src/mesa/program/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
  33 #include <talloc.h>
  34
  35 #include "main/imports.h"
  36 #include "main/macros.h"
  37 #include "main/mtypes.h"
  38 #include "register_allocate.h"
  39
  40 struct ra_reg {
  41    char *name;
  42    GLboolean *conflicts;
  43 };
  44
  45 struct ra_regs {
  46    struct ra_reg *regs;
  47    unsigned int count;
  48
  49    struct ra_class **classes;
  50    unsigned int class_count;
  51 };
  52
  53 struct ra_class {
  54    GLboolean *regs;
  55
  56    /**
  57     * p_B in Runeson/Nyström paper.
  58     *
  59     * This is "how many regs are in the set."
  60     */
  61    unsigned int p;
  62
  63    /**
  64     * q_B,C in Runeson/Nyström paper.
  65     */
  66    unsigned int *q;
  67 };
  68
  69 struct ra_node {
  70    GLboolean *adjacency;
  71    unsigned int class;
  72    unsigned int adjacency_count;
  73    unsigned int reg;
  74    GLboolean in_stack;
  75    float spill_cost;
  76 };
  77
  78 struct ra_graph {
  79    struct ra_regs *regs;
  80    /**
  81     * the variables that need register allocation.
  82     */
  83    struct ra_node *nodes;
  84    unsigned int count; /**< count of nodes. */
  85
  86    unsigned int *stack;
  87    unsigned int stack_count;
  88 };
  89
  90 struct ra_regs *
  91 ra_alloc_reg_set(unsigned int count)
  92 {
  93    unsigned int i;
  94    struct ra_regs *regs;
  95
  96    regs = talloc_zero(NULL, struct ra_regs);
  97    regs->count = count;
  98    regs->regs = talloc_zero_array(regs, struct ra_reg, count);
  99
 100    for (i = 0; i < count; i++) {
 101       regs->regs[i].conflicts = talloc_zero_array(regs->regs, GLboolean, count);
 102       regs->regs[i].conflicts[i] = GL_TRUE;
 103    }
 104
 105    return regs;
 106 }
 107
 108 void
 109 ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2)
 110 {
 111    regs->regs[r1].conflicts[r2] = GL_TRUE;
 112    regs->regs[r2].conflicts[r1] = GL_TRUE;
 113 }
 114
 115 unsigned int
 116 ra_alloc_reg_class(struct ra_regs *regs)
 117 {
 118    struct ra_class *class;
 119
 120    regs->classes = talloc_realloc(regs, regs->classes,
 121                                   struct ra_class *,
 122                                   regs->class_count + 1);
 123
 124    class = talloc_zero(regs, struct ra_class);
 125    regs->classes[regs->class_count] = class;
 126
 127    class->regs = talloc_zero_array(class, GLboolean, regs->count);
 128
 129    return regs->class_count++;
 130 }
 131
 132 void
 133 ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r)
 134 {
 135    struct ra_class *class = regs->classes[c];
 136
 137    class->regs[r] = GL_TRUE;
 138    class->p++;
 139 }
 140
 141 /**
 142  * Must be called after all conflicts and register classes have been
 143  * set up and before the register set is used for allocation.
 144  */
 145 void
 146 ra_set_finalize(struct ra_regs *regs)
 147 {
 148    unsigned int b, c;
 149
 150    for (b = 0; b < regs->class_count; b++) {
 151       regs->classes[b]->q = talloc_array(regs, unsigned int, regs->class_count);
 152    }
 153
 154    /* Compute, for each class B and C, how many regs of B an
 155     * allocation to C could conflict with.
 156     */
 157    for (b = 0; b < regs->class_count; b++) {
 158       for (c = 0; c < regs->class_count; c++) {
 159          unsigned int rc;
 160          int max_conflicts = 0;
 161
 162          for (rc = 0; rc < regs->count; rc++) {
 163             unsigned int rb;
 164             int conflicts = 0;
 165
 166             if (!regs->classes[c]->regs[rc])
 167                continue;
 168
 169             for (rb = 0; rb < regs->count; rb++) {
 170                if (regs->classes[b]->regs[rb] &&
 171                    regs->regs[rb].conflicts[rc])
 172                   conflicts++;
 173             }
 174             max_conflicts = MAX2(max_conflicts, conflicts);
 175          }
 176          regs->classes[b]->q[c] = max_conflicts;
 177       }
 178    }
 179 }
 180
 181 struct ra_graph *
 182 ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count)
 183 {
 184    struct ra_graph *g;
 185    unsigned int i;
 186
 187    g = talloc_zero(regs, struct ra_graph);
 188    g->regs = regs;
 189    g->nodes = talloc_zero_array(g, struct ra_node, count);
 190    g->count = count;
 191
 192    g->stack = talloc_zero_array(g, unsigned int, count);
 193
 194    for (i = 0; i < count; i++) {
 195       g->nodes[i].adjacency = talloc_zero_array(g, GLboolean, count);
 196       g->nodes[i].adjacency[i] = GL_TRUE;
 197       g->nodes[i].reg = ~0;
 198    }
 199
 200    return g;
 201 }
 202
 203 void
 204 ra_set_node_class(struct ra_graph *g,
 205                   unsigned int n, unsigned int class)
 206 {
 207    g->nodes[n].class = class;
 208 }
 209
 210 void
 211 ra_add_node_interference(struct ra_graph *g,
 212                          unsigned int n1, unsigned int n2)
 213 {
 214    if (g->nodes[n1].adjacency[n2])
 215       return;
 216
 217    g->nodes[n1].adjacency[n2] = GL_TRUE;
 218    g->nodes[n2].adjacency_count++;
 219    g->nodes[n2].adjacency[n1] = GL_TRUE;
 220    g->nodes[n2].adjacency_count++;
 221 }
 222
 223 static GLboolean pq_test(struct ra_graph *g, unsigned int n)
 224 {
 225    unsigned int j;
 226    unsigned int q = 0;
 227    int n_class = g->nodes[n].class;
 228
 229    for (j = 0; j < g->count; j++) {
 230       if (j == n || g->nodes[j].in_stack)
 231          continue;
 232
 233       if (g->nodes[n].adjacency[j]) {
 234          unsigned int j_class = g->nodes[j].class;
 235          q += g->regs->classes[n_class]->q[j_class];
 236       }
 237    }
 238
 239    return q < g->regs->classes[n_class]->p;
 240 }
 241
 242 /**
 243  * Simplifies the interference graph by pushing all
 244  * trivially-colorable nodes into a stack of nodes to be colored,
 245  * removing them from the graph, and rinsing and repeating.
 246  *
 247  * Returns GL_TRUE if all nodes were removed from the graph.  GL_FALSE
 248  * means that either spilling will be required, or optimistic coloring
 249  * should be applied.
 250  */
 251 GLboolean
 252 ra_simplify(struct ra_graph *g)
 253 {
 254    GLboolean progress = GL_TRUE;
 255    int i;
 256
 257    while (progress) {
 258       progress = GL_FALSE;
 259
 260       for (i = g->count - 1; i >= 0; i--) {
 261          if (g->nodes[i].in_stack)
 262             continue;
 263
 264          if (pq_test(g, i)) {
 265             g->stack[g->stack_count] = i;
 266             g->stack_count++;
 267             g->nodes[i].in_stack = GL_TRUE;
 268             progress = GL_TRUE;
 269          }
 270       }
 271    }
 272
 273    for (i = 0; i < g->count; i++) {
 274       if (!g->nodes[i].in_stack)
 275          return GL_FALSE;
 276    }
 277
 278    return GL_TRUE;
 279 }
 280
 281 /**
 282  * Pops nodes from the stack back into the graph, coloring them with
 283  * registers as they go.
 284  *
 285  * If all nodes were trivially colorable, then this must succeed.  If
 286  * not (optimistic coloring), then it may return GL_FALSE;
 287  */
 288 GLboolean
 289 ra_select(struct ra_graph *g)
 290 {
 291    int i;
 292
 293    while (g->stack_count != 0) {
 294       unsigned int r;
 295       int n = g->stack[g->stack_count - 1];
 296       struct ra_class *c = g->regs->classes[g->nodes[n].class];
 297
 298       /* Find the lowest-numbered reg which is not used by a member
 299        * of the graph adjacent to us.
 300        */
 301       for (r = 0; r < g->regs->count; r++) {
 302          if (!c->regs[r])
 303             continue;
 304
 305          /* Check if any of our neighbors conflict with this register choice. */
 306          for (i = 0; i < g->count; i++) {
 307             if (g->nodes[n].adjacency[i] &&
 308                !g->nodes[i].in_stack &&
 309                 g->regs->regs[r].conflicts[g->nodes[i].reg]) {
 310                break;
 311             }
 312          }
 313          if (i == g->count)
 314             break;
 315       }
 316       if (r == g->regs->count)
 317          return GL_FALSE;
 318
 319       g->nodes[n].reg = r;
 320       g->nodes[n].in_stack = GL_FALSE;
 321       g->stack_count--;
 322    }
 323
 324    return GL_TRUE;
 325 }
 326
 327 /**
 328  * Optimistic register coloring: Just push the remaining nodes
 329  * on the stack.  They'll be colored first in ra_select(), and
 330  * if they succeed then the locally-colorable nodes are still
 331  * locally-colorable and the rest of the register allocation
 332  * will succeed.
 333  */
 334 void
 335 ra_optimistic_color(struct ra_graph *g)
 336 {
 337    unsigned int i;
 338
 339    for (i = 0; i < g->count; i++) {
 340       if (g->nodes[i].in_stack)
 341          continue;
 342
 343       g->stack[g->stack_count] = i;
 344       g->stack_count++;
 345       g->nodes[i].in_stack = GL_TRUE;
 346    }
 347 }
 348
 349 GLboolean
 350 ra_allocate_no_spills(struct ra_graph *g)
 351 {
 352    if (!ra_simplify(g)) {
 353       ra_optimistic_color(g);
 354    }
 355    return ra_select(g);
 356 }
 357
 358 unsigned int
 359 ra_get_node_reg(struct ra_graph *g, unsigned int n)
 360 {
 361    return g->nodes[n].reg;
 362 }
 363
 364 static float
 365 ra_get_spill_benefit(struct ra_graph *g, unsigned int n)
 366 {
 367    int j;
 368    float benefit = 0;
 369    int n_class = g->nodes[n].class;
 370
 371    /* Define the benefit of eliminating an interference between n, j
 372     * through spilling as q(C, B) / p(C).  This is similar to the
 373     * "count number of edges" approach of traditional graph coloring,
 374     * but takes classes into account.
 375     */
 376    for (j = 0; j < g->count; j++) {
 377       if (j != n && g->nodes[n].adjacency[j]) {
 378          unsigned int j_class = g->nodes[j].class;
 379          benefit += ((float)g->regs->classes[n_class]->q[j_class] /
 380                      g->regs->classes[n_class]->p);
 381          break;
 382       }
 383    }
 384
 385    return benefit;
 386 }
 387
 388 /**
 389  * Returns a node number to be spilled according to the cost/benefit using
 390  * the pq test, or -1 if there are no spillable nodes.
 391  */
 392 int
 393 ra_get_best_spill_node(struct ra_graph *g)
 394 {
 395    unsigned int best_node = -1;
 396    unsigned int best_benefit = 0.0;
 397    unsigned int n;
 398
 399    for (n = 0; n < g->count; n++) {
 400       float cost = g->nodes[n].spill_cost;
 401       float benefit;
 402
 403       if (cost <= 0.0)
 404          continue;
 405
 406       benefit = ra_get_spill_benefit(g, n);
 407
 408       if (benefit / cost > best_benefit) {
 409          best_benefit = benefit / cost;
 410          best_node = n;
 411       }
 412    }
 413
 414    return best_node;
 415 }
 416
 417 /**
 418  * Only nodes with a spill cost set (cost != 0.0) will be considered
 419  * for register spilling.
 420  */
 421 void
 422 ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost)
 423 {
 424    g->nodes[n].spill_cost = cost;
 425 }