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