2 * Copyright © 2010 Intel Corporation
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:
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
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
24 * Eric Anholt <eric@anholt.net>
28 /** @file register_allocate.c
30 * Graph-coloring register allocator.
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).
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.
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.
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.
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".
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().
76 #include "main/imports.h"
77 #include "main/macros.h"
78 #include "main/mtypes.h"
79 #include "main/bitset.h"
80 #include "register_allocate.h"
85 BITSET_WORD *conflicts;
86 unsigned int *conflict_list;
87 unsigned int conflict_list_size;
88 unsigned int num_conflicts;
95 struct ra_class **classes;
96 unsigned int class_count;
103 * Bitset indicating which registers belong to this class.
105 * (If bit N is set, then register N belongs to this class.)
110 * p(B) in Runeson/Nyström paper.
112 * This is "how many regs are in the set."
117 * q(B,C) (indexed by C, B is this register class) in
118 * Runeson/Nyström paper. This is "how many registers of B could
119 * the worst choice register from C conflict with".
127 * List of which nodes this node interferes with. This should be
128 * symmetric with the other node.
130 BITSET_WORD *adjacency;
131 unsigned int *adjacency_list;
132 unsigned int adjacency_list_size;
133 unsigned int adjacency_count;
138 /* Register, if assigned, or NO_REG. */
142 * Set when the node is in the trivially colorable stack. When
143 * set, the adjacency to this node is ignored, to implement the
144 * "remove the edge from the graph" in simplification without
145 * having to actually modify the adjacency_list.
150 * The q total, as defined in the Runeson/Nyström paper, for all the
151 * interfering nodes not in the stack.
153 unsigned int q_total;
155 /* For an implementation that needs register spilling, this is the
156 * approximate cost of spilling this node.
162 struct ra_regs *regs;
164 * the variables that need register allocation.
166 struct ra_node *nodes;
167 unsigned int count; /**< count of nodes. */
170 unsigned int stack_count;
174 * Creates a set of registers for the allocator.
176 * mem_ctx is a ralloc context for the allocator. The reg set may be freed
177 * using ralloc_free().
180 ra_alloc_reg_set(void *mem_ctx, unsigned int count)
183 struct ra_regs *regs;
185 regs = rzalloc(mem_ctx, struct ra_regs);
187 regs->regs = rzalloc_array(regs, struct ra_reg, count);
189 for (i = 0; i < count; i++) {
190 regs->regs[i].conflicts = rzalloc_array(regs->regs, BITSET_WORD,
191 BITSET_WORDS(count));
192 BITSET_SET(regs->regs[i].conflicts, i);
194 regs->regs[i].conflict_list = ralloc_array(regs->regs, unsigned int, 4);
195 regs->regs[i].conflict_list_size = 4;
196 regs->regs[i].conflict_list[0] = i;
197 regs->regs[i].num_conflicts = 1;
204 * The register allocator by default prefers to allocate low register numbers,
205 * since it was written for hardware (gen4/5 Intel) that is limited in its
206 * multithreadedness by the number of registers used in a given shader.
208 * However, for hardware without that restriction, densely packed register
209 * allocation can put serious constraints on instruction scheduling. This
210 * function tells the allocator to rotate around the registers if possible as
211 * it allocates the nodes.
214 ra_set_allocate_round_robin(struct ra_regs *regs)
216 regs->round_robin = true;
220 ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2)
222 struct ra_reg *reg1 = ®s->regs[r1];
224 if (reg1->conflict_list_size == reg1->num_conflicts) {
225 reg1->conflict_list_size *= 2;
226 reg1->conflict_list = reralloc(regs->regs, reg1->conflict_list,
227 unsigned int, reg1->conflict_list_size);
229 reg1->conflict_list[reg1->num_conflicts++] = r2;
230 BITSET_SET(reg1->conflicts, r2);
234 ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2)
236 if (!BITSET_TEST(regs->regs[r1].conflicts, r2)) {
237 ra_add_conflict_list(regs, r1, r2);
238 ra_add_conflict_list(regs, r2, r1);
243 * Adds a conflict between base_reg and reg, and also between reg and
244 * anything that base_reg conflicts with.
246 * This can simplify code for setting up multiple register classes
247 * which are aggregates of some base hardware registers, compared to
248 * explicitly using ra_add_reg_conflict.
251 ra_add_transitive_reg_conflict(struct ra_regs *regs,
252 unsigned int base_reg, unsigned int reg)
256 ra_add_reg_conflict(regs, reg, base_reg);
258 for (i = 0; i < regs->regs[base_reg].num_conflicts; i++) {
259 ra_add_reg_conflict(regs, reg, regs->regs[base_reg].conflict_list[i]);
264 ra_alloc_reg_class(struct ra_regs *regs)
266 struct ra_class *class;
268 regs->classes = reralloc(regs->regs, regs->classes, struct ra_class *,
269 regs->class_count + 1);
271 class = rzalloc(regs, struct ra_class);
272 regs->classes[regs->class_count] = class;
274 class->regs = rzalloc_array(class, BITSET_WORD, BITSET_WORDS(regs->count));
276 return regs->class_count++;
280 ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r)
282 struct ra_class *class = regs->classes[c];
284 BITSET_SET(class->regs, r);
289 * Returns true if the register belongs to the given class.
292 reg_belongs_to_class(unsigned int r, struct ra_class *c)
294 return BITSET_TEST(c->regs, r);
298 * Must be called after all conflicts and register classes have been
299 * set up and before the register set is used for allocation.
300 * To avoid costly q value computation, use the q_values paramater
301 * to pass precomputed q values to this function.
304 ra_set_finalize(struct ra_regs *regs, unsigned int **q_values)
308 for (b = 0; b < regs->class_count; b++) {
309 regs->classes[b]->q = ralloc_array(regs, unsigned int, regs->class_count);
313 for (b = 0; b < regs->class_count; b++) {
314 for (c = 0; c < regs->class_count; c++) {
315 regs->classes[b]->q[c] = q_values[b][c];
321 /* Compute, for each class B and C, how many regs of B an
322 * allocation to C could conflict with.
324 for (b = 0; b < regs->class_count; b++) {
325 for (c = 0; c < regs->class_count; c++) {
327 int max_conflicts = 0;
329 for (rc = 0; rc < regs->count; rc++) {
333 if (!reg_belongs_to_class(rc, regs->classes[c]))
336 for (i = 0; i < regs->regs[rc].num_conflicts; i++) {
337 unsigned int rb = regs->regs[rc].conflict_list[i];
338 if (BITSET_TEST(regs->classes[b]->regs, rb))
341 max_conflicts = MAX2(max_conflicts, conflicts);
343 regs->classes[b]->q[c] = max_conflicts;
349 ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2)
351 BITSET_SET(g->nodes[n1].adjacency, n2);
354 int n1_class = g->nodes[n1].class;
355 int n2_class = g->nodes[n2].class;
356 g->nodes[n1].q_total += g->regs->classes[n1_class]->q[n2_class];
359 if (g->nodes[n1].adjacency_count >=
360 g->nodes[n1].adjacency_list_size) {
361 g->nodes[n1].adjacency_list_size *= 2;
362 g->nodes[n1].adjacency_list = reralloc(g, g->nodes[n1].adjacency_list,
364 g->nodes[n1].adjacency_list_size);
367 g->nodes[n1].adjacency_list[g->nodes[n1].adjacency_count] = n2;
368 g->nodes[n1].adjacency_count++;
372 ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count)
377 g = rzalloc(regs, struct ra_graph);
379 g->nodes = rzalloc_array(g, struct ra_node, count);
382 g->stack = rzalloc_array(g, unsigned int, count);
384 for (i = 0; i < count; i++) {
385 int bitset_count = BITSET_WORDS(count);
386 g->nodes[i].adjacency = rzalloc_array(g, BITSET_WORD, bitset_count);
388 g->nodes[i].adjacency_list_size = 4;
389 g->nodes[i].adjacency_list =
390 ralloc_array(g, unsigned int, g->nodes[i].adjacency_list_size);
391 g->nodes[i].adjacency_count = 0;
392 g->nodes[i].q_total = 0;
394 ra_add_node_adjacency(g, i, i);
395 g->nodes[i].reg = NO_REG;
402 ra_set_node_class(struct ra_graph *g,
403 unsigned int n, unsigned int class)
405 g->nodes[n].class = class;
409 ra_add_node_interference(struct ra_graph *g,
410 unsigned int n1, unsigned int n2)
412 if (!BITSET_TEST(g->nodes[n1].adjacency, n2)) {
413 ra_add_node_adjacency(g, n1, n2);
414 ra_add_node_adjacency(g, n2, n1);
419 pq_test(struct ra_graph *g, unsigned int n)
421 int n_class = g->nodes[n].class;
423 return g->nodes[n].q_total < g->regs->classes[n_class]->p;
427 decrement_q(struct ra_graph *g, unsigned int n)
430 int n_class = g->nodes[n].class;
432 for (i = 0; i < g->nodes[n].adjacency_count; i++) {
433 unsigned int n2 = g->nodes[n].adjacency_list[i];
434 unsigned int n2_class = g->nodes[n2].class;
436 if (n != n2 && !g->nodes[n2].in_stack) {
437 assert(g->nodes[n2].q_total >= g->regs->classes[n2_class]->q[n_class]);
438 g->nodes[n2].q_total -= g->regs->classes[n2_class]->q[n_class];
444 * Simplifies the interference graph by pushing all
445 * trivially-colorable nodes into a stack of nodes to be colored,
446 * removing them from the graph, and rinsing and repeating.
448 * If we encounter a case where we can't push any nodes on the stack, then
449 * we optimistically choose a node and push it on the stack. We heuristically
450 * push the node with the lowest total q value, since it has the fewest
451 * neighbors and therefore is most likely to be allocated.
454 ra_simplify(struct ra_graph *g)
456 bool progress = true;
460 unsigned int best_optimistic_node = ~0;
461 unsigned int lowest_q_total = ~0;
465 for (i = g->count - 1; i >= 0; i--) {
466 if (g->nodes[i].in_stack || g->nodes[i].reg != NO_REG)
471 g->stack[g->stack_count] = i;
473 g->nodes[i].in_stack = true;
476 unsigned int new_q_total = g->nodes[i].q_total;
477 if (new_q_total < lowest_q_total) {
478 best_optimistic_node = i;
479 lowest_q_total = new_q_total;
484 if (!progress && best_optimistic_node != ~0) {
485 decrement_q(g, best_optimistic_node);
486 g->stack[g->stack_count] = best_optimistic_node;
488 g->nodes[best_optimistic_node].in_stack = true;
495 * Pops nodes from the stack back into the graph, coloring them with
496 * registers as they go.
498 * If all nodes were trivially colorable, then this must succeed. If
499 * not (optimistic coloring), then it may return false;
502 ra_select(struct ra_graph *g)
505 int start_search_reg = 0;
507 while (g->stack_count != 0) {
510 int n = g->stack[g->stack_count - 1];
511 struct ra_class *c = g->regs->classes[g->nodes[n].class];
513 /* Find the lowest-numbered reg which is not used by a member
514 * of the graph adjacent to us.
516 for (ri = 0; ri < g->regs->count; ri++) {
517 r = (start_search_reg + ri) % g->regs->count;
518 if (!reg_belongs_to_class(r, c))
521 /* Check if any of our neighbors conflict with this register choice. */
522 for (i = 0; i < g->nodes[n].adjacency_count; i++) {
523 unsigned int n2 = g->nodes[n].adjacency_list[i];
525 if (!g->nodes[n2].in_stack &&
526 BITSET_TEST(g->regs->regs[r].conflicts, g->nodes[n2].reg)) {
530 if (i == g->nodes[n].adjacency_count)
534 /* set this to false even if we return here so that
535 * ra_get_best_spill_node() considers this node later.
537 g->nodes[n].in_stack = false;
539 if (ri == g->regs->count)
545 if (g->regs->round_robin)
546 start_search_reg = r + 1;
553 ra_allocate(struct ra_graph *g)
560 ra_get_node_reg(struct ra_graph *g, unsigned int n)
562 return g->nodes[n].reg;
566 * Forces a node to a specific register. This can be used to avoid
567 * creating a register class containing one node when handling data
568 * that must live in a fixed location and is known to not conflict
569 * with other forced register assignment (as is common with shader
570 * input data). These nodes do not end up in the stack during
571 * ra_simplify(), and thus at ra_select() time it is as if they were
572 * the first popped off the stack and assigned their fixed locations.
573 * Nodes that use this function do not need to be assigned a register
576 * Must be called before ra_simplify().
579 ra_set_node_reg(struct ra_graph *g, unsigned int n, unsigned int reg)
581 g->nodes[n].reg = reg;
582 g->nodes[n].in_stack = false;
586 ra_get_spill_benefit(struct ra_graph *g, unsigned int n)
590 int n_class = g->nodes[n].class;
592 /* Define the benefit of eliminating an interference between n, n2
593 * through spilling as q(C, B) / p(C). This is similar to the
594 * "count number of edges" approach of traditional graph coloring,
595 * but takes classes into account.
597 for (j = 0; j < g->nodes[n].adjacency_count; j++) {
598 unsigned int n2 = g->nodes[n].adjacency_list[j];
600 unsigned int n2_class = g->nodes[n2].class;
601 benefit += ((float)g->regs->classes[n_class]->q[n2_class] /
602 g->regs->classes[n_class]->p);
610 * Returns a node number to be spilled according to the cost/benefit using
611 * the pq test, or -1 if there are no spillable nodes.
614 ra_get_best_spill_node(struct ra_graph *g)
616 unsigned int best_node = -1;
617 float best_benefit = 0.0;
620 /* Consider any nodes that we colored successfully or the node we failed to
621 * color for spilling. When we failed to color a node in ra_select(), we
622 * only considered these nodes, so spilling any other ones would not result
623 * in us making progress.
625 for (n = 0; n < g->count; n++) {
626 float cost = g->nodes[n].spill_cost;
632 if (g->nodes[n].in_stack)
635 benefit = ra_get_spill_benefit(g, n);
637 if (benefit / cost > best_benefit) {
638 best_benefit = benefit / cost;
647 * Only nodes with a spill cost set (cost != 0.0) will be considered
648 * for register spilling.
651 ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost)
653 g->nodes[n].spill_cost = cost;