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"
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
;
105 * p(B) in Runeson/Nyström paper.
107 * This is "how many regs are in the set."
112 * q(B,C) (indexed by C, B is this register class) in
113 * Runeson/Nyström paper. This is "how many registers of B could
114 * the worst choice register from C conflict with".
122 * List of which nodes this node interferes with. This should be
123 * symmetric with the other node.
125 BITSET_WORD
*adjacency
;
126 unsigned int *adjacency_list
;
127 unsigned int adjacency_list_size
;
128 unsigned int adjacency_count
;
133 /* Register, if assigned, or NO_REG. */
137 * Set when the node is in the trivially colorable stack. When
138 * set, the adjacency to this node is ignored, to implement the
139 * "remove the edge from the graph" in simplification without
140 * having to actually modify the adjacency_list.
144 /* For an implementation that needs register spilling, this is the
145 * approximate cost of spilling this node.
151 struct ra_regs
*regs
;
153 * the variables that need register allocation.
155 struct ra_node
*nodes
;
156 unsigned int count
; /**< count of nodes. */
159 unsigned int stack_count
;
163 * Creates a set of registers for the allocator.
165 * mem_ctx is a ralloc context for the allocator. The reg set may be freed
166 * using ralloc_free().
169 ra_alloc_reg_set(void *mem_ctx
, unsigned int count
)
172 struct ra_regs
*regs
;
174 regs
= rzalloc(mem_ctx
, struct ra_regs
);
176 regs
->regs
= rzalloc_array(regs
, struct ra_reg
, count
);
178 for (i
= 0; i
< count
; i
++) {
179 regs
->regs
[i
].conflicts
= rzalloc_array(regs
->regs
, GLboolean
, count
);
180 regs
->regs
[i
].conflicts
[i
] = GL_TRUE
;
182 regs
->regs
[i
].conflict_list
= ralloc_array(regs
->regs
, unsigned int, 4);
183 regs
->regs
[i
].conflict_list_size
= 4;
184 regs
->regs
[i
].conflict_list
[0] = i
;
185 regs
->regs
[i
].num_conflicts
= 1;
192 * The register allocator by default prefers to allocate low register numbers,
193 * since it was written for hardware (gen4/5 Intel) that is limited in its
194 * multithreadedness by the number of registers used in a given shader.
196 * However, for hardware without that restriction, densely packed register
197 * allocation can put serious constraints on instruction scheduling. This
198 * function tells the allocator to rotate around the registers if possible as
199 * it allocates the nodes.
202 ra_set_allocate_round_robin(struct ra_regs
*regs
)
204 regs
->round_robin
= true;
208 ra_add_conflict_list(struct ra_regs
*regs
, unsigned int r1
, unsigned int r2
)
210 struct ra_reg
*reg1
= ®s
->regs
[r1
];
212 if (reg1
->conflict_list_size
== reg1
->num_conflicts
) {
213 reg1
->conflict_list_size
*= 2;
214 reg1
->conflict_list
= reralloc(regs
->regs
, reg1
->conflict_list
,
215 unsigned int, reg1
->conflict_list_size
);
217 reg1
->conflict_list
[reg1
->num_conflicts
++] = r2
;
218 reg1
->conflicts
[r2
] = GL_TRUE
;
222 ra_add_reg_conflict(struct ra_regs
*regs
, unsigned int r1
, unsigned int r2
)
224 if (!regs
->regs
[r1
].conflicts
[r2
]) {
225 ra_add_conflict_list(regs
, r1
, r2
);
226 ra_add_conflict_list(regs
, r2
, r1
);
231 * Adds a conflict between base_reg and reg, and also between reg and
232 * anything that base_reg conflicts with.
234 * This can simplify code for setting up multiple register classes
235 * which are aggregates of some base hardware registers, compared to
236 * explicitly using ra_add_reg_conflict.
239 ra_add_transitive_reg_conflict(struct ra_regs
*regs
,
240 unsigned int base_reg
, unsigned int reg
)
244 ra_add_reg_conflict(regs
, reg
, base_reg
);
246 for (i
= 0; i
< regs
->regs
[base_reg
].num_conflicts
; i
++) {
247 ra_add_reg_conflict(regs
, reg
, regs
->regs
[base_reg
].conflict_list
[i
]);
252 ra_alloc_reg_class(struct ra_regs
*regs
)
254 struct ra_class
*class;
256 regs
->classes
= reralloc(regs
->regs
, regs
->classes
, struct ra_class
*,
257 regs
->class_count
+ 1);
259 class = rzalloc(regs
, struct ra_class
);
260 regs
->classes
[regs
->class_count
] = class;
262 class->regs
= rzalloc_array(class, GLboolean
, regs
->count
);
264 return regs
->class_count
++;
268 ra_class_add_reg(struct ra_regs
*regs
, unsigned int c
, unsigned int r
)
270 struct ra_class
*class = regs
->classes
[c
];
272 class->regs
[r
] = GL_TRUE
;
277 * Must be called after all conflicts and register classes have been
278 * set up and before the register set is used for allocation.
279 * To avoid costly q value computation, use the q_values paramater
280 * to pass precomputed q values to this function.
283 ra_set_finalize(struct ra_regs
*regs
, unsigned int **q_values
)
287 for (b
= 0; b
< regs
->class_count
; b
++) {
288 regs
->classes
[b
]->q
= ralloc_array(regs
, unsigned int, regs
->class_count
);
292 for (b
= 0; b
< regs
->class_count
; b
++) {
293 for (c
= 0; c
< regs
->class_count
; c
++) {
294 regs
->classes
[b
]->q
[c
] = q_values
[b
][c
];
300 /* Compute, for each class B and C, how many regs of B an
301 * allocation to C could conflict with.
303 for (b
= 0; b
< regs
->class_count
; b
++) {
304 for (c
= 0; c
< regs
->class_count
; c
++) {
306 int max_conflicts
= 0;
308 for (rc
= 0; rc
< regs
->count
; rc
++) {
312 if (!regs
->classes
[c
]->regs
[rc
])
315 for (i
= 0; i
< regs
->regs
[rc
].num_conflicts
; i
++) {
316 unsigned int rb
= regs
->regs
[rc
].conflict_list
[i
];
317 if (regs
->classes
[b
]->regs
[rb
])
320 max_conflicts
= MAX2(max_conflicts
, conflicts
);
322 regs
->classes
[b
]->q
[c
] = max_conflicts
;
328 ra_add_node_adjacency(struct ra_graph
*g
, unsigned int n1
, unsigned int n2
)
330 BITSET_SET(g
->nodes
[n1
].adjacency
, n2
);
332 if (g
->nodes
[n1
].adjacency_count
>=
333 g
->nodes
[n1
].adjacency_list_size
) {
334 g
->nodes
[n1
].adjacency_list_size
*= 2;
335 g
->nodes
[n1
].adjacency_list
= reralloc(g
, g
->nodes
[n1
].adjacency_list
,
337 g
->nodes
[n1
].adjacency_list_size
);
340 g
->nodes
[n1
].adjacency_list
[g
->nodes
[n1
].adjacency_count
] = n2
;
341 g
->nodes
[n1
].adjacency_count
++;
345 ra_alloc_interference_graph(struct ra_regs
*regs
, unsigned int count
)
350 g
= rzalloc(regs
, struct ra_graph
);
352 g
->nodes
= rzalloc_array(g
, struct ra_node
, count
);
355 g
->stack
= rzalloc_array(g
, unsigned int, count
);
357 for (i
= 0; i
< count
; i
++) {
358 int bitset_count
= BITSET_WORDS(count
);
359 g
->nodes
[i
].adjacency
= rzalloc_array(g
, BITSET_WORD
, bitset_count
);
361 g
->nodes
[i
].adjacency_list_size
= 4;
362 g
->nodes
[i
].adjacency_list
=
363 ralloc_array(g
, unsigned int, g
->nodes
[i
].adjacency_list_size
);
364 g
->nodes
[i
].adjacency_count
= 0;
366 ra_add_node_adjacency(g
, i
, i
);
367 g
->nodes
[i
].reg
= NO_REG
;
374 ra_set_node_class(struct ra_graph
*g
,
375 unsigned int n
, unsigned int class)
377 g
->nodes
[n
].class = class;
381 ra_add_node_interference(struct ra_graph
*g
,
382 unsigned int n1
, unsigned int n2
)
384 if (!BITSET_TEST(g
->nodes
[n1
].adjacency
, n2
)) {
385 ra_add_node_adjacency(g
, n1
, n2
);
386 ra_add_node_adjacency(g
, n2
, n1
);
390 static GLboolean
pq_test(struct ra_graph
*g
, unsigned int n
)
394 int n_class
= g
->nodes
[n
].class;
396 for (j
= 0; j
< g
->nodes
[n
].adjacency_count
; j
++) {
397 unsigned int n2
= g
->nodes
[n
].adjacency_list
[j
];
398 unsigned int n2_class
= g
->nodes
[n2
].class;
400 if (n
!= n2
&& !g
->nodes
[n2
].in_stack
) {
401 q
+= g
->regs
->classes
[n_class
]->q
[n2_class
];
405 return q
< g
->regs
->classes
[n_class
]->p
;
409 * Simplifies the interference graph by pushing all
410 * trivially-colorable nodes into a stack of nodes to be colored,
411 * removing them from the graph, and rinsing and repeating.
413 * Returns GL_TRUE if all nodes were removed from the graph. GL_FALSE
414 * means that either spilling will be required, or optimistic coloring
418 ra_simplify(struct ra_graph
*g
)
420 GLboolean progress
= GL_TRUE
;
426 for (i
= g
->count
- 1; i
>= 0; i
--) {
427 if (g
->nodes
[i
].in_stack
|| g
->nodes
[i
].reg
!= NO_REG
)
431 g
->stack
[g
->stack_count
] = i
;
433 g
->nodes
[i
].in_stack
= GL_TRUE
;
439 for (i
= 0; i
< g
->count
; i
++) {
440 if (!g
->nodes
[i
].in_stack
&& g
->nodes
[i
].reg
== -1)
448 * Pops nodes from the stack back into the graph, coloring them with
449 * registers as they go.
451 * If all nodes were trivially colorable, then this must succeed. If
452 * not (optimistic coloring), then it may return GL_FALSE;
455 ra_select(struct ra_graph
*g
)
458 int start_search_reg
= 0;
460 while (g
->stack_count
!= 0) {
463 int n
= g
->stack
[g
->stack_count
- 1];
464 struct ra_class
*c
= g
->regs
->classes
[g
->nodes
[n
].class];
466 /* Find the lowest-numbered reg which is not used by a member
467 * of the graph adjacent to us.
469 for (ri
= 0; ri
< g
->regs
->count
; ri
++) {
470 r
= (start_search_reg
+ ri
) % g
->regs
->count
;
474 /* Check if any of our neighbors conflict with this register choice. */
475 for (i
= 0; i
< g
->nodes
[n
].adjacency_count
; i
++) {
476 unsigned int n2
= g
->nodes
[n
].adjacency_list
[i
];
478 if (!g
->nodes
[n2
].in_stack
&&
479 g
->regs
->regs
[r
].conflicts
[g
->nodes
[n2
].reg
]) {
483 if (i
== g
->nodes
[n
].adjacency_count
)
486 if (ri
== g
->regs
->count
)
490 g
->nodes
[n
].in_stack
= GL_FALSE
;
493 if (g
->regs
->round_robin
)
494 start_search_reg
= r
+ 1;
501 * Optimistic register coloring: Just push the remaining nodes
502 * on the stack. They'll be colored first in ra_select(), and
503 * if they succeed then the locally-colorable nodes are still
504 * locally-colorable and the rest of the register allocation
508 ra_optimistic_color(struct ra_graph
*g
)
512 for (i
= 0; i
< g
->count
; i
++) {
513 if (g
->nodes
[i
].in_stack
|| g
->nodes
[i
].reg
!= NO_REG
)
516 g
->stack
[g
->stack_count
] = i
;
518 g
->nodes
[i
].in_stack
= GL_TRUE
;
523 ra_allocate_no_spills(struct ra_graph
*g
)
525 if (!ra_simplify(g
)) {
526 ra_optimistic_color(g
);
532 ra_get_node_reg(struct ra_graph
*g
, unsigned int n
)
534 return g
->nodes
[n
].reg
;
538 * Forces a node to a specific register. This can be used to avoid
539 * creating a register class containing one node when handling data
540 * that must live in a fixed location and is known to not conflict
541 * with other forced register assignment (as is common with shader
542 * input data). These nodes do not end up in the stack during
543 * ra_simplify(), and thus at ra_select() time it is as if they were
544 * the first popped off the stack and assigned their fixed locations.
545 * Nodes that use this function do not need to be assigned a register
548 * Must be called before ra_simplify().
551 ra_set_node_reg(struct ra_graph
*g
, unsigned int n
, unsigned int reg
)
553 g
->nodes
[n
].reg
= reg
;
554 g
->nodes
[n
].in_stack
= GL_FALSE
;
558 ra_get_spill_benefit(struct ra_graph
*g
, unsigned int n
)
562 int n_class
= g
->nodes
[n
].class;
564 /* Define the benefit of eliminating an interference between n, n2
565 * through spilling as q(C, B) / p(C). This is similar to the
566 * "count number of edges" approach of traditional graph coloring,
567 * but takes classes into account.
569 for (j
= 0; j
< g
->nodes
[n
].adjacency_count
; j
++) {
570 unsigned int n2
= g
->nodes
[n
].adjacency_list
[j
];
572 unsigned int n2_class
= g
->nodes
[n2
].class;
573 benefit
+= ((float)g
->regs
->classes
[n_class
]->q
[n2_class
] /
574 g
->regs
->classes
[n_class
]->p
);
582 * Returns a node number to be spilled according to the cost/benefit using
583 * the pq test, or -1 if there are no spillable nodes.
586 ra_get_best_spill_node(struct ra_graph
*g
)
588 unsigned int best_node
= -1;
589 float best_benefit
= 0.0;
592 for (n
= 0; n
< g
->count
; n
++) {
593 float cost
= g
->nodes
[n
].spill_cost
;
599 /* Only consider registers for spilling if they are still in the
600 * interference graph (those on the stack have already been proven to be
601 * allocatable without spilling).
603 if (g
->nodes
[n
].in_stack
)
606 benefit
= ra_get_spill_benefit(g
, n
);
608 if (benefit
/ cost
> best_benefit
) {
609 best_benefit
= benefit
/ cost
;
618 * Only nodes with a spill cost set (cost != 0.0) will be considered
619 * for register spilling.
622 ra_set_node_spill_cost(struct ra_graph
*g
, unsigned int n
, float cost
)
624 g
->nodes
[n
].spill_cost
= cost
;