src/intel/compiler/brw_cfg.cpp

   1 /*
   2  * Copyright © 2012 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 #include "brw_cfg.h"
  29 #include "brw_shader.h"
  30
  31 /** @file brw_cfg.cpp
  32  *
  33  * Walks the shader instructions generated and creates a set of basic
  34  * blocks with successor/predecessor edges connecting them.
  35  */
  36
  37 using namespace brw;
  38
  39 static bblock_t *
  40 pop_stack(exec_list *list)
  41 {
  42    bblock_link *link = (bblock_link *)list->get_tail();
  43    bblock_t *block = link->block;
  44    link->link.remove();
  45
  46    return block;
  47 }
  48
  49 static exec_node *
  50 link(void *mem_ctx, bblock_t *block, enum bblock_link_kind kind)
  51 {
  52    bblock_link *l = new(mem_ctx) bblock_link(block, kind);
  53    return &l->link;
  54 }
  55
  56 void
  57 push_stack(exec_list *list, void *mem_ctx, bblock_t *block)
  58 {
  59    /* The kind of the link is immaterial, but we need to provide one since
  60     * this is (ab)using the edge data structure in order to implement a stack.
  61     */
  62    list->push_tail(link(mem_ctx, block, bblock_link_logical));
  63 }
  64
  65 bblock_t::bblock_t(cfg_t *cfg) :
  66    cfg(cfg), start_ip(0), end_ip(0), num(0), cycle_count(0)
  67 {
  68    instructions.make_empty();
  69    parents.make_empty();
  70    children.make_empty();
  71 }
  72
  73 void
  74 bblock_t::add_successor(void *mem_ctx, bblock_t *successor,
  75                         enum bblock_link_kind kind)
  76 {
  77    successor->parents.push_tail(::link(mem_ctx, this, kind));
  78    children.push_tail(::link(mem_ctx, successor, kind));
  79 }
  80
  81 bool
  82 bblock_t::is_predecessor_of(const bblock_t *block,
  83                             enum bblock_link_kind kind) const
  84 {
  85    foreach_list_typed_safe (bblock_link, parent, link, &block->parents) {
  86       if (parent->block == this && parent->kind <= kind) {
  87          return true;
  88       }
  89    }
  90
  91    return false;
  92 }
  93
  94 bool
  95 bblock_t::is_successor_of(const bblock_t *block,
  96                           enum bblock_link_kind kind) const
  97 {
  98    foreach_list_typed_safe (bblock_link, child, link, &block->children) {
  99       if (child->block == this && child->kind <= kind) {
 100          return true;
 101       }
 102    }
 103
 104    return false;
 105 }
 106
 107 static bool
 108 ends_block(const backend_instruction *inst)
 109 {
 110    enum opcode op = inst->opcode;
 111
 112    return op == BRW_OPCODE_IF ||
 113           op == BRW_OPCODE_ELSE ||
 114           op == BRW_OPCODE_CONTINUE ||
 115           op == BRW_OPCODE_BREAK ||
 116           op == BRW_OPCODE_DO ||
 117           op == BRW_OPCODE_WHILE;
 118 }
 119
 120 static bool
 121 starts_block(const backend_instruction *inst)
 122 {
 123    enum opcode op = inst->opcode;
 124
 125    return op == BRW_OPCODE_DO ||
 126           op == BRW_OPCODE_ENDIF;
 127 }
 128
 129 bool
 130 bblock_t::can_combine_with(const bblock_t *that) const
 131 {
 132    if ((const bblock_t *)this->link.next != that)
 133       return false;
 134
 135    if (ends_block(this->end()) ||
 136        starts_block(that->start()))
 137       return false;
 138
 139    return true;
 140 }
 141
 142 void
 143 bblock_t::combine_with(bblock_t *that)
 144 {
 145    assert(this->can_combine_with(that));
 146    foreach_list_typed (bblock_link, link, link, &that->parents) {
 147       assert(link->block == this);
 148    }
 149
 150    this->end_ip = that->end_ip;
 151    this->instructions.append_list(&that->instructions);
 152
 153    this->cfg->remove_block(that);
 154 }
 155
 156 void
 157 bblock_t::dump() const
 158 {
 159    const backend_shader *s = this->cfg->s;
 160
 161    int ip = this->start_ip;
 162    foreach_inst_in_block(backend_instruction, inst, this) {
 163       fprintf(stderr, "%5d: ", ip);
 164       s->dump_instruction(inst);
 165       ip++;
 166    }
 167 }
 168
 169 cfg_t::cfg_t(const backend_shader *s, exec_list *instructions) :
 170    s(s)
 171 {
 172    mem_ctx = ralloc_context(NULL);
 173    block_list.make_empty();
 174    blocks = NULL;
 175    num_blocks = 0;
 176    cycle_count = 0;
 177
 178    bblock_t *cur = NULL;
 179    int ip = 0;
 180
 181    bblock_t *entry = new_block();
 182    bblock_t *cur_if = NULL;    /**< BB ending with IF. */
 183    bblock_t *cur_else = NULL;  /**< BB ending with ELSE. */
 184    bblock_t *cur_endif = NULL; /**< BB starting with ENDIF. */
 185    bblock_t *cur_do = NULL;    /**< BB starting with DO. */
 186    bblock_t *cur_while = NULL; /**< BB immediately following WHILE. */
 187    exec_list if_stack, else_stack, do_stack, while_stack;
 188    bblock_t *next;
 189
 190    set_next_block(&cur, entry, ip);
 191
 192    foreach_in_list_safe(backend_instruction, inst, instructions) {
 193       /* set_next_block wants the post-incremented ip */
 194       ip++;
 195
 196       inst->exec_node::remove();
 197
 198       switch (inst->opcode) {
 199       case BRW_OPCODE_IF:
 200          cur->instructions.push_tail(inst);
 201
 202          /* Push our information onto a stack so we can recover from
 203           * nested ifs.
 204           */
 205          push_stack(&if_stack, mem_ctx, cur_if);
 206          push_stack(&else_stack, mem_ctx, cur_else);
 207
 208          cur_if = cur;
 209          cur_else = NULL;
 210          cur_endif = NULL;
 211
 212          /* Set up our immediately following block, full of "then"
 213           * instructions.
 214           */
 215          next = new_block();
 216          cur_if->add_successor(mem_ctx, next, bblock_link_logical);
 217
 218          set_next_block(&cur, next, ip);
 219          break;
 220
 221       case BRW_OPCODE_ELSE:
 222          cur->instructions.push_tail(inst);
 223
 224          cur_else = cur;
 225
 226          next = new_block();
 227          assert(cur_if != NULL);
 228          cur_if->add_successor(mem_ctx, next, bblock_link_logical);
 229          cur_else->add_successor(mem_ctx, next, bblock_link_physical);
 230
 231          set_next_block(&cur, next, ip);
 232          break;
 233
 234       case BRW_OPCODE_ENDIF: {
 235          if (cur->instructions.is_empty()) {
 236             /* New block was just created; use it. */
 237             cur_endif = cur;
 238          } else {
 239             cur_endif = new_block();
 240
 241             cur->add_successor(mem_ctx, cur_endif, bblock_link_logical);
 242
 243             set_next_block(&cur, cur_endif, ip - 1);
 244          }
 245
 246          cur->instructions.push_tail(inst);
 247
 248          if (cur_else) {
 249             cur_else->add_successor(mem_ctx, cur_endif, bblock_link_logical);
 250          } else {
 251             assert(cur_if != NULL);
 252             cur_if->add_successor(mem_ctx, cur_endif, bblock_link_logical);
 253          }
 254
 255          assert(cur_if->end()->opcode == BRW_OPCODE_IF);
 256          assert(!cur_else || cur_else->end()->opcode == BRW_OPCODE_ELSE);
 257
 258          /* Pop the stack so we're in the previous if/else/endif */
 259          cur_if = pop_stack(&if_stack);
 260          cur_else = pop_stack(&else_stack);
 261          break;
 262       }
 263       case BRW_OPCODE_DO:
 264          /* Push our information onto a stack so we can recover from
 265           * nested loops.
 266           */
 267          push_stack(&do_stack, mem_ctx, cur_do);
 268          push_stack(&while_stack, mem_ctx, cur_while);
 269
 270          /* Set up the block just after the while.  Don't know when exactly
 271           * it will start, yet.
 272           */
 273          cur_while = new_block();
 274
 275          if (cur->instructions.is_empty()) {
 276             /* New block was just created; use it. */
 277             cur_do = cur;
 278          } else {
 279             cur_do = new_block();
 280
 281             cur->add_successor(mem_ctx, cur_do, bblock_link_logical);
 282
 283             set_next_block(&cur, cur_do, ip - 1);
 284          }
 285
 286          cur->instructions.push_tail(inst);
 287
 288          /* Represent divergent execution of the loop as a pair of alternative
 289           * edges coming out of the DO instruction: For any physical iteration
 290           * of the loop a given logical thread can either start off enabled
 291           * (which is represented as the "next" successor), or disabled (if it
 292           * has reached a non-uniform exit of the loop during a previous
 293           * iteration, which is represented as the "cur_while" successor).
 294           *
 295           * The disabled edge will be taken by the logical thread anytime we
 296           * arrive at the DO instruction through a back-edge coming from a
 297           * conditional exit of the loop where divergent control flow started.
 298           *
 299           * This guarantees that there is a control-flow path from any
 300           * divergence point of the loop into the convergence point
 301           * (immediately past the WHILE instruction) such that it overlaps the
 302           * whole IP region of divergent control flow (potentially the whole
 303           * loop) *and* doesn't imply the execution of any instructions part
 304           * of the loop (since the corresponding execution mask bit will be
 305           * disabled for a diverging thread).
 306           *
 307           * This way we make sure that any variables that are live throughout
 308           * the region of divergence for an inactive logical thread are also
 309           * considered to interfere with any other variables assigned by
 310           * active logical threads within the same physical region of the
 311           * program, since otherwise we would risk cross-channel data
 312           * corruption.
 313           */
 314          next = new_block();
 315          cur->add_successor(mem_ctx, next, bblock_link_logical);
 316          cur->add_successor(mem_ctx, cur_while, bblock_link_physical);
 317          set_next_block(&cur, next, ip);
 318          break;
 319
 320       case BRW_OPCODE_CONTINUE:
 321          cur->instructions.push_tail(inst);
 322
 323          /* A conditional CONTINUE may start a region of divergent control
 324           * flow until the start of the next loop iteration (*not* until the
 325           * end of the loop which is why the successor is not the top-level
 326           * divergence point at cur_do).  The live interval of any variable
 327           * extending through a CONTINUE edge is guaranteed to overlap the
 328           * whole region of divergent execution, because any variable live-out
 329           * at the CONTINUE instruction will also be live-in at the top of the
 330           * loop, and therefore also live-out at the bottom-most point of the
 331           * loop which is reachable from the top (since a control flow path
 332           * exists from a definition of the variable through this CONTINUE
 333           * instruction, the top of the loop, the (reachable) bottom of the
 334           * loop, the top of the loop again, into a use of the variable).
 335           */
 336          assert(cur_do != NULL);
 337          cur->add_successor(mem_ctx, cur_do->next(), bblock_link_logical);
 338
 339          next = new_block();
 340          if (inst->predicate)
 341             cur->add_successor(mem_ctx, next, bblock_link_logical);
 342          else
 343             cur->add_successor(mem_ctx, next, bblock_link_physical);
 344
 345          set_next_block(&cur, next, ip);
 346          break;
 347
 348       case BRW_OPCODE_BREAK:
 349          cur->instructions.push_tail(inst);
 350
 351          /* A conditional BREAK instruction may start a region of divergent
 352           * control flow until the end of the loop if the condition is
 353           * non-uniform, in which case the loop will execute additional
 354           * iterations with the present channel disabled.  We model this as a
 355           * control flow path from the divergence point to the convergence
 356           * point that overlaps the whole IP range of the loop and skips over
 357           * the execution of any other instructions part of the loop.
 358           *
 359           * See the DO case for additional explanation.
 360           */
 361          assert(cur_do != NULL);
 362          cur->add_successor(mem_ctx, cur_do, bblock_link_physical);
 363          cur->add_successor(mem_ctx, cur_while, bblock_link_logical);
 364
 365          next = new_block();
 366          if (inst->predicate)
 367             cur->add_successor(mem_ctx, next, bblock_link_logical);
 368
 369          set_next_block(&cur, next, ip);
 370          break;
 371
 372       case BRW_OPCODE_WHILE:
 373          cur->instructions.push_tail(inst);
 374
 375          assert(cur_do != NULL && cur_while != NULL);
 376
 377          /* A conditional WHILE instruction may start a region of divergent
 378           * control flow until the end of the loop, just like the BREAK
 379           * instruction.  See the BREAK case for more details.  OTOH an
 380           * unconditional WHILE instruction is non-divergent (just like an
 381           * unconditional CONTINUE), and will necessarily lead to the
 382           * execution of an additional iteration of the loop for all enabled
 383           * channels, so we may skip over the divergence point at the top of
 384           * the loop to keep the CFG as unambiguous as possible.
 385           */
 386          if (inst->predicate) {
 387             cur->add_successor(mem_ctx, cur_do, bblock_link_logical);
 388          } else {
 389             cur->add_successor(mem_ctx, cur_do->next(), bblock_link_logical);
 390          }
 391
 392          set_next_block(&cur, cur_while, ip);
 393
 394          /* Pop the stack so we're in the previous loop */
 395          cur_do = pop_stack(&do_stack);
 396          cur_while = pop_stack(&while_stack);
 397          break;
 398
 399       default:
 400          cur->instructions.push_tail(inst);
 401          break;
 402       }
 403    }
 404
 405    cur->end_ip = ip - 1;
 406
 407    make_block_array();
 408 }
 409
 410 cfg_t::~cfg_t()
 411 {
 412    ralloc_free(mem_ctx);
 413 }
 414
 415 void
 416 cfg_t::remove_block(bblock_t *block)
 417 {
 418    foreach_list_typed_safe (bblock_link, predecessor, link, &block->parents) {
 419       /* Remove block from all of its predecessors' successor lists. */
 420       foreach_list_typed_safe (bblock_link, successor, link,
 421                                &predecessor->block->children) {
 422          if (block == successor->block) {
 423             successor->link.remove();
 424             ralloc_free(successor);
 425          }
 426       }
 427
 428       /* Add removed-block's successors to its predecessors' successor lists. */
 429       foreach_list_typed (bblock_link, successor, link, &block->children) {
 430          if (!successor->block->is_successor_of(predecessor->block,
 431                                                 successor->kind)) {
 432             predecessor->block->children.push_tail(link(mem_ctx,
 433                                                         successor->block,
 434                                                         successor->kind));
 435          }
 436       }
 437    }
 438
 439    foreach_list_typed_safe (bblock_link, successor, link, &block->children) {
 440       /* Remove block from all of its childrens' parents lists. */
 441       foreach_list_typed_safe (bblock_link, predecessor, link,
 442                                &successor->block->parents) {
 443          if (block == predecessor->block) {
 444             predecessor->link.remove();
 445             ralloc_free(predecessor);
 446          }
 447       }
 448
 449       /* Add removed-block's predecessors to its successors' predecessor lists. */
 450       foreach_list_typed (bblock_link, predecessor, link, &block->parents) {
 451          if (!predecessor->block->is_predecessor_of(successor->block,
 452                                                     predecessor->kind)) {
 453             successor->block->parents.push_tail(link(mem_ctx,
 454                                                      predecessor->block,
 455                                                      predecessor->kind));
 456          }
 457       }
 458    }
 459
 460    block->link.remove();
 461
 462    for (int b = block->num; b < this->num_blocks - 1; b++) {
 463       this->blocks[b] = this->blocks[b + 1];
 464       this->blocks[b]->num = b;
 465    }
 466
 467    this->blocks[this->num_blocks - 1]->num = this->num_blocks - 2;
 468    this->num_blocks--;
 469 }
 470
 471 bblock_t *
 472 cfg_t::new_block()
 473 {
 474    bblock_t *block = new(mem_ctx) bblock_t(this);
 475
 476    return block;
 477 }
 478
 479 void
 480 cfg_t::set_next_block(bblock_t **cur, bblock_t *block, int ip)
 481 {
 482    if (*cur) {
 483       (*cur)->end_ip = ip - 1;
 484    }
 485
 486    block->start_ip = ip;
 487    block->num = num_blocks++;
 488    block_list.push_tail(&block->link);
 489    *cur = block;
 490 }
 491
 492 void
 493 cfg_t::make_block_array()
 494 {
 495    blocks = ralloc_array(mem_ctx, bblock_t *, num_blocks);
 496
 497    int i = 0;
 498    foreach_block (block, this) {
 499       blocks[i++] = block;
 500    }
 501    assert(i == num_blocks);
 502 }
 503
 504 void
 505 cfg_t::dump()
 506 {
 507    const idom_tree *idom = (s ? &s->idom_analysis.require() : NULL);
 508
 509    foreach_block (block, this) {
 510       if (idom && idom->parent(block))
 511          fprintf(stderr, "START B%d IDOM(B%d)", block->num,
 512                  idom->parent(block)->num);
 513       else
 514          fprintf(stderr, "START B%d IDOM(none)", block->num);
 515
 516       foreach_list_typed(bblock_link, link, link, &block->parents) {
 517          fprintf(stderr, " <%cB%d",
 518                  link->kind == bblock_link_logical ? '-' : '~',
 519                  link->block->num);
 520       }
 521       fprintf(stderr, "\n");
 522       if (s != NULL)
 523          block->dump();
 524       fprintf(stderr, "END B%d", block->num);
 525       foreach_list_typed(bblock_link, link, link, &block->children) {
 526          fprintf(stderr, " %c>B%d",
 527                  link->kind == bblock_link_logical ? '-' : '~',
 528                  link->block->num);
 529       }
 530       fprintf(stderr, "\n");
 531    }
 532 }
 533
 534 /* Calculates the immediate dominator of each block, according to "A Simple,
 535  * Fast Dominance Algorithm" by Keith D. Cooper, Timothy J. Harvey, and Ken
 536  * Kennedy.
 537  *
 538  * The authors claim that for control flow graphs of sizes normally encountered
 539  * (less than 1000 nodes) that this algorithm is significantly faster than
 540  * others like Lengauer-Tarjan.
 541  */
 542 idom_tree::idom_tree(const backend_shader *s) :
 543    num_parents(s->cfg->num_blocks),
 544    parents(new bblock_t *[num_parents]())
 545 {
 546    bool changed;
 547
 548    parents[0] = s->cfg->blocks[0];
 549
 550    do {
 551       changed = false;
 552
 553       foreach_block(block, s->cfg) {
 554          if (block->num == 0)
 555             continue;
 556
 557          bblock_t *new_idom = NULL;
 558          foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
 559             if (parent(parent_link->block)) {
 560                new_idom = (new_idom ? intersect(new_idom, parent_link->block) :
 561                            parent_link->block);
 562             }
 563          }
 564
 565          if (parent(block) != new_idom) {
 566             parents[block->num] = new_idom;
 567             changed = true;
 568          }
 569       }
 570    } while (changed);
 571 }
 572
 573 idom_tree::~idom_tree()
 574 {
 575    delete[] parents;
 576 }
 577
 578 bblock_t *
 579 idom_tree::intersect(bblock_t *b1, bblock_t *b2) const
 580 {
 581    /* Note, the comparisons here are the opposite of what the paper says
 582     * because we index blocks from beginning -> end (i.e. reverse post-order)
 583     * instead of post-order like they assume.
 584     */
 585    while (b1->num != b2->num) {
 586       while (b1->num > b2->num)
 587          b1 = parent(b1);
 588       while (b2->num > b1->num)
 589          b2 = parent(b2);
 590    }
 591    assert(b1);
 592    return b1;
 593 }
 594
 595 void
 596 idom_tree::dump() const
 597 {
 598    printf("digraph DominanceTree {\n");
 599    for (unsigned i = 0; i < num_parents; i++)
 600       printf("\t%d -> %d\n", parents[i]->num, i);
 601    printf("}\n");
 602 }
 603
 604 void
 605 cfg_t::dump_cfg()
 606 {
 607    printf("digraph CFG {\n");
 608    for (int b = 0; b < num_blocks; b++) {
 609       bblock_t *block = this->blocks[b];
 610
 611       foreach_list_typed_safe (bblock_link, child, link, &block->children) {
 612          printf("\t%d -> %d\n", b, child->block->num);
 613       }
 614    }
 615    printf("}\n");
 616 }