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