), we
998 * only considered these nodes, so spilling any other ones would not result
999 * in us making progress.
1000 */
1001 for (n = 0; n < g->count; n++) {
1002 float cost = g->nodes[n].spill_cost;
1003 float benefit;
1004
1005 if (cost <= 0.0f)
1006 continue;
1007
1008 if (BITSET_TEST(g->tmp.in_stack, n))
1009 continue;
1010
1011 benefit = ra_get_spill_benefit(g, n);
1012
1013 if (benefit / cost > best_benefit) {
1014 best_benefit = benefit / cost;
1015 best_node = n;
1016 }
1017 }
1018
1019 return best_node;
1020 }
1021
1022 /**
1023 * Only nodes with a spill cost set (cost != 0.0) will be considered
1024 * for register spilling.
1025 */
1026 void
1027 ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost)
1028 {
1029 g->nodes[n].spill_cost = cost;
1030 }