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 }; 178 179 /** 180 * Creates a set of registers for the allocator. 181 * 182 * mem_ctx is a ralloc context for the allocator. The reg set may be freed 183 * using ralloc_free(). 184 */ 185 struct ra_regs * 186 ra_alloc_reg_set(void *mem_ctx, unsigned int count, bool need_conflict_lists) 187 { 188 unsigned int i; 189 struct ra_regs *regs; 190 191 regs = rzalloc(mem_ctx, struct ra_regs); 192 regs->count = count; 193 regs->regs = rzalloc_array(regs, struct ra_reg, count); 194 195 for (i = 0; i < count; i++) { 196 regs->regs[i].conflicts = rzalloc_array(regs->regs, BITSET_WORD, 197 BITSET_WORDS(count)); 198 BITSET_SET(regs->regs[i].conflicts, i); 199 200 if (need_conflict_lists) { 201 regs->regs[i].conflict_list = ralloc_array(regs->regs, 202 unsigned int, 4); 203 regs->regs[i].conflict_list_size = 4; 204 regs->regs[i].conflict_list[0] = i; 205 } else { 206 regs->regs[i].conflict_list = NULL; 207 regs->regs[i].conflict_list_size = 0; 208 } 209 regs->regs[i].num_conflicts = 1; 210 } 211 212 return regs; 213 } 214 215 /** 216 * The register allocator by default prefers to allocate low register numbers, 217 * since it was written for hardware (gen4/5 Intel) that is limited in its 218 * multithreadedness by the number of registers used in a given shader. 219 * 220 * However, for hardware without that restriction, densely packed register 221 * allocation can put serious constraints on instruction scheduling. This 222 * function tells the allocator to rotate around the registers if possible as 223 * it allocates the nodes. 224 */ 225 void 226 ra_set_allocate_round_robin(struct ra_regs *regs) 227 { 228 regs->round_robin = true; 229 } 230 231 static void 232 ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2) 233 { 234 struct ra_reg *reg1 = ®s->regs[r1]; 235 236 if (reg1->conflict_list) { 237 if (reg1->conflict_list_size == reg1->num_conflicts) { 238 reg1->conflict_list_size *= 2; 239 reg1->conflict_list = reralloc(regs->regs, reg1->conflict_list, 240 unsigned int, reg1->conflict_list_size); 241 } 242 reg1->conflict_list[reg1->num_conflicts++] = r2; 243 } 244 BITSET_SET(reg1->conflicts, r2); 245 } 246 247 void 248 ra_add_reg_conflict(struct ra_regs *regs, unsigned int r1, unsigned int r2) 249 { 250 if (!BITSET_TEST(regs->regs[r1].conflicts, r2)) { 251 ra_add_conflict_list(regs, r1, r2); 252 ra_add_conflict_list(regs, r2, r1); 253 } 254 } 255 256 /** 257 * Adds a conflict between base_reg and reg, and also between reg and 258 * anything that base_reg conflicts with. 259 * 260 * This can simplify code for setting up multiple register classes 261 * which are aggregates of some base hardware registers, compared to 262 * explicitly using ra_add_reg_conflict. 263 */ 264 void 265 ra_add_transitive_reg_conflict(struct ra_regs *regs, 266 unsigned int base_reg, unsigned int reg) 267 { 268 unsigned int i; 269 270 ra_add_reg_conflict(regs, reg, base_reg); 271 272 for (i = 0; i < regs->regs[base_reg].num_conflicts; i++) { 273 ra_add_reg_conflict(regs, reg, regs->regs[base_reg].conflict_list[i]); 274 } 275 } 276 277 /** 278 * Makes every conflict on the given register transitive. In other words, 279 * every register that conflicts with r will now conflict with every other 280 * register conflicting with r. 281 * 282 * This can simplify code for setting up multiple register classes 283 * which are aggregates of some base hardware registers, compared to 284 * explicitly using ra_add_reg_conflict. 285 */ 286 void 287 ra_make_reg_conflicts_transitive(struct ra_regs *regs, unsigned int r) 288 { 289 struct ra_reg *reg = ®s->regs[r]; 290 BITSET_WORD tmp; 291 int c; 292 293 BITSET_FOREACH_SET(c, tmp, reg->conflicts, regs->count) { 294 struct ra_reg *other = ®s->regs[c]; 295 unsigned i; 296 for (i = 0; i < BITSET_WORDS(regs->count); i++) 297 other->conflicts[i] |= reg->conflicts[i]; 298 } 299 } 300 301 unsigned int 302 ra_alloc_reg_class(struct ra_regs *regs) 303 { 304 struct ra_class *class; 305 306 regs->classes = reralloc(regs->regs, regs->classes, struct ra_class *, 307 regs->class_count + 1); 308 309 class = rzalloc(regs, struct ra_class); 310 regs->classes[regs->class_count] = class; 311 312 class->regs = rzalloc_array(class, BITSET_WORD, BITSET_WORDS(regs->count)); 313 314 return regs->class_count++; 315 } 316 317 void 318 ra_class_add_reg(struct ra_regs *regs, unsigned int c, unsigned int r) 319 { 320 struct ra_class *class = regs->classes[c]; 321 322 BITSET_SET(class->regs, r); 323 class->p++; 324 } 325 326 /** 327 * Returns true if the register belongs to the given class. 328 */ 329 static bool 330 reg_belongs_to_class(unsigned int r, struct ra_class *c) 331 { 332 return BITSET_TEST(c->regs, r); 333 } 334 335 /** 336 * Must be called after all conflicts and register classes have been 337 * set up and before the register set is used for allocation. 338 * To avoid costly q value computation, use the q_values paramater 339 * to pass precomputed q values to this function. 340 */ 341 void 342 ra_set_finalize(struct ra_regs *regs, unsigned int **q_values) 343 { 344 unsigned int b, c; 345 346 for (b = 0; b < regs->class_count; b++) { 347 regs->classes[b]->q = ralloc_array(regs, unsigned int, regs->class_count); 348 } 349 350 if (q_values) { 351 for (b = 0; b < regs->class_count; b++) { 352 for (c = 0; c < regs->class_count; c++) { 353 regs->classes[b]->q[c] = q_values[b][c]; 354 } 355 } 356 } else { 357 /* Compute, for each class B and C, how many regs of B an 358 * allocation to C could conflict with. 359 */ 360 for (b = 0; b < regs->class_count; b++) { 361 for (c = 0; c < regs->class_count; c++) { 362 unsigned int rc; 363 int max_conflicts = 0; 364 365 for (rc = 0; rc < regs->count; rc++) { 366 int conflicts = 0; 367 unsigned int i; 368 369 if (!reg_belongs_to_class(rc, regs->classes[c])) 370 continue; 371 372 for (i = 0; i < regs->regs[rc].num_conflicts; i++) { 373 unsigned int rb = regs->regs[rc].conflict_list[i]; 374 if (reg_belongs_to_class(rb, regs->classes[b])) 375 conflicts++; 376 } 377 max_conflicts = MAX2(max_conflicts, conflicts); 378 } 379 regs->classes[b]->q[c] = max_conflicts; 380 } 381 } 382 } 383 384 for (b = 0; b < regs->count; b++) { 385 ralloc_free(regs->regs[b].conflict_list); 386 regs->regs[b].conflict_list = NULL; 387 } 388 } 389 390 static void 391 ra_add_node_adjacency(struct ra_graph *g, unsigned int n1, unsigned int n2) 392 { 393 BITSET_SET(g->nodes[n1].adjacency, n2); 394 395 if (n1 != n2) { 396 int n1_class = g->nodes[n1].class; 397 int n2_class = g->nodes[n2].class; 398 g->nodes[n1].q_total += g->regs->classes[n1_class]->q[n2_class]; 399 } 400 401 if (g->nodes[n1].adjacency_count >= 402 g->nodes[n1].adjacency_list_size) { 403 g->nodes[n1].adjacency_list_size *= 2; 404 g->nodes[n1].adjacency_list = reralloc(g, g->nodes[n1].adjacency_list, 405 unsigned int, 406 g->nodes[n1].adjacency_list_size); 407 } 408 409 g->nodes[n1].adjacency_list[g->nodes[n1].adjacency_count] = n2; 410 g->nodes[n1].adjacency_count++; 411 } 412 413 struct ra_graph * 414 ra_alloc_interference_graph(struct ra_regs *regs, unsigned int count) 415 { 416 struct ra_graph *g; 417 unsigned int i; 418 419 g = rzalloc(NULL, struct ra_graph); 420 g->regs = regs; 421 g->nodes = rzalloc_array(g, struct ra_node, count); 422 g->count = count; 423 424 g->stack = rzalloc_array(g, unsigned int, count); 425 426 for (i = 0; i < count; i++) { 427 int bitset_count = BITSET_WORDS(count); 428 g->nodes[i].adjacency = rzalloc_array(g, BITSET_WORD, bitset_count); 429 430 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; 435 436 ra_add_node_adjacency(g, i, i); 437 g->nodes[i].reg = NO_REG; 438 } 439 440 return g; 441 } 442 443 void 444 ra_set_node_class(struct ra_graph *g, 445 unsigned int n, unsigned int class) 446 { 447 g->nodes[n].class = class; 448 } 449 450 void 451 ra_add_node_interference(struct ra_graph *g, 452 unsigned int n1, unsigned int n2) 453 { 454 if (!BITSET_TEST(g->nodes[n1].adjacency, n2)) { 455 ra_add_node_adjacency(g, n1, n2); 456 ra_add_node_adjacency(g, n2, n1); 457 } 458 } 459 460 static bool 461 pq_test(struct ra_graph *g, unsigned int n) 462 { 463 int n_class = g->nodes[n].class; 464 465 return g->nodes[n].q_total < g->regs->classes[n_class]->p; 466 } 467 468 static void 469 decrement_q(struct ra_graph *g, unsigned int n) 470 { 471 unsigned int i; 472 int n_class = g->nodes[n].class; 473 474 for (i = 0; i < g->nodes[n].adjacency_count; i++) { 475 unsigned int n2 = g->nodes[n].adjacency_list[i]; 476 unsigned int n2_class = g->nodes[n2].class; 477 478 if (n != n2 && !g->nodes[n2].in_stack) { 479 assert(g->nodes[n2].q_total >= g->regs->classes[n2_class]->q[n_class]); 480 g->nodes[n2].q_total -= g->regs->classes[n2_class]->q[n_class]; 481 } 482 } 483 } 484 485 /** 486 * Simplifies the interference graph by pushing all 487 * trivially-colorable nodes into a stack of nodes to be colored, 488 * removing them from the graph, and rinsing and repeating. 489 * 490 * If we encounter a case where we can't push any nodes on the stack, then 491 * we optimistically choose a node and push it on the stack. We heuristically 492 * push the node with the lowest total q value, since it has the fewest 493 * neighbors and therefore is most likely to be allocated. 494 */ 495 static void 496 ra_simplify(struct ra_graph *g) 497 { 498 bool progress = true; 499 unsigned int stack_optimistic_start = UINT_MAX; 500 int i; 501 502 while (progress) { 503 unsigned int best_optimistic_node = ~0; 504 unsigned int lowest_q_total = ~0; 505 506 progress = false; 507 508 for (i = g->count - 1; i >= 0; i--) { 509 if (g->nodes[i].in_stack || g->nodes[i].reg != NO_REG) 510 continue; 511 512 if (pq_test(g, i)) { 513 decrement_q(g, i); 514 g->stack[g->stack_count] = i; 515 g->stack_count++; 516 g->nodes[i].in_stack = true; 517 progress = true; 518 } else { 519 unsigned int new_q_total = g->nodes[i].q_total; 520 if (new_q_total < lowest_q_total) { 521 best_optimistic_node = i; 522 lowest_q_total = new_q_total; 523 } 524 } 525 } 526 527 if (!progress && best_optimistic_node != ~0U) { 528 if (stack_optimistic_start == UINT_MAX) 529 stack_optimistic_start = g->stack_count; 530 531 decrement_q(g, best_optimistic_node); 532 g->stack[g->stack_count] = best_optimistic_node; 533 g->stack_count++; 534 g->nodes[best_optimistic_node].in_stack = true; 535 progress = true; 536 } 537 } 538 539 g->stack_optimistic_start = stack_optimistic_start; 540 } 541 542 /** 543 * Pops nodes from the stack back into the graph, coloring them with 544 * registers as they go. 545 * 546 * If all nodes were trivially colorable, then this must succeed. If 547 * not (optimistic coloring), then it may return false; 548 */ 549 static bool 550 ra_select(struct ra_graph *g) 551 { 552 int start_search_reg = 0; 553 554 while (g->stack_count != 0) { 555 unsigned int i; 556 unsigned int ri; 557 unsigned int r = -1; 558 int n = g->stack[g->stack_count - 1]; 559 struct ra_class *c = g->regs->classes[g->nodes[n].class]; 560 561 /* Find the lowest-numbered reg which is not used by a member 562 * of the graph adjacent to us. 563 */ 564 for (ri = 0; ri < g->regs->count; ri++) { 565 r = (start_search_reg + ri) % g->regs->count; 566 if (!reg_belongs_to_class(r, c)) 567 continue; 568 569 /* Check if any of our neighbors conflict with this register choice. */ 570 for (i = 0; i < g->nodes[n].adjacency_count; i++) { 571 unsigned int n2 = g->nodes[n].adjacency_list[i]; 572 573 if (!g->nodes[n2].in_stack && 574 BITSET_TEST(g->regs->regs[r].conflicts, g->nodes[n2].reg)) { 575 break; 576 } 577 } 578 if (i == g->nodes[n].adjacency_count) 579 break; 580 } 581 582 /* set this to false even if we return here so that 583 * ra_get_best_spill_node() considers this node later. 584 */ 585 g->nodes[n].in_stack = false; 586 587 if (ri == g->regs->count) 588 return false; 589 590 g->nodes[n].reg = r; 591 g->stack_count--; 592 593 /* Rotate the starting point except for any nodes above the lowest 594 * optimistically colorable node. The likelihood that we will succeed 595 * at allocating optimistically colorable nodes is highly dependent on 596 * the way that the previous nodes popped off the stack are laid out. 597 * The round-robin strategy increases the fragmentation of the register 598 * file and decreases the number of nearby nodes assigned to the same 599 * color, what increases the likelihood of spilling with respect to the 600 * dense packing strategy. 601 */ 602 if (g->regs->round_robin && 603 g->stack_count - 1 <= g->stack_optimistic_start) 604 start_search_reg = r + 1; 605 } 606 607 return true; 608 } 609 610 bool 611 ra_allocate(struct ra_graph *g) 612 { 613 ra_simplify(g); 614 return ra_select(g); 615 } 616 617 unsigned int 618 ra_get_node_reg(struct ra_graph *g, unsigned int n) 619 { 620 return g->nodes[n].reg; 621 } 622 623 /** 624 * Forces a node to a specific register. This can be used to avoid 625 * creating a register class containing one node when handling data 626 * that must live in a fixed location and is known to not conflict 627 * with other forced register assignment (as is common with shader 628 * input data). These nodes do not end up in the stack during 629 * ra_simplify(), and thus at ra_select() time it is as if they were 630 * the first popped off the stack and assigned their fixed locations. 631 * Nodes that use this function do not need to be assigned a register 632 * class. 633 * 634 * Must be called before ra_simplify(). 635 */ 636 void 637 ra_set_node_reg(struct ra_graph *g, unsigned int n, unsigned int reg) 638 { 639 g->nodes[n].reg = reg; 640 g->nodes[n].in_stack = false; 641 } 642 643 static float 644 ra_get_spill_benefit(struct ra_graph *g, unsigned int n) 645 { 646 unsigned int j; 647 float benefit = 0; 648 int n_class = g->nodes[n].class; 649 650 /* Define the benefit of eliminating an interference between n, n2 651 * through spilling as q(C, B) / p(C). This is similar to the 652 * "count number of edges" approach of traditional graph coloring, 653 * but takes classes into account. 654 */ 655 for (j = 0; j < g->nodes[n].adjacency_count; j++) { 656 unsigned int n2 = g->nodes[n].adjacency_list[j]; 657 if (n != n2) { 658 unsigned int n2_class = g->nodes[n2].class; 659 benefit += ((float)g->regs->classes[n_class]->q[n2_class] / 660 g->regs->classes[n_class]->p); 661 } 662 } 663 664 return benefit; 665 } 666 667 /** 668 * Returns a node number to be spilled according to the cost/benefit using 669 * the pq test, or -1 if there are no spillable nodes. 670 */ 671 int 672 ra_get_best_spill_node(struct ra_graph *g) 673 { 674 unsigned int best_node = -1; 675 float best_benefit = 0.0; 676 unsigned int n; 677 678 /* Consider any nodes that we colored successfully or the node we failed to 679 * color for spilling. When we failed to color a node in ra_select(), we 680 * only considered these nodes, so spilling any other ones would not result 681 * in us making progress. 682 */ 683 for (n = 0; n < g->count; n++) { 684 float cost = g->nodes[n].spill_cost; 685 float benefit; 686 687 if (cost <= 0.0f) 688 continue; 689 690 if (g->nodes[n].in_stack) 691 continue; 692 693 benefit = ra_get_spill_benefit(g, n); 694 695 if (benefit / cost > best_benefit) { 696 best_benefit = benefit / cost; 697 best_node = n; 698 } 699 } 700 701 return best_node; 702 } 703 704 /** 705 * Only nodes with a spill cost set (cost != 0.0) will be considered 706 * for register spilling. 707 */ 708 void 709 ra_set_node_spill_cost(struct ra_graph *g, unsigned int n, float cost) 710 { 711 g->nodes[n].spill_cost = cost; 712 } 713
