src/mesa/drivers/dri/i965/brw_fs_copy_propagation.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
  24 /** @file brw_fs_copy_propagation.cpp
  25  *
  26  * Support for global copy propagation in two passes: A local pass that does
  27  * intra-block copy (and constant) propagation, and a global pass that uses
  28  * dataflow analysis on the copies available at the end of each block to re-do
  29  * local copy propagation with more copies available.
  30  *
  31  * See Muchnick's Advanced Compiler Design and Implementation, section
  32  * 12.5 (p356).
  33  */
  34
  35 #define ACP_HASH_SIZE 16
  36
  37 #include "main/bitset.h"
  38 #include "brw_fs.h"
  39 #include "brw_cfg.h"
  40
  41 namespace { /* avoid conflict with opt_copy_propagation_elements */
  42 struct acp_entry : public exec_node {
  43    fs_reg dst;
  44    fs_reg src;
  45    uint8_t regs_written;
  46    enum opcode opcode;
  47    bool saturate;
  48 };
  49
  50 struct block_data {
  51    /**
  52     * Which entries in the fs_copy_prop_dataflow acp table are live at the
  53     * start of this block.  This is the useful output of the analysis, since
  54     * it lets us plug those into the local copy propagation on the second
  55     * pass.
  56     */
  57    BITSET_WORD *livein;
  58
  59    /**
  60     * Which entries in the fs_copy_prop_dataflow acp table are live at the end
  61     * of this block.  This is done in initial setup from the per-block acps
  62     * returned by the first local copy prop pass.
  63     */
  64    BITSET_WORD *liveout;
  65
  66    /**
  67     * Which entries in the fs_copy_prop_dataflow acp table are generated by
  68     * instructions in this block which reach the end of the block without
  69     * being killed.
  70     */
  71    BITSET_WORD *copy;
  72
  73    /**
  74     * Which entries in the fs_copy_prop_dataflow acp table are killed over the
  75     * course of this block.
  76     */
  77    BITSET_WORD *kill;
  78 };
  79
  80 class fs_copy_prop_dataflow
  81 {
  82 public:
  83    fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
  84                          exec_list *out_acp[ACP_HASH_SIZE]);
  85
  86    void setup_initial_values();
  87    void run();
  88
  89    void dump_block_data() const;
  90
  91    void *mem_ctx;
  92    cfg_t *cfg;
  93
  94    acp_entry **acp;
  95    int num_acp;
  96    int bitset_words;
  97
  98   struct block_data *bd;
  99 };
 100 } /* anonymous namespace */
 101
 102 fs_copy_prop_dataflow::fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
 103                                              exec_list *out_acp[ACP_HASH_SIZE])
 104    : mem_ctx(mem_ctx), cfg(cfg)
 105 {
 106    bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
 107
 108    num_acp = 0;
 109    foreach_block (block, cfg) {
 110       for (int i = 0; i < ACP_HASH_SIZE; i++) {
 111          num_acp += out_acp[block->num][i].length();
 112       }
 113    }
 114
 115    acp = rzalloc_array(mem_ctx, struct acp_entry *, num_acp);
 116
 117    bitset_words = BITSET_WORDS(num_acp);
 118
 119    int next_acp = 0;
 120    foreach_block (block, cfg) {
 121       bd[block->num].livein = rzalloc_array(bd, BITSET_WORD, bitset_words);
 122       bd[block->num].liveout = rzalloc_array(bd, BITSET_WORD, bitset_words);
 123       bd[block->num].copy = rzalloc_array(bd, BITSET_WORD, bitset_words);
 124       bd[block->num].kill = rzalloc_array(bd, BITSET_WORD, bitset_words);
 125
 126       for (int i = 0; i < ACP_HASH_SIZE; i++) {
 127          foreach_in_list(acp_entry, entry, &out_acp[block->num][i]) {
 128             acp[next_acp] = entry;
 129
 130             /* opt_copy_propagate_local populates out_acp with copies created
 131              * in a block which are still live at the end of the block.  This
 132              * is exactly what we want in the COPY set.
 133              */
 134             BITSET_SET(bd[block->num].copy, next_acp);
 135
 136             next_acp++;
 137          }
 138       }
 139    }
 140
 141    assert(next_acp == num_acp);
 142
 143    setup_initial_values();
 144    run();
 145 }
 146
 147 /**
 148  * Set up initial values for each of the data flow sets, prior to running
 149  * the fixed-point algorithm.
 150  */
 151 void
 152 fs_copy_prop_dataflow::setup_initial_values()
 153 {
 154    /* Initialize the COPY and KILL sets. */
 155    foreach_block (block, cfg) {
 156       foreach_inst_in_block(fs_inst, inst, block) {
 157          if (inst->dst.file != GRF)
 158             continue;
 159
 160          /* Mark ACP entries which are killed by this instruction. */
 161          for (int i = 0; i < num_acp; i++) {
 162             if (inst->overwrites_reg(acp[i]->dst) ||
 163                 inst->overwrites_reg(acp[i]->src)) {
 164                BITSET_SET(bd[block->num].kill, i);
 165             }
 166          }
 167       }
 168    }
 169
 170    /* Populate the initial values for the livein and liveout sets.  For the
 171     * block at the start of the program, livein = 0 and liveout = copy.
 172     * For the others, set liveout to 0 (the empty set) and livein to ~0
 173     * (the universal set).
 174     */
 175    foreach_block (block, cfg) {
 176       if (block->parents.is_empty()) {
 177          for (int i = 0; i < bitset_words; i++) {
 178             bd[block->num].livein[i] = 0u;
 179             bd[block->num].liveout[i] = bd[block->num].copy[i];
 180          }
 181       } else {
 182          for (int i = 0; i < bitset_words; i++) {
 183             bd[block->num].liveout[i] = 0u;
 184             bd[block->num].livein[i] = ~0u;
 185          }
 186       }
 187    }
 188 }
 189
 190 /**
 191  * Walk the set of instructions in the block, marking which entries in the acp
 192  * are killed by the block.
 193  */
 194 void
 195 fs_copy_prop_dataflow::run()
 196 {
 197    bool progress;
 198
 199    do {
 200       progress = false;
 201
 202       /* Update liveout for all blocks. */
 203       foreach_block (block, cfg) {
 204          if (block->parents.is_empty())
 205             continue;
 206
 207          for (int i = 0; i < bitset_words; i++) {
 208             const BITSET_WORD old_liveout = bd[block->num].liveout[i];
 209
 210             bd[block->num].liveout[i] =
 211                bd[block->num].copy[i] | (bd[block->num].livein[i] &
 212                                          ~bd[block->num].kill[i]);
 213
 214             if (old_liveout != bd[block->num].liveout[i])
 215                progress = true;
 216          }
 217       }
 218
 219       /* Update livein for all blocks.  If a copy is live out of all parent
 220        * blocks, it's live coming in to this block.
 221        */
 222       foreach_block (block, cfg) {
 223          if (block->parents.is_empty())
 224             continue;
 225
 226          for (int i = 0; i < bitset_words; i++) {
 227             const BITSET_WORD old_livein = bd[block->num].livein[i];
 228
 229             bd[block->num].livein[i] = ~0u;
 230             foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
 231                bblock_t *parent = parent_link->block;
 232                bd[block->num].livein[i] &= bd[parent->num].liveout[i];
 233             }
 234
 235             if (old_livein != bd[block->num].livein[i])
 236                progress = true;
 237          }
 238       }
 239    } while (progress);
 240 }
 241
 242 void
 243 fs_copy_prop_dataflow::dump_block_data() const
 244 {
 245    foreach_block (block, cfg) {
 246       fprintf(stderr, "Block %d [%d, %d] (parents ", block->num,
 247              block->start_ip, block->end_ip);
 248       foreach_list_typed(bblock_link, link, link, &block->parents) {
 249          bblock_t *parent = link->block;
 250          fprintf(stderr, "%d ", parent->num);
 251       }
 252       fprintf(stderr, "):\n");
 253       fprintf(stderr, "       livein = 0x");
 254       for (int i = 0; i < bitset_words; i++)
 255          fprintf(stderr, "%08x", bd[block->num].livein[i]);
 256       fprintf(stderr, ", liveout = 0x");
 257       for (int i = 0; i < bitset_words; i++)
 258          fprintf(stderr, "%08x", bd[block->num].liveout[i]);
 259       fprintf(stderr, ",\n       copy   = 0x");
 260       for (int i = 0; i < bitset_words; i++)
 261          fprintf(stderr, "%08x", bd[block->num].copy[i]);
 262       fprintf(stderr, ", kill    = 0x");
 263       for (int i = 0; i < bitset_words; i++)
 264          fprintf(stderr, "%08x", bd[block->num].kill[i]);
 265       fprintf(stderr, "\n");
 266    }
 267 }
 268
 269 static bool
 270 is_logic_op(enum opcode opcode)
 271 {
 272    return (opcode == BRW_OPCODE_AND ||
 273            opcode == BRW_OPCODE_OR  ||
 274            opcode == BRW_OPCODE_XOR ||
 275            opcode == BRW_OPCODE_NOT);
 276 }
 277
 278 bool
 279 fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
 280 {
 281    if (inst->src[arg].file != GRF)
 282       return false;
 283
 284    if (entry->src.file == IMM)
 285       return false;
 286    assert(entry->src.file == GRF || entry->src.file == UNIFORM);
 287
 288    if (entry->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
 289        inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD)
 290       return false;
 291
 292    assert(entry->dst.file == GRF);
 293    if (inst->src[arg].reg != entry->dst.reg)
 294       return false;
 295
 296    /* Bail if inst is reading a range that isn't contained in the range
 297     * that entry is writing.
 298     */
 299    if (inst->src[arg].reg_offset < entry->dst.reg_offset ||
 300        (inst->src[arg].reg_offset * 32 + inst->src[arg].subreg_offset +
 301         inst->regs_read(this, arg) * inst->src[arg].stride * 32) >
 302        (entry->dst.reg_offset + entry->regs_written) * 32)
 303       return false;
 304
 305    /* See resolve_ud_negate() and comment in brw_fs_emit.cpp. */
 306    if (inst->conditional_mod &&
 307        inst->src[arg].type == BRW_REGISTER_TYPE_UD &&
 308        entry->src.negate)
 309       return false;
 310
 311    bool has_source_modifiers = entry->src.abs || entry->src.negate;
 312
 313    if ((has_source_modifiers || entry->src.file == UNIFORM ||
 314         !entry->src.is_contiguous()) &&
 315        !inst->can_do_source_mods(brw))
 316       return false;
 317
 318    if (has_source_modifiers &&
 319        inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
 320       return false;
 321
 322    /* Bail if the result of composing both strides would exceed the
 323     * hardware limit.
 324     */
 325    if (entry->src.stride * inst->src[arg].stride > 4)
 326       return false;
 327
 328    /* Bail if the result of composing both strides cannot be expressed
 329     * as another stride. This avoids, for example, trying to transform
 330     * this:
 331     *
 332     *     MOV (8) rX<1>UD rY<0;1,0>UD
 333     *     FOO (8) ...     rX<8;8,1>UW
 334     *
 335     * into this:
 336     *
 337     *     FOO (8) ...     rY<0;1,0>UW
 338     *
 339     * Which would have different semantics.
 340     */
 341    if (entry->src.stride != 1 &&
 342        (inst->src[arg].stride *
 343         type_sz(inst->src[arg].type)) % type_sz(entry->src.type) != 0)
 344       return false;
 345
 346    if (has_source_modifiers && entry->dst.type != inst->src[arg].type)
 347       return false;
 348
 349    if (brw->gen >= 8 && (entry->src.negate || entry->src.abs) &&
 350        is_logic_op(inst->opcode)) {
 351       return false;
 352    }
 353
 354    if (entry->saturate) {
 355       switch(inst->opcode) {
 356       case BRW_OPCODE_SEL:
 357          if (inst->src[1].file != IMM ||
 358              inst->src[1].fixed_hw_reg.dw1.f < 0.0 ||
 359              inst->src[1].fixed_hw_reg.dw1.f > 1.0) {
 360             return false;
 361          }
 362          break;
 363       default:
 364          return false;
 365       }
 366    }
 367
 368    inst->src[arg].file = entry->src.file;
 369    inst->src[arg].reg = entry->src.reg;
 370    inst->src[arg].stride *= entry->src.stride;
 371    inst->saturate = inst->saturate || entry->saturate;
 372
 373    switch (entry->src.file) {
 374    case UNIFORM:
 375       assert(entry->src.width == 1);
 376    case BAD_FILE:
 377    case HW_REG:
 378       inst->src[arg].width = entry->src.width;
 379       inst->src[arg].reg_offset = entry->src.reg_offset;
 380       inst->src[arg].subreg_offset = entry->src.subreg_offset;
 381       break;
 382    case GRF:
 383       {
 384          assert(entry->src.width % inst->src[arg].width == 0);
 385          /* In this case, we'll just leave the width alone.  The source
 386           * register could have different widths depending on how it is
 387           * being used.  For instance, if only half of the register was
 388           * used then we want to preserve that and continue to only use
 389           * half.
 390           *
 391           * Also, we have to deal with mapping parts of vgrfs to other
 392           * parts of vgrfs so we have to do some reg_offset magic.
 393           */
 394
 395          /* Compute the offset of inst->src[arg] relative to inst->dst */
 396          assert(entry->dst.subreg_offset == 0);
 397          int rel_offset = inst->src[arg].reg_offset - entry->dst.reg_offset;
 398          int rel_suboffset = inst->src[arg].subreg_offset;
 399
 400          /* Compute the final register offset (in bytes) */
 401          int offset = entry->src.reg_offset * 32 + entry->src.subreg_offset;
 402          offset += rel_offset * 32 + rel_suboffset;
 403          inst->src[arg].reg_offset = offset / 32;
 404          inst->src[arg].subreg_offset = offset % 32;
 405       }
 406       break;
 407    default:
 408       unreachable("Invalid register file");
 409       break;
 410    }
 411
 412    if (!inst->src[arg].abs) {
 413       inst->src[arg].abs = entry->src.abs;
 414       inst->src[arg].negate ^= entry->src.negate;
 415    }
 416
 417    return true;
 418 }
 419
 420
 421 bool
 422 fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
 423 {
 424    bool progress = false;
 425
 426    if (entry->src.file != IMM)
 427       return false;
 428    if (entry->saturate)
 429       return false;
 430
 431    for (int i = inst->sources - 1; i >= 0; i--) {
 432       if (inst->src[i].file != GRF)
 433          continue;
 434
 435       assert(entry->dst.file == GRF);
 436       if (inst->src[i].reg != entry->dst.reg)
 437          continue;
 438
 439       /* Bail if inst is reading a range that isn't contained in the range
 440        * that entry is writing.
 441        */
 442       if (inst->src[i].reg_offset < entry->dst.reg_offset ||
 443           (inst->src[i].reg_offset * 32 + inst->src[i].subreg_offset +
 444            inst->regs_read(this, i) * inst->src[i].stride * 32) >
 445           (entry->dst.reg_offset + entry->regs_written) * 32)
 446          continue;
 447
 448       /* Don't bother with cases that should have been taken care of by the
 449        * GLSL compiler's constant folding pass.
 450        */
 451       if (inst->src[i].negate || inst->src[i].abs)
 452          continue;
 453
 454       fs_reg val = entry->src;
 455       val.effective_width = inst->src[i].effective_width;
 456       val.type = inst->src[i].type;
 457
 458       switch (inst->opcode) {
 459       case BRW_OPCODE_MOV:
 460       case SHADER_OPCODE_LOAD_PAYLOAD:
 461          inst->src[i] = val;
 462          progress = true;
 463          break;
 464
 465       case SHADER_OPCODE_POW:
 466       case SHADER_OPCODE_INT_QUOTIENT:
 467       case SHADER_OPCODE_INT_REMAINDER:
 468          if (brw->gen < 8)
 469             break;
 470          /* fallthrough */
 471       case BRW_OPCODE_BFI1:
 472       case BRW_OPCODE_ASR:
 473       case BRW_OPCODE_SHL:
 474       case BRW_OPCODE_SHR:
 475       case BRW_OPCODE_SUBB:
 476          if (i == 1) {
 477             inst->src[i] = val;
 478             progress = true;
 479          }
 480          break;
 481
 482       case BRW_OPCODE_MACH:
 483       case BRW_OPCODE_MUL:
 484       case BRW_OPCODE_ADD:
 485       case BRW_OPCODE_OR:
 486       case BRW_OPCODE_AND:
 487       case BRW_OPCODE_XOR:
 488       case BRW_OPCODE_ADDC:
 489          if (i == 1) {
 490             inst->src[i] = val;
 491             progress = true;
 492          } else if (i == 0 && inst->src[1].file != IMM) {
 493             /* Fit this constant in by commuting the operands.
 494              * Exception: we can't do this for 32-bit integer MUL/MACH
 495              * because it's asymmetric.
 496              */
 497             if ((inst->opcode == BRW_OPCODE_MUL ||
 498                  inst->opcode == BRW_OPCODE_MACH) &&
 499                 (inst->src[1].type == BRW_REGISTER_TYPE_D ||
 500                  inst->src[1].type == BRW_REGISTER_TYPE_UD))
 501                break;
 502             inst->src[0] = inst->src[1];
 503             inst->src[1] = val;
 504             progress = true;
 505          }
 506          break;
 507
 508       case BRW_OPCODE_CMP:
 509       case BRW_OPCODE_IF:
 510          if (i == 1) {
 511             inst->src[i] = val;
 512             progress = true;
 513          } else if (i == 0 && inst->src[1].file != IMM) {
 514             enum brw_conditional_mod new_cmod;
 515
 516             new_cmod = brw_swap_cmod(inst->conditional_mod);
 517             if (new_cmod != BRW_CONDITIONAL_NONE) {
 518                /* Fit this constant in by swapping the operands and
 519                 * flipping the test
 520                 */
 521                inst->src[0] = inst->src[1];
 522                inst->src[1] = val;
 523                inst->conditional_mod = new_cmod;
 524                progress = true;
 525             }
 526          }
 527          break;
 528
 529       case BRW_OPCODE_SEL:
 530          if (i == 1) {
 531             inst->src[i] = val;
 532             progress = true;
 533          } else if (i == 0 && inst->src[1].file != IMM) {
 534             inst->src[0] = inst->src[1];
 535             inst->src[1] = val;
 536
 537             /* If this was predicated, flipping operands means
 538              * we also need to flip the predicate.
 539              */
 540             if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
 541                inst->predicate_inverse =
 542                   !inst->predicate_inverse;
 543             }
 544             progress = true;
 545          }
 546          break;
 547
 548       case SHADER_OPCODE_RCP:
 549          /* The hardware doesn't do math on immediate values
 550           * (because why are you doing that, seriously?), but
 551           * the correct answer is to just constant fold it
 552           * anyway.
 553           */
 554          assert(i == 0);
 555          if (inst->src[0].fixed_hw_reg.dw1.f != 0.0f) {
 556             inst->opcode = BRW_OPCODE_MOV;
 557             inst->src[0] = val;
 558             inst->src[0].fixed_hw_reg.dw1.f = 1.0f / inst->src[0].fixed_hw_reg.dw1.f;
 559             progress = true;
 560          }
 561          break;
 562
 563       case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
 564          inst->src[i] = val;
 565          progress = true;
 566          break;
 567
 568       default:
 569          break;
 570       }
 571    }
 572
 573    return progress;
 574 }
 575
 576 static bool
 577 can_propagate_from(fs_inst *inst)
 578 {
 579    return (inst->opcode == BRW_OPCODE_MOV &&
 580            inst->dst.file == GRF &&
 581            ((inst->src[0].file == GRF &&
 582              (inst->src[0].reg != inst->dst.reg ||
 583               inst->src[0].reg_offset != inst->dst.reg_offset)) ||
 584             inst->src[0].file == UNIFORM ||
 585             inst->src[0].file == IMM) &&
 586            inst->src[0].type == inst->dst.type &&
 587            !inst->is_partial_write());
 588 }
 589
 590 /* Walks a basic block and does copy propagation on it using the acp
 591  * list.
 592  */
 593 bool
 594 fs_visitor::opt_copy_propagate_local(void *copy_prop_ctx, bblock_t *block,
 595                                      exec_list *acp)
 596 {
 597    bool progress = false;
 598
 599    foreach_inst_in_block(fs_inst, inst, block) {
 600       /* Try propagating into this instruction. */
 601       for (int i = 0; i < inst->sources; i++) {
 602          if (inst->src[i].file != GRF)
 603             continue;
 604
 605          foreach_in_list(acp_entry, entry, &acp[inst->src[i].reg % ACP_HASH_SIZE]) {
 606             if (try_constant_propagate(inst, entry))
 607                progress = true;
 608
 609             if (try_copy_propagate(inst, i, entry))
 610                progress = true;
 611          }
 612       }
 613
 614       /* kill the destination from the ACP */
 615       if (inst->dst.file == GRF) {
 616          foreach_in_list_safe(acp_entry, entry, &acp[inst->dst.reg % ACP_HASH_SIZE]) {
 617             if (inst->overwrites_reg(entry->dst)) {
 618                entry->remove();
 619             }
 620          }
 621
 622          /* Oops, we only have the chaining hash based on the destination, not
 623           * the source, so walk across the entire table.
 624           */
 625          for (int i = 0; i < ACP_HASH_SIZE; i++) {
 626             foreach_in_list_safe(acp_entry, entry, &acp[i]) {
 627                if (inst->overwrites_reg(entry->src))
 628                   entry->remove();
 629             }
 630          }
 631       }
 632
 633       /* If this instruction's source could potentially be folded into the
 634        * operand of another instruction, add it to the ACP.
 635        */
 636       if (can_propagate_from(inst)) {
 637          acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
 638          entry->dst = inst->dst;
 639          entry->src = inst->src[0];
 640          entry->regs_written = inst->regs_written;
 641          entry->opcode = inst->opcode;
 642          entry->saturate = inst->saturate;
 643          acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 644       } else if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
 645                  inst->dst.file == GRF) {
 646          int offset = 0;
 647          for (int i = 0; i < inst->sources; i++) {
 648             int regs_written = ((inst->src[i].effective_width *
 649                                  type_sz(inst->src[i].type)) + 31) / 32;
 650             if (inst->src[i].file == GRF) {
 651                acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
 652                entry->dst = inst->dst;
 653                entry->dst.reg_offset = offset;
 654                entry->dst.width = inst->src[i].effective_width;
 655                entry->src = inst->src[i];
 656                entry->regs_written = regs_written;
 657                entry->opcode = inst->opcode;
 658                if (!entry->dst.equals(inst->src[i])) {
 659                   acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 660                } else {
 661                   ralloc_free(entry);
 662                }
 663             }
 664             offset += regs_written;
 665          }
 666       }
 667    }
 668
 669    return progress;
 670 }
 671
 672 bool
 673 fs_visitor::opt_copy_propagate()
 674 {
 675    bool progress = false;
 676    void *copy_prop_ctx = ralloc_context(NULL);
 677    exec_list *out_acp[cfg->num_blocks];
 678
 679    for (int i = 0; i < cfg->num_blocks; i++)
 680       out_acp[i] = new exec_list [ACP_HASH_SIZE];
 681
 682    /* First, walk through each block doing local copy propagation and getting
 683     * the set of copies available at the end of the block.
 684     */
 685    foreach_block (block, cfg) {
 686       progress = opt_copy_propagate_local(copy_prop_ctx, block,
 687                                           out_acp[block->num]) || progress;
 688    }
 689
 690    /* Do dataflow analysis for those available copies. */
 691    fs_copy_prop_dataflow dataflow(copy_prop_ctx, cfg, out_acp);
 692
 693    /* Next, re-run local copy propagation, this time with the set of copies
 694     * provided by the dataflow analysis available at the start of a block.
 695     */
 696    foreach_block (block, cfg) {
 697       exec_list in_acp[ACP_HASH_SIZE];
 698
 699       for (int i = 0; i < dataflow.num_acp; i++) {
 700          if (BITSET_TEST(dataflow.bd[block->num].livein, i)) {
 701             struct acp_entry *entry = dataflow.acp[i];
 702             in_acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 703          }
 704       }
 705
 706       progress = opt_copy_propagate_local(copy_prop_ctx, block, in_acp) || progress;
 707    }
 708
 709    for (int i = 0; i < cfg->num_blocks; i++)
 710       delete [] out_acp[i];
 711    ralloc_free(copy_prop_ctx);
 712
 713    if (progress)
 714       invalidate_live_intervals();
 715
 716    return progress;
 717 }