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 "util/bitset.h"
79 #include "register_allocate.h"
84 BITSET_WORD
*conflicts
;
85 unsigned int *conflict_list
;
86 unsigned int conflict_list_size
;
87 unsigned int num_conflicts
;
94 struct ra_class
**classes
;
95 unsigned int class_count
;
102 * Bitset indicating which registers belong to this class.
104 * (If bit N is set, then register N belongs to this class.)
109 * p(B) in Runeson/Nyström paper.
111 * This is "how many regs are in the set."
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".
126 * List of which nodes this node interferes with. This should be
127 * symmetric with the other node.
129 BITSET_WORD
*adjacency
;
130 unsigned int *adjacency_list
;
131 unsigned int adjacency_list_size
;
132 unsigned int adjacency_count
;
137 /* Register, if assigned, or NO_REG. */
141 * The q total, as defined in the Runeson/Nyström paper, for all the
142 * interfering nodes not in the stack.
144 unsigned int q_total
;
146 /* For an implementation that needs register spilling, this is the
147 * approximate cost of spilling this node.
153 struct ra_regs
*regs
;
155 * the variables that need register allocation.
157 struct ra_node
*nodes
;
158 unsigned int count
; /**< count of nodes. */
161 unsigned int stack_count
;
163 /** Bit-set indicating, for each register, if it's in the stack */
164 BITSET_WORD
*in_stack
;
167 * Tracks the start of the set of optimistically-colored registers in the
170 unsigned int stack_optimistic_start
;
172 unsigned int (*select_reg_callback
)(struct ra_graph
*g
, BITSET_WORD
*regs
,
174 void *select_reg_callback_data
;
178 * Creates a set of registers for the allocator.
180 * mem_ctx is a ralloc context for the allocator. The reg set may be freed
181 * using ralloc_free().
184 ra_alloc_reg_set(void *mem_ctx
, unsigned int count
, bool need_conflict_lists
)
187 struct ra_regs
*regs
;
189 regs
= rzalloc(mem_ctx
, struct ra_regs
);
191 regs
->regs
= rzalloc_array(regs
, struct ra_reg
, count
);
193 for (i
= 0; i
< count
; i
++) {
194 regs
->regs
[i
].conflicts
= rzalloc_array(regs
->regs
, BITSET_WORD
,
195 BITSET_WORDS(count
));
196 BITSET_SET(regs
->regs
[i
].conflicts
, i
);
198 if (need_conflict_lists
) {
199 regs
->regs
[i
].conflict_list
= ralloc_array(regs
->regs
,
201 regs
->regs
[i
].conflict_list_size
= 4;
202 regs
->regs
[i
].conflict_list
[0] = i
;
204 regs
->regs
[i
].conflict_list
= NULL
;
205 regs
->regs
[i
].conflict_list_size
= 0;
207 regs
->regs
[i
].num_conflicts
= 1;
214 * The register allocator by default prefers to allocate low register numbers,
215 * since it was written for hardware (gen4/5 Intel) that is limited in its
216 * multithreadedness by the number of registers used in a given shader.
218 * However, for hardware without that restriction, densely packed register
219 * allocation can put serious constraints on instruction scheduling. This
220 * function tells the allocator to rotate around the registers if possible as
221 * it allocates the nodes.
224 ra_set_allocate_round_robin(struct ra_regs
*regs
)
226 regs
->round_robin
= true;
230 ra_add_conflict_list(struct ra_regs
*regs
, unsigned int r1
, unsigned int r2
)
232 struct ra_reg
*reg1
= ®s
->regs
[r1
];
234 if (reg1
->conflict_list
) {
235 if (reg1
->conflict_list_size
== reg1
->num_conflicts
) {
236 reg1
->conflict_list_size
*= 2;
237 reg1
->conflict_list
= reralloc(regs
->regs
, reg1
->conflict_list
,
238 unsigned int, reg1
->conflict_list_size
);
240 reg1
->conflict_list
[reg1
->num_conflicts
++] = r2
;
242 BITSET_SET(reg1
->conflicts
, r2
);
246 ra_add_reg_conflict(struct ra_regs
*regs
, unsigned int r1
, unsigned int r2
)
248 if (!BITSET_TEST(regs
->regs
[r1
].conflicts
, r2
)) {
249 ra_add_conflict_list(regs
, r1
, r2
);
250 ra_add_conflict_list(regs
, r2
, r1
);
255 * Adds a conflict between base_reg and reg, and also between reg and
256 * anything that base_reg conflicts with.
258 * This can simplify code for setting up multiple register classes
259 * which are aggregates of some base hardware registers, compared to
260 * explicitly using ra_add_reg_conflict.
263 ra_add_transitive_reg_conflict(struct ra_regs
*regs
,
264 unsigned int base_reg
, unsigned int reg
)
268 ra_add_reg_conflict(regs
, reg
, base_reg
);
270 for (i
= 0; i
< regs
->regs
[base_reg
].num_conflicts
; i
++) {
271 ra_add_reg_conflict(regs
, reg
, regs
->regs
[base_reg
].conflict_list
[i
]);
276 * Makes every conflict on the given register transitive. In other words,
277 * every register that conflicts with r will now conflict with every other
278 * register conflicting with r.
280 * This can simplify code for setting up multiple register classes
281 * which are aggregates of some base hardware registers, compared to
282 * explicitly using ra_add_reg_conflict.
285 ra_make_reg_conflicts_transitive(struct ra_regs
*regs
, unsigned int r
)
287 struct ra_reg
*reg
= ®s
->regs
[r
];
291 BITSET_FOREACH_SET(c
, tmp
, reg
->conflicts
, regs
->count
) {
292 struct ra_reg
*other
= ®s
->regs
[c
];
294 for (i
= 0; i
< BITSET_WORDS(regs
->count
); i
++)
295 other
->conflicts
[i
] |= reg
->conflicts
[i
];
300 ra_alloc_reg_class(struct ra_regs
*regs
)
302 struct ra_class
*class;
304 regs
->classes
= reralloc(regs
->regs
, regs
->classes
, struct ra_class
*,
305 regs
->class_count
+ 1);
307 class = rzalloc(regs
, struct ra_class
);
308 regs
->classes
[regs
->class_count
] = class;
310 class->regs
= rzalloc_array(class, BITSET_WORD
, BITSET_WORDS(regs
->count
));
312 return regs
->class_count
++;
316 ra_class_add_reg(struct ra_regs
*regs
, unsigned int c
, unsigned int r
)
318 struct ra_class
*class = regs
->classes
[c
];
320 BITSET_SET(class->regs
, r
);
325 * Returns true if the register belongs to the given class.
328 reg_belongs_to_class(unsigned int r
, struct ra_class
*c
)
330 return BITSET_TEST(c
->regs
, r
);
334 * Must be called after all conflicts and register classes have been
335 * set up and before the register set is used for allocation.
336 * To avoid costly q value computation, use the q_values paramater
337 * to pass precomputed q values to this function.
340 ra_set_finalize(struct ra_regs
*regs
, unsigned int **q_values
)
344 for (b
= 0; b
< regs
->class_count
; b
++) {
345 regs
->classes
[b
]->q
= ralloc_array(regs
, unsigned int, regs
->class_count
);
349 for (b
= 0; b
< regs
->class_count
; b
++) {
350 for (c
= 0; c
< regs
->class_count
; c
++) {
351 regs
->classes
[b
]->q
[c
] = q_values
[b
][c
];
355 /* Compute, for each class B and C, how many regs of B an
356 * allocation to C could conflict with.
358 for (b
= 0; b
< regs
->class_count
; b
++) {
359 for (c
= 0; c
< regs
->class_count
; c
++) {
361 int max_conflicts
= 0;
363 for (rc
= 0; rc
< regs
->count
; rc
++) {
367 if (!reg_belongs_to_class(rc
, regs
->classes
[c
]))
370 for (i
= 0; i
< regs
->regs
[rc
].num_conflicts
; i
++) {
371 unsigned int rb
= regs
->regs
[rc
].conflict_list
[i
];
372 if (reg_belongs_to_class(rb
, regs
->classes
[b
]))
375 max_conflicts
= MAX2(max_conflicts
, conflicts
);
377 regs
->classes
[b
]->q
[c
] = max_conflicts
;
382 for (b
= 0; b
< regs
->count
; b
++) {
383 ralloc_free(regs
->regs
[b
].conflict_list
);
384 regs
->regs
[b
].conflict_list
= NULL
;
389 ra_add_node_adjacency(struct ra_graph
*g
, unsigned int n1
, unsigned int n2
)
391 BITSET_SET(g
->nodes
[n1
].adjacency
, n2
);
395 int n1_class
= g
->nodes
[n1
].class;
396 int n2_class
= g
->nodes
[n2
].class;
397 g
->nodes
[n1
].q_total
+= g
->regs
->classes
[n1_class
]->q
[n2_class
];
399 if (g
->nodes
[n1
].adjacency_count
>=
400 g
->nodes
[n1
].adjacency_list_size
) {
401 g
->nodes
[n1
].adjacency_list_size
*= 2;
402 g
->nodes
[n1
].adjacency_list
= reralloc(g
, g
->nodes
[n1
].adjacency_list
,
404 g
->nodes
[n1
].adjacency_list_size
);
407 g
->nodes
[n1
].adjacency_list
[g
->nodes
[n1
].adjacency_count
] = n2
;
408 g
->nodes
[n1
].adjacency_count
++;
412 ra_alloc_interference_graph(struct ra_regs
*regs
, unsigned int count
)
417 g
= rzalloc(NULL
, struct ra_graph
);
419 g
->nodes
= rzalloc_array(g
, struct ra_node
, count
);
422 g
->stack
= rzalloc_array(g
, unsigned int, count
);
424 int bitset_count
= BITSET_WORDS(count
);
425 g
->in_stack
= rzalloc_array(g
, BITSET_WORD
, bitset_count
);
427 for (i
= 0; i
< count
; i
++) {
428 g
->nodes
[i
].adjacency
= rzalloc_array(g
, BITSET_WORD
, bitset_count
);
430 g
->nodes
[i
].adjacency_list_size
= 4;
431 g
->nodes
[i
].adjacency_list
=
432 ralloc_array(g
, unsigned int, g
->nodes
[i
].adjacency_list_size
);
433 g
->nodes
[i
].adjacency_count
= 0;
434 g
->nodes
[i
].q_total
= 0;
436 g
->nodes
[i
].reg
= NO_REG
;
442 void ra_set_select_reg_callback(struct ra_graph
*g
,
443 unsigned int (*callback
)(struct ra_graph
*g
,
448 g
->select_reg_callback
= callback
;
449 g
->select_reg_callback_data
= data
;
453 ra_set_node_class(struct ra_graph
*g
,
454 unsigned int n
, unsigned int class)
456 g
->nodes
[n
].class = class;
460 ra_add_node_interference(struct ra_graph
*g
,
461 unsigned int n1
, unsigned int n2
)
463 if (n1
!= n2
&& !BITSET_TEST(g
->nodes
[n1
].adjacency
, n2
)) {
464 ra_add_node_adjacency(g
, n1
, n2
);
465 ra_add_node_adjacency(g
, n2
, n1
);
470 pq_test(struct ra_graph
*g
, unsigned int n
)
472 int n_class
= g
->nodes
[n
].class;
474 return g
->nodes
[n
].q_total
< g
->regs
->classes
[n_class
]->p
;
478 decrement_q(struct ra_graph
*g
, unsigned int n
)
481 int n_class
= g
->nodes
[n
].class;
483 for (i
= 0; i
< g
->nodes
[n
].adjacency_count
; i
++) {
484 unsigned int n2
= g
->nodes
[n
].adjacency_list
[i
];
485 unsigned int n2_class
= g
->nodes
[n2
].class;
487 if (!BITSET_TEST(g
->in_stack
, n2
)) {
488 assert(g
->nodes
[n2
].q_total
>= g
->regs
->classes
[n2_class
]->q
[n_class
]);
489 g
->nodes
[n2
].q_total
-= g
->regs
->classes
[n2_class
]->q
[n_class
];
495 * Simplifies the interference graph by pushing all
496 * trivially-colorable nodes into a stack of nodes to be colored,
497 * removing them from the graph, and rinsing and repeating.
499 * If we encounter a case where we can't push any nodes on the stack, then
500 * we optimistically choose a node and push it on the stack. We heuristically
501 * push the node with the lowest total q value, since it has the fewest
502 * neighbors and therefore is most likely to be allocated.
505 ra_simplify(struct ra_graph
*g
)
507 bool progress
= true;
508 unsigned int stack_optimistic_start
= UINT_MAX
;
512 unsigned int best_optimistic_node
= ~0;
513 unsigned int lowest_q_total
= ~0;
517 for (i
= g
->count
- 1; i
>= 0; i
--) {
518 if (BITSET_TEST(g
->in_stack
, i
) || g
->nodes
[i
].reg
!= NO_REG
)
523 g
->stack
[g
->stack_count
] = i
;
525 BITSET_SET(g
->in_stack
, i
);
527 } else if (!progress
) {
528 /* We only need to do this if we haven't made progress. If we
529 * have made progress, we'll throw the data away and loop again
532 unsigned int new_q_total
= g
->nodes
[i
].q_total
;
533 if (new_q_total
< lowest_q_total
) {
534 best_optimistic_node
= i
;
535 lowest_q_total
= new_q_total
;
540 if (!progress
&& best_optimistic_node
!= ~0U) {
541 if (stack_optimistic_start
== UINT_MAX
)
542 stack_optimistic_start
= g
->stack_count
;
544 decrement_q(g
, best_optimistic_node
);
545 g
->stack
[g
->stack_count
] = best_optimistic_node
;
547 BITSET_SET(g
->in_stack
, best_optimistic_node
);
552 g
->stack_optimistic_start
= stack_optimistic_start
;
556 ra_any_neighbors_conflict(struct ra_graph
*g
, unsigned int n
, unsigned int r
)
560 for (i
= 0; i
< g
->nodes
[n
].adjacency_count
; i
++) {
561 unsigned int n2
= g
->nodes
[n
].adjacency_list
[i
];
563 if (!BITSET_TEST(g
->in_stack
, n2
) &&
564 BITSET_TEST(g
->regs
->regs
[r
].conflicts
, g
->nodes
[n2
].reg
)) {
572 /* Computes a bitfield of what regs are available for a given register
575 * This lets drivers implement a more complicated policy than our simple first
576 * or round robin policies (which don't require knowing the whole bitset)
579 ra_compute_available_regs(struct ra_graph
*g
, unsigned int n
, BITSET_WORD
*regs
)
581 struct ra_class
*c
= g
->regs
->classes
[g
->nodes
[n
].class];
583 /* Populate with the set of regs that are in the node's class. */
584 memcpy(regs
, c
->regs
, BITSET_WORDS(g
->regs
->count
) * sizeof(BITSET_WORD
));
586 /* Remove any regs that conflict with nodes that we're adjacent to and have
589 for (int i
= 0; i
< g
->nodes
[n
].adjacency_count
; i
++) {
590 unsigned int n2
= g
->nodes
[n
].adjacency_list
[i
];
591 unsigned int r
= g
->nodes
[n2
].reg
;
593 if (!BITSET_TEST(g
->in_stack
, n2
)) {
594 for (int j
= 0; j
< BITSET_WORDS(g
->regs
->count
); j
++)
595 regs
[j
] &= ~g
->regs
->regs
[r
].conflicts
[j
];
599 for (int i
= 0; i
< BITSET_WORDS(g
->regs
->count
); i
++) {
608 * Pops nodes from the stack back into the graph, coloring them with
609 * registers as they go.
611 * If all nodes were trivially colorable, then this must succeed. If
612 * not (optimistic coloring), then it may return false;
615 ra_select(struct ra_graph
*g
)
617 int start_search_reg
= 0;
618 BITSET_WORD
*select_regs
= NULL
;
620 if (g
->select_reg_callback
)
621 select_regs
= malloc(BITSET_WORDS(g
->regs
->count
) * sizeof(BITSET_WORD
));
623 while (g
->stack_count
!= 0) {
626 int n
= g
->stack
[g
->stack_count
- 1];
627 struct ra_class
*c
= g
->regs
->classes
[g
->nodes
[n
].class];
629 /* set this to false even if we return here so that
630 * ra_get_best_spill_node() considers this node later.
632 BITSET_CLEAR(g
->in_stack
, n
);
634 if (g
->select_reg_callback
) {
635 if (!ra_compute_available_regs(g
, n
, select_regs
)) {
640 r
= g
->select_reg_callback(g
, select_regs
, g
->select_reg_callback_data
);
642 /* Find the lowest-numbered reg which is not used by a member
643 * of the graph adjacent to us.
645 for (ri
= 0; ri
< g
->regs
->count
; ri
++) {
646 r
= (start_search_reg
+ ri
) % g
->regs
->count
;
647 if (!reg_belongs_to_class(r
, c
))
650 if (!ra_any_neighbors_conflict(g
, n
, r
))
654 if (ri
>= g
->regs
->count
)
661 /* Rotate the starting point except for any nodes above the lowest
662 * optimistically colorable node. The likelihood that we will succeed
663 * at allocating optimistically colorable nodes is highly dependent on
664 * the way that the previous nodes popped off the stack are laid out.
665 * The round-robin strategy increases the fragmentation of the register
666 * file and decreases the number of nearby nodes assigned to the same
667 * color, what increases the likelihood of spilling with respect to the
668 * dense packing strategy.
670 if (g
->regs
->round_robin
&&
671 g
->stack_count
- 1 <= g
->stack_optimistic_start
)
672 start_search_reg
= r
+ 1;
681 ra_allocate(struct ra_graph
*g
)
688 ra_get_node_reg(struct ra_graph
*g
, unsigned int n
)
690 return g
->nodes
[n
].reg
;
694 * Forces a node to a specific register. This can be used to avoid
695 * creating a register class containing one node when handling data
696 * that must live in a fixed location and is known to not conflict
697 * with other forced register assignment (as is common with shader
698 * input data). These nodes do not end up in the stack during
699 * ra_simplify(), and thus at ra_select() time it is as if they were
700 * the first popped off the stack and assigned their fixed locations.
701 * Nodes that use this function do not need to be assigned a register
704 * Must be called before ra_simplify().
707 ra_set_node_reg(struct ra_graph
*g
, unsigned int n
, unsigned int reg
)
709 g
->nodes
[n
].reg
= reg
;
710 BITSET_CLEAR(g
->in_stack
, n
);
714 ra_get_spill_benefit(struct ra_graph
*g
, unsigned int n
)
718 int n_class
= g
->nodes
[n
].class;
720 /* Define the benefit of eliminating an interference between n, n2
721 * through spilling as q(C, B) / p(C). This is similar to the
722 * "count number of edges" approach of traditional graph coloring,
723 * but takes classes into account.
725 for (j
= 0; j
< g
->nodes
[n
].adjacency_count
; j
++) {
726 unsigned int n2
= g
->nodes
[n
].adjacency_list
[j
];
727 unsigned int n2_class
= g
->nodes
[n2
].class;
728 benefit
+= ((float)g
->regs
->classes
[n_class
]->q
[n2_class
] /
729 g
->regs
->classes
[n_class
]->p
);
736 * Returns a node number to be spilled according to the cost/benefit using
737 * the pq test, or -1 if there are no spillable nodes.
740 ra_get_best_spill_node(struct ra_graph
*g
)
742 unsigned int best_node
= -1;
743 float best_benefit
= 0.0;
746 /* Consider any nodes that we colored successfully or the node we failed to
747 * color for spilling. When we failed to color a node in ra_select(), we
748 * only considered these nodes, so spilling any other ones would not result
749 * in us making progress.
751 for (n
= 0; n
< g
->count
; n
++) {
752 float cost
= g
->nodes
[n
].spill_cost
;
758 if (BITSET_TEST(g
->in_stack
, n
))
761 benefit
= ra_get_spill_benefit(g
, n
);
763 if (benefit
/ cost
> best_benefit
) {
764 best_benefit
= benefit
/ cost
;
773 * Only nodes with a spill cost set (cost != 0.0) will be considered
774 * for register spilling.
777 ra_set_node_spill_cost(struct ra_graph
*g
, unsigned int n
, float cost
)
779 g
->nodes
[n
].spill_cost
= cost
;