4eed0b5aab77929599903c8fa5fc4755a255fa78
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
;
162 * Tracks the start of the set of optimistically-colored registers in the
165 * Along with any registers not in the stack (if one called ra_simplify()
166 * and didn't do optimistic coloring), these need to be considered for
169 unsigned int stack_optimistic_start
;
173 * Creates a set of registers for the allocator.
175 * mem_ctx is a ralloc context for the allocator. The reg set may be freed
176 * using ralloc_free().
179 ra_alloc_reg_set(void *mem_ctx
, unsigned int count
)
182 struct ra_regs
*regs
;
184 regs
= rzalloc(mem_ctx
, struct ra_regs
);
186 regs
->regs
= rzalloc_array(regs
, struct ra_reg
, count
);
188 for (i
= 0; i
< count
; i
++) {
189 regs
->regs
[i
].conflicts
= rzalloc_array(regs
->regs
, GLboolean
, count
);
190 regs
->regs
[i
].conflicts
[i
] = GL_TRUE
;
192 regs
->regs
[i
].conflict_list
= ralloc_array(regs
->regs
, unsigned int, 4);
193 regs
->regs
[i
].conflict_list_size
= 4;
194 regs
->regs
[i
].conflict_list
[0] = i
;
195 regs
->regs
[i
].num_conflicts
= 1;
202 * The register allocator by default prefers to allocate low register numbers,
203 * since it was written for hardware (gen4/5 Intel) that is limited in its
204 * multithreadedness by the number of registers used in a given shader.
206 * However, for hardware without that restriction, densely packed register
207 * allocation can put serious constraints on instruction scheduling. This
208 * function tells the allocator to rotate around the registers if possible as
209 * it allocates the nodes.
212 ra_set_allocate_round_robin(struct ra_regs
*regs
)
214 regs
->round_robin
= true;
218 ra_add_conflict_list(struct ra_regs
*regs
, unsigned int r1
, unsigned int r2
)
220 struct ra_reg
*reg1
= ®s
->regs
[r1
];
222 if (reg1
->conflict_list_size
== reg1
->num_conflicts
) {
223 reg1
->conflict_list_size
*= 2;
224 reg1
->conflict_list
= reralloc(regs
->regs
, reg1
->conflict_list
,
225 unsigned int, reg1
->conflict_list_size
);
227 reg1
->conflict_list
[reg1
->num_conflicts
++] = r2
;
228 reg1
->conflicts
[r2
] = GL_TRUE
;
232 ra_add_reg_conflict(struct ra_regs
*regs
, unsigned int r1
, unsigned int r2
)
234 if (!regs
->regs
[r1
].conflicts
[r2
]) {
235 ra_add_conflict_list(regs
, r1
, r2
);
236 ra_add_conflict_list(regs
, r2
, r1
);
241 * Adds a conflict between base_reg and reg, and also between reg and
242 * anything that base_reg conflicts with.
244 * This can simplify code for setting up multiple register classes
245 * which are aggregates of some base hardware registers, compared to
246 * explicitly using ra_add_reg_conflict.
249 ra_add_transitive_reg_conflict(struct ra_regs
*regs
,
250 unsigned int base_reg
, unsigned int reg
)
254 ra_add_reg_conflict(regs
, reg
, base_reg
);
256 for (i
= 0; i
< regs
->regs
[base_reg
].num_conflicts
; i
++) {
257 ra_add_reg_conflict(regs
, reg
, regs
->regs
[base_reg
].conflict_list
[i
]);
262 ra_alloc_reg_class(struct ra_regs
*regs
)
264 struct ra_class
*class;
266 regs
->classes
= reralloc(regs
->regs
, regs
->classes
, struct ra_class
*,
267 regs
->class_count
+ 1);
269 class = rzalloc(regs
, struct ra_class
);
270 regs
->classes
[regs
->class_count
] = class;
272 class->regs
= rzalloc_array(class, GLboolean
, regs
->count
);
274 return regs
->class_count
++;
278 ra_class_add_reg(struct ra_regs
*regs
, unsigned int c
, unsigned int r
)
280 struct ra_class
*class = regs
->classes
[c
];
282 class->regs
[r
] = GL_TRUE
;
287 * Must be called after all conflicts and register classes have been
288 * set up and before the register set is used for allocation.
289 * To avoid costly q value computation, use the q_values paramater
290 * to pass precomputed q values to this function.
293 ra_set_finalize(struct ra_regs
*regs
, unsigned int **q_values
)
297 for (b
= 0; b
< regs
->class_count
; b
++) {
298 regs
->classes
[b
]->q
= ralloc_array(regs
, unsigned int, regs
->class_count
);
302 for (b
= 0; b
< regs
->class_count
; b
++) {
303 for (c
= 0; c
< regs
->class_count
; c
++) {
304 regs
->classes
[b
]->q
[c
] = q_values
[b
][c
];
310 /* Compute, for each class B and C, how many regs of B an
311 * allocation to C could conflict with.
313 for (b
= 0; b
< regs
->class_count
; b
++) {
314 for (c
= 0; c
< regs
->class_count
; c
++) {
316 int max_conflicts
= 0;
318 for (rc
= 0; rc
< regs
->count
; rc
++) {
322 if (!regs
->classes
[c
]->regs
[rc
])
325 for (i
= 0; i
< regs
->regs
[rc
].num_conflicts
; i
++) {
326 unsigned int rb
= regs
->regs
[rc
].conflict_list
[i
];
327 if (regs
->classes
[b
]->regs
[rb
])
330 max_conflicts
= MAX2(max_conflicts
, conflicts
);
332 regs
->classes
[b
]->q
[c
] = max_conflicts
;
338 ra_add_node_adjacency(struct ra_graph
*g
, unsigned int n1
, unsigned int n2
)
340 BITSET_SET(g
->nodes
[n1
].adjacency
, n2
);
342 if (g
->nodes
[n1
].adjacency_count
>=
343 g
->nodes
[n1
].adjacency_list_size
) {
344 g
->nodes
[n1
].adjacency_list_size
*= 2;
345 g
->nodes
[n1
].adjacency_list
= reralloc(g
, g
->nodes
[n1
].adjacency_list
,
347 g
->nodes
[n1
].adjacency_list_size
);
350 g
->nodes
[n1
].adjacency_list
[g
->nodes
[n1
].adjacency_count
] = n2
;
351 g
->nodes
[n1
].adjacency_count
++;
355 ra_alloc_interference_graph(struct ra_regs
*regs
, unsigned int count
)
360 g
= rzalloc(regs
, struct ra_graph
);
362 g
->nodes
= rzalloc_array(g
, struct ra_node
, count
);
365 g
->stack
= rzalloc_array(g
, unsigned int, count
);
367 for (i
= 0; i
< count
; i
++) {
368 int bitset_count
= BITSET_WORDS(count
);
369 g
->nodes
[i
].adjacency
= rzalloc_array(g
, BITSET_WORD
, bitset_count
);
371 g
->nodes
[i
].adjacency_list_size
= 4;
372 g
->nodes
[i
].adjacency_list
=
373 ralloc_array(g
, unsigned int, g
->nodes
[i
].adjacency_list_size
);
374 g
->nodes
[i
].adjacency_count
= 0;
376 ra_add_node_adjacency(g
, i
, i
);
377 g
->nodes
[i
].reg
= NO_REG
;
384 ra_set_node_class(struct ra_graph
*g
,
385 unsigned int n
, unsigned int class)
387 g
->nodes
[n
].class = class;
391 ra_add_node_interference(struct ra_graph
*g
,
392 unsigned int n1
, unsigned int n2
)
394 if (!BITSET_TEST(g
->nodes
[n1
].adjacency
, n2
)) {
395 ra_add_node_adjacency(g
, n1
, n2
);
396 ra_add_node_adjacency(g
, n2
, n1
);
400 static GLboolean
pq_test(struct ra_graph
*g
, unsigned int n
)
404 int n_class
= g
->nodes
[n
].class;
406 for (j
= 0; j
< g
->nodes
[n
].adjacency_count
; j
++) {
407 unsigned int n2
= g
->nodes
[n
].adjacency_list
[j
];
408 unsigned int n2_class
= g
->nodes
[n2
].class;
410 if (n
!= n2
&& !g
->nodes
[n2
].in_stack
) {
411 q
+= g
->regs
->classes
[n_class
]->q
[n2_class
];
415 return q
< g
->regs
->classes
[n_class
]->p
;
419 * Simplifies the interference graph by pushing all
420 * trivially-colorable nodes into a stack of nodes to be colored,
421 * removing them from the graph, and rinsing and repeating.
423 * Returns GL_TRUE if all nodes were removed from the graph. GL_FALSE
424 * means that either spilling will be required, or optimistic coloring
428 ra_simplify(struct ra_graph
*g
)
430 GLboolean progress
= GL_TRUE
;
436 for (i
= g
->count
- 1; i
>= 0; i
--) {
437 if (g
->nodes
[i
].in_stack
|| g
->nodes
[i
].reg
!= NO_REG
)
441 g
->stack
[g
->stack_count
] = i
;
443 g
->nodes
[i
].in_stack
= GL_TRUE
;
449 for (i
= 0; i
< g
->count
; i
++) {
450 if (!g
->nodes
[i
].in_stack
&& g
->nodes
[i
].reg
== -1)
458 * Pops nodes from the stack back into the graph, coloring them with
459 * registers as they go.
461 * If all nodes were trivially colorable, then this must succeed. If
462 * not (optimistic coloring), then it may return GL_FALSE;
465 ra_select(struct ra_graph
*g
)
468 int start_search_reg
= 0;
470 while (g
->stack_count
!= 0) {
473 int n
= g
->stack
[g
->stack_count
- 1];
474 struct ra_class
*c
= g
->regs
->classes
[g
->nodes
[n
].class];
476 /* Find the lowest-numbered reg which is not used by a member
477 * of the graph adjacent to us.
479 for (ri
= 0; ri
< g
->regs
->count
; ri
++) {
480 r
= (start_search_reg
+ ri
) % g
->regs
->count
;
484 /* Check if any of our neighbors conflict with this register choice. */
485 for (i
= 0; i
< g
->nodes
[n
].adjacency_count
; i
++) {
486 unsigned int n2
= g
->nodes
[n
].adjacency_list
[i
];
488 if (!g
->nodes
[n2
].in_stack
&&
489 g
->regs
->regs
[r
].conflicts
[g
->nodes
[n2
].reg
]) {
493 if (i
== g
->nodes
[n
].adjacency_count
)
496 if (ri
== g
->regs
->count
)
500 g
->nodes
[n
].in_stack
= GL_FALSE
;
503 if (g
->regs
->round_robin
)
504 start_search_reg
= r
+ 1;
511 * Optimistic register coloring: Just push the remaining nodes
512 * on the stack. They'll be colored first in ra_select(), and
513 * if they succeed then the locally-colorable nodes are still
514 * locally-colorable and the rest of the register allocation
518 ra_optimistic_color(struct ra_graph
*g
)
522 g
->stack_optimistic_start
= g
->stack_count
;
523 for (i
= 0; i
< g
->count
; i
++) {
524 if (g
->nodes
[i
].in_stack
|| g
->nodes
[i
].reg
!= NO_REG
)
527 g
->stack
[g
->stack_count
] = i
;
529 g
->nodes
[i
].in_stack
= GL_TRUE
;
534 ra_allocate_no_spills(struct ra_graph
*g
)
536 if (!ra_simplify(g
)) {
537 ra_optimistic_color(g
);
543 ra_get_node_reg(struct ra_graph
*g
, unsigned int n
)
545 return g
->nodes
[n
].reg
;
549 * Forces a node to a specific register. This can be used to avoid
550 * creating a register class containing one node when handling data
551 * that must live in a fixed location and is known to not conflict
552 * with other forced register assignment (as is common with shader
553 * input data). These nodes do not end up in the stack during
554 * ra_simplify(), and thus at ra_select() time it is as if they were
555 * the first popped off the stack and assigned their fixed locations.
556 * Nodes that use this function do not need to be assigned a register
559 * Must be called before ra_simplify().
562 ra_set_node_reg(struct ra_graph
*g
, unsigned int n
, unsigned int reg
)
564 g
->nodes
[n
].reg
= reg
;
565 g
->nodes
[n
].in_stack
= GL_FALSE
;
569 ra_get_spill_benefit(struct ra_graph
*g
, unsigned int n
)
573 int n_class
= g
->nodes
[n
].class;
575 /* Define the benefit of eliminating an interference between n, n2
576 * through spilling as q(C, B) / p(C). This is similar to the
577 * "count number of edges" approach of traditional graph coloring,
578 * but takes classes into account.
580 for (j
= 0; j
< g
->nodes
[n
].adjacency_count
; j
++) {
581 unsigned int n2
= g
->nodes
[n
].adjacency_list
[j
];
583 unsigned int n2_class
= g
->nodes
[n2
].class;
584 benefit
+= ((float)g
->regs
->classes
[n_class
]->q
[n2_class
] /
585 g
->regs
->classes
[n_class
]->p
);
593 * Returns a node number to be spilled according to the cost/benefit using
594 * the pq test, or -1 if there are no spillable nodes.
597 ra_get_best_spill_node(struct ra_graph
*g
)
599 unsigned int best_node
= -1;
600 float best_benefit
= 0.0;
603 /* For any registers not in the stack to be colored, consider them for
604 * spilling. This will mostly collect nodes that were being optimistally
605 * colored as part of ra_allocate_no_spills() if we didn't successfully
606 * optimistically color.
608 * It also includes nodes not trivially colorable by ra_simplify() if it
609 * was used directly instead of as part of ra_allocate_no_spills().
611 for (n
= 0; n
< g
->count
; n
++) {
612 float cost
= g
->nodes
[n
].spill_cost
;
618 if (g
->nodes
[n
].in_stack
)
621 benefit
= ra_get_spill_benefit(g
, n
);
623 if (benefit
/ cost
> best_benefit
) {
624 best_benefit
= benefit
/ cost
;
629 /* Also consider spilling any nodes that were set up to be optimistically
630 * colored that we couldn't manage to color in ra_select().
632 for (i
= g
->stack_optimistic_start
; i
< g
->stack_count
; i
++) {
636 cost
= g
->nodes
[n
].spill_cost
;
641 benefit
= ra_get_spill_benefit(g
, n
);
643 if (benefit
/ cost
> best_benefit
) {
644 best_benefit
= benefit
/ cost
;
653 * Only nodes with a spill cost set (cost != 0.0) will be considered
654 * for register spilling.
657 ra_set_node_spill_cost(struct ra_graph
*g
, unsigned int n
, float cost
)
659 g
->nodes
[n
].spill_cost
= cost
;