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