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 (at) anholt.net> 25 * 26 */ 27 28 /** @file register_allocate.c 29 * 30 * Graph-coloring register allocator. 31 * 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). 36 * 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. 42 * 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. 48 * 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 Nystrm. 54 * 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". 66 * 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(). 71 */ 72 73 #include <stdbool.h> 74 75 #include "ralloc.h" 76 #include "main/imports.h" 77 #include "main/macros.h" 78 #include "main/mtypes.h" 79 #include "util/bitset.h" 80 #include "register_allocate.h" 81 82 #define NO_REG ~0U 83 84 struct ra_reg { 85 BITSET_WORD *conflicts; 86 unsigned int *conflict_list; 87 unsigned int conflict_list_size; 88 unsigned int num_conflicts; 89 }; 90 91 struct ra_regs { 92 struct ra_reg *regs; 93 unsigned int count; 94 95 struct ra_class **classes; 96 unsigned int class_count; 97 98 bool round_robin; 99 }; 100 101 struct ra_class { 102 /** 103 * Bitset indicating which registers belong to this class. 104 * 105 * (If bit N is set, then register N belongs to this class.) 106 */ 107 BITSET_WORD *regs; 108 109 /** 110 * p(B) in Runeson/Nystrm paper. 111 * 112 * This is "how many regs are in the set." 113 */ 114 unsigned int p; 115 116 /** 117 * q(B,C) (indexed by C, B is this register class) in 118 * Runeson/Nystrm paper. This is "how many registers of B could 119 * the worst choice register from C conflict with". 120 */ 121 unsigned int *q; 122 }; 123 124 struct ra_node { 125 /** @{ 126 * 127 * List of which nodes this node interferes with. This should be 128 * symmetric with the other node. 129 */ 130 BITSET_WORD *adjacency; 131 unsigned int *adjacency_list; 132 unsigned int adjacency_list_size; 133 unsigned int adjacency_count; 134 /** @} */ 135 136 unsigned int class; 137 138 /* Register, if assigned, or NO_REG. */ 139 unsigned int reg; 140 141 /** 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. 146 */ 147 bool in_stack; 148 149 /** 150 * The q total, as defined in the Runeson/Nystrm paper, for all the 151 * interfering nodes not in the stack. 152 */ 153 unsigned int q_total; 154 155 /* For an implementation that needs register spilling, this is the 156 * approximate cost of spilling this node. 157 */ 158 float spill_cost; 159 }; 160 161 struct ra_graph { 162 struct ra_regs *regs; 163 /** 164 * the variables that need register allocation. 165 */ 166 struct ra_node *nodes; 167 unsigned int count; /**< count of nodes. */ 168 169 unsigned int *stack; 170 unsigned int stack_count; 171 172 /** 173 * Tracks the start of the set of optimistically-colored registers in the 174 * stack. 175 */ 176 unsigned int stack_optimistic_start; 177 178 unsigned int (*select_reg_callback)(struct ra_graph *g, BITSET_WORD *regs, 179 void *data); 180 void *select_reg_callback_data; 181 }; 182 183 /** 184 * Creates a set of registers for the allocator. 185 * 186 * mem_ctx is a ralloc context for the allocator. The reg set may be freed 187 * using ralloc_free(). 188 */ 189 struct ra_regs * 190 ra_alloc_reg_set(void *mem_ctx, unsigned int count, bool need_conflict_lists) 191 { 192 unsigned int i; 193 struct ra_regs *regs; 194 195 regs = rzalloc(mem_ctx, struct ra_regs); 196 regs->count = count; 197 regs->regs = rzalloc_array(regs, struct ra_reg, count); 198 199 for (i = 0; i < count; i++) { 200 regs->regs[i].conflicts = rzalloc_array(regs->regs, BITSET_WORD, 201 BITSET_WORDS(count)); 202 BITSET_SET(regs->regs[i].conflicts, i); 203 204 if (need_conflict_lists) { 205 regs->regs[i].conflict_list = ralloc_array(regs->regs, 206 unsigned int, 4); 207 regs->regs[i].conflict_list_size = 4; 208 regs->regs[i].conflict_list[0] = i; 209 } else { 210 regs->regs[i].conflict_list = NULL; 211 regs->regs[i].conflict_list_size = 0; 212 } 213 regs->regs[i].num_conflicts = 1; 214 } 215 216 return regs; 217 } 218 219 /** 220 * The register allocator by default prefers to allocate low register numbers, 221 * since it was written for hardware (gen4/5 Intel) that is limited in its 222 * multithreadedness by the number of registers used in a given shader. 223 * 224 * However, for hardware without that restriction, densely packed register 225 * allocation can put serious constraints on instruction scheduling. This 226 * function tells the allocator to rotate around the registers if possible as 227 * it allocates the nodes. 228 */ 229 void 230 ra_set_allocate_round_robin(struct ra_regs *regs) 231 { 232 regs->round_robin = true; 233 } 234 235 static void 236 ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2) 237 { 238 struct ra_reg *reg1 = ®s->regs[r1]; 239 240 if (reg1->conflict_list) { 241 if (reg1->conflict_list_size == reg1->num_conflicts) { 242 reg1->conflict_list_size *= 2; 243 reg1->conflict_list = reralloc(regs->regs, reg1->conflict_list, 244 unsigned int, reg1->conflict_list_size); 245 } 246 reg1->conflict_list[reg1->num_conflicts++] = r2; 247 } 248 BITSET_SET(reg1->conflicts, r2); 249 } 250 251 void 252 ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2) 253 { 254 if (!BITSET_TEST(regs->regs[r1].conflicts, r2)) { 255 ra_add_conflict_list(regs, r1, r2); 256 ra_add_conflict_list(regs, r2, r1); 257 } 258 } 259 260 /** 261 * Adds a conflict between base_reg and reg, and also between reg and 262 * anything that base_reg conflicts with. 263 * 264 * This can simplify code for setting up multiple register classes 265 * which are aggregates of some base hardware registers, compared to 266 * explicitly using ra_add_reg_conflict. 267 */ 268 void 269 ra_add_transitive_reg_conflict(struct ra_regs *regs, 270 unsigned int base_reg, unsigned int reg) 271 { 272 unsigned int i; 273 274 ra_add_reg_conflict(regs, reg, base_reg); 275 276 for (i = 0; i < regs->regs[base_reg].num_conflicts; i++) { 277 ra_add_reg_conflict(regs, reg, regs->regs[base_reg].conflict_list[i]); 278 } 279 } 280 281 /** 282 * Makes every conflict on the given register transitive. In other words, 283 * every register that conflicts with r will now conflict with every other 284 * register conflicting with r. 285 * 286 * This can simplify code for setting up multiple register classes 287 * which are aggregates of some base hardware registers, compared to 288 * explicitly using ra_add_reg_conflict. 289 */ 290 void 291 ra_make_reg_conflicts_transitive(struct ra_regs *regs, unsigned int r) 292 { 293 struct ra_reg *reg = ®s->regs[r]; 294 BITSET_WORD tmp; 295 int c; 296 297 BITSET_FOREACH_SET(c, tmp, reg->conflicts, regs->count) { 298 struct ra_reg *other = ®s->regs[c]; 299 unsigned i; 300 for (i = 0; i < BITSET_WORDS(regs->count); i++) 301 other->conflicts[i] |= reg->conflicts[i]; 302 } 303 } 304 305 unsigned int 306 ra_alloc_reg_class(struct ra_regs *regs) 307 { 308 struct ra_class *class; 309 310 regs->classes = reralloc(regs->regs, regs->classes, struct ra_class *, 311 regs->class_count + 1); 312 313 class = rzalloc(regs, struct ra_class); 314 regs->classes[regs->class_count] = class; 315 316 class->regs = rzalloc_array(class, BITSET_WORD, BITSET_WORDS(regs->count)); 317 318 return regs->class_count++; 319 } 320 321 void 322 ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r) 323 { 324 struct ra_class *class = regs->classes[c]; 325 326 BITSET_SET(class->regs, r); 327 class->p++; 328 } 329 330 /** 331 * Returns true if the register belongs to the given class. 332 */ 333 static bool 334 reg_belongs_to_class(unsigned int r, struct ra_class *c) 335 { 336 return BITSET_TEST(c->regs, r); 337 } 338 339 /** 340 * Must be called after all conflicts and register classes have been 341 * set up and before the register set is used for allocation. 342 * To avoid costly q value computation, use the q_values paramater 343 * to pass precomputed q values to this function. 344 */ 345 void 346 ra_set_finalize(struct ra_regs *regs, unsigned int **q_values) 347 { 348 unsigned int b, c; 349 350 for (b = 0; b < regs->class_count; b++) { 351 regs->classes[b]->q = ralloc_array(regs, unsigned int, regs->class_count); 352 } 353 354 if (q_values) { 355 for (b = 0; b < regs->class_count; b++) { 356 for (c = 0; c < regs->class_count; c++) { 357 regs->classes[b]->q[c] = q_values[b][c]; 358 } 359 } 360 } else { 361 /* Compute, for each class B and C, how many regs of B an 362 * allocation to C could conflict with. 363 */ 364 for (b = 0; b < regs->class_count; b++) { 365 for (c = 0; c < regs->class_count; c++) { 366 unsigned int rc; 367 int max_conflicts = 0; 368 369 for (rc = 0; rc < regs->count; rc++) { 370 int conflicts = 0; 371 unsigned int i; 372 373 if (!reg_belongs_to_class(rc, regs->classes[c])) 374 continue; 375 376 for (i = 0; i < regs->regs[rc].num_conflicts; i++) { 377 unsigned int rb = regs->regs[rc].conflict_list[i]; 378 if (reg_belongs_to_class(rb, regs->classes[b])) 379 conflicts++; 380 } 381 max_conflicts = MAX2(max_conflicts, conflicts); 382 } 383 regs->classes[b]->q[c] = max_conflicts; 384 } 385 } 386 } 387 388 for (b = 0; b < regs->count; b++) { 389 ralloc_free(regs->regs[b].conflict_list); 390 regs->regs[b].conflict_list = NULL; 391 } 392 } 393 394 static void 395 ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2) 396 { 397 BITSET_SET(g->nodes[n1].adjacency, n2); 398 399 assert(n1 != n2); 400 401 int n1_class = g->nodes[n1].class; 402 int n2_class = g->nodes[n2].class; 403 g->nodes[n1].q_total += g->regs->classes[n1_class]->q[n2_class]; 404 405 if (g->nodes[n1].adjacency_count >= 406 g->nodes[n1].adjacency_list_size) { 407 g->nodes[n1].adjacency_list_size *= 2; 408 g->nodes[n1].adjacency_list = reralloc(g, g->nodes[n1].adjacency_list, 409 unsigned int, 410 g->nodes[n1].adjacency_list_size); 411 } 412 413 g->nodes[n1].adjacency_list[g->nodes[n1].adjacency_count] = n2; 414 g->nodes[n1].adjacency_count++; 415 } 416 417 struct ra_graph * 418 ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count) 419 { 420 struct ra_graph *g; 421 unsigned int i; 422 423 g = rzalloc(NULL, struct ra_graph); 424 g->regs = regs; 425 g->nodes = rzalloc_array(g, struct ra_node, count); 426 g->count = count; 427 428 g->stack = rzalloc_array(g, unsigned int, count); 429 430 for (i = 0; i < count; i++) { 431 int bitset_count = BITSET_WORDS(count); 432 g->nodes[i].adjacency = rzalloc_array(g, BITSET_WORD, bitset_count); 433 434 g->nodes[i].adjacency_list_size = 4; 435 g->nodes[i].adjacency_list = 436 ralloc_array(g, unsigned int, g->nodes[i].adjacency_list_size); 437 g->nodes[i].adjacency_count = 0; 438 g->nodes[i].q_total = 0; 439 440 g->nodes[i].reg = NO_REG; 441 } 442 443 return g; 444 } 445 446 void ra_set_select_reg_callback(struct ra_graph *g, 447 unsigned int (*callback)(struct ra_graph *g, 448 BITSET_WORD *regs, 449 void *data), 450 void *data) 451 { 452 g->select_reg_callback = callback; 453 g->select_reg_callback_data = data; 454 } 455 456 void 457 ra_set_node_class(struct ra_graph *g, 458 unsigned int n, unsigned int class) 459 { 460 g->nodes[n].class = class; 461 } 462 463 void 464 ra_add_node_interference(struct ra_graph *g, 465 unsigned int n1, unsigned int n2) 466 { 467 if (n1 != n2 && !BITSET_TEST(g->nodes[n1].adjacency, n2)) { 468 ra_add_node_adjacency(g, n1, n2); 469 ra_add_node_adjacency(g, n2, n1); 470 } 471 } 472 473 static bool 474 pq_test(struct ra_graph *g, unsigned int n) 475 { 476 int n_class = g->nodes[n].class; 477 478 return g->nodes[n].q_total < g->regs->classes[n_class]->p; 479 } 480 481 static void 482 decrement_q(struct ra_graph *g, unsigned int n) 483 { 484 unsigned int i; 485 int n_class = g->nodes[n].class; 486 487 for (i = 0; i < g->nodes[n].adjacency_count; i++) { 488 unsigned int n2 = g->nodes[n].adjacency_list[i]; 489 unsigned int n2_class = g->nodes[n2].class; 490 491 if (!g->nodes[n2].in_stack) { 492 assert(g->nodes[n2].q_total >= g->regs->classes[n2_class]->q[n_class]); 493 g->nodes[n2].q_total -= g->regs->classes[n2_class]->q[n_class]; 494 } 495 } 496 } 497 498 /** 499 * Simplifies the interference graph by pushing all 500 * trivially-colorable nodes into a stack of nodes to be colored, 501 * removing them from the graph, and rinsing and repeating. 502 * 503 * If we encounter a case where we can't push any nodes on the stack, then 504 * we optimistically choose a node and push it on the stack. We heuristically 505 * push the node with the lowest total q value, since it has the fewest 506 * neighbors and therefore is most likely to be allocated. 507 */ 508 static void 509 ra_simplify(struct ra_graph *g) 510 { 511 bool progress = true; 512 unsigned int stack_optimistic_start = UINT_MAX; 513 int i; 514 515 while (progress) { 516 unsigned int best_optimistic_node = ~0; 517 unsigned int lowest_q_total = ~0; 518 519 progress = false; 520 521 for (i = g->count - 1; i >= 0; i--) { 522 if (g->nodes[i].in_stack || g->nodes[i].reg != NO_REG) 523 continue; 524 525 if (pq_test(g, i)) { 526 decrement_q(g, i); 527 g->stack[g->stack_count] = i; 528 g->stack_count++; 529 g->nodes[i].in_stack = true; 530 progress = true; 531 } else { 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; 536 } 537 } 538 } 539 540 if (!progress && best_optimistic_node != ~0U) { 541 if (stack_optimistic_start == UINT_MAX) 542 stack_optimistic_start = g->stack_count; 543 544 decrement_q(g, best_optimistic_node); 545 g->stack[g->stack_count] = best_optimistic_node; 546 g->stack_count++; 547 g->nodes[best_optimistic_node].in_stack = true; 548 progress = true; 549 } 550 } 551 552 g->stack_optimistic_start = stack_optimistic_start; 553 } 554 555 static bool 556 ra_any_neighbors_conflict(struct ra_graph *g, unsigned int n, unsigned int r) 557 { 558 unsigned int i; 559 560 for (i = 0; i < g->nodes[n].adjacency_count; i++) { 561 unsigned int n2 = g->nodes[n].adjacency_list[i]; 562 563 if (!g->nodes[n2].in_stack && 564 BITSET_TEST(g->regs->regs[r].conflicts, g->nodes[n2].reg)) { 565 return true; 566 } 567 } 568 569 return false; 570 } 571 572 /* Computes a bitfield of what regs are available for a given register 573 * selection. 574 * 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) 577 */ 578 static bool 579 ra_compute_available_regs(struct ra_graph *g, unsigned int n, BITSET_WORD *regs) 580 { 581 struct ra_class *c = g->regs->classes[g->nodes[n].class]; 582 583 /* 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)); 585 586 /* Remove any regs that conflict with nodes that we're adjacent to and have 587 * already colored. 588 */ 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; 592 593 if (!g->nodes[n2].in_stack) { 594 for (int j = 0; j < BITSET_WORDS(g->regs->count); j++) 595 regs[j] &= ~g->regs->regs[r].conflicts[j]; 596 } 597 } 598 599 for (int i = 0; i < BITSET_WORDS(g->regs->count); i++) { 600 if (regs[i]) 601 return true; 602 } 603 604 return false; 605 } 606 607 /** 608 * Pops nodes from the stack back into the graph, coloring them with 609 * registers as they go. 610 * 611 * If all nodes were trivially colorable, then this must succeed. If 612 * not (optimistic coloring), then it may return false; 613 */ 614 static bool 615 ra_select(struct ra_graph *g) 616 { 617 int start_search_reg = 0; 618 BITSET_WORD *select_regs = NULL; 619 620 if (g->select_reg_callback) 621 select_regs = malloc(BITSET_WORDS(g->regs->count) * sizeof(BITSET_WORD)); 622 623 while (g->stack_count != 0) { 624 unsigned int ri; 625 unsigned int r = -1; 626 int n = g->stack[g->stack_count - 1]; 627 struct ra_class *c = g->regs->classes[g->nodes[n].class]; 628 629 /* set this to false even if we return here so that 630 * ra_get_best_spill_node() considers this node later. 631 */ 632 g->nodes[n].in_stack = false; 633 634 if (g->select_reg_callback) { 635 if (!ra_compute_available_regs(g, n, select_regs)) { 636 free(select_regs); 637 return false; 638 } 639 640 r = g->select_reg_callback(g, select_regs, g->select_reg_callback_data); 641 } else { 642 /* Find the lowest-numbered reg which is not used by a member 643 * of the graph adjacent to us. 644 */ 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)) 648 continue; 649 650 if (!ra_any_neighbors_conflict(g, n, r)) 651 break; 652 } 653 654 if (ri >= g->regs->count) 655 return false; 656 } 657 658 g->nodes[n].reg = r; 659 g->stack_count--; 660 661 /* 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. 669 */ 670 if (g->regs->round_robin && 671 g->stack_count - 1 <= g->stack_optimistic_start) 672 start_search_reg = r + 1; 673 } 674 675 free(select_regs); 676 677 return true; 678 } 679 680 bool 681 ra_allocate(struct ra_graph *g) 682 { 683 ra_simplify(g); 684 return ra_select(g); 685 } 686 687 unsigned int 688 ra_get_node_reg(struct ra_graph *g, unsigned int n) 689 { 690 return g->nodes[n].reg; 691 } 692 693 /** 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 702 * class. 703 * 704 * Must be called before ra_simplify(). 705 */ 706 void 707 ra_set_node_reg(struct ra_graph *g, unsigned int n, unsigned int reg) 708 { 709 g->nodes[n].reg = reg; 710 g->nodes[n].in_stack = false; 711 } 712 713 static float 714 ra_get_spill_benefit(struct ra_graph *g, unsigned int n) 715 { 716 unsigned int j; 717 float benefit = 0; 718 int n_class = g->nodes[n].class; 719 720 /* 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. 724 */ 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); 730 } 731 732 return benefit; 733 } 734 735 /** 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. 738 */ 739 int 740 ra_get_best_spill_node(struct ra_graph *g) 741 { 742 unsigned int best_node = -1; 743 float best_benefit = 0.0; 744 unsigned int n; 745 746 /* 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. 750 */ 751 for (n = 0; n < g->count; n++) { 752 float cost = g->nodes[n].spill_cost; 753 float benefit; 754 755 if (cost <= 0.0f) 756 continue; 757 758 if (g->nodes[n].in_stack) 759 continue; 760 761 benefit = ra_get_spill_benefit(g, n); 762 763 if (benefit / cost > best_benefit) { 764 best_benefit = benefit / cost; 765 best_node = n; 766 } 767 } 768 769 return best_node; 770 } 771 772 /** 773 * Only nodes with a spill cost set (cost != 0.0) will be considered 774 * for register spilling. 775 */ 776 void 777 ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost) 778 { 779 g->nodes[n].spill_cost = cost; 780 } 781