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 "util/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 static bool
 279 can_change_source_types(fs_inst *inst)
 280 {
 281    return !inst->src[0].abs && !inst->src[0].negate &&
 282           (inst->opcode == BRW_OPCODE_MOV ||
 283            (inst->opcode == BRW_OPCODE_SEL &&
 284             inst->predicate != BRW_PREDICATE_NONE &&
 285             !inst->src[1].abs && !inst->src[1].negate));
 286 }
 287
 288 bool
 289 fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
 290 {
 291    if (inst->src[arg].file != GRF)
 292       return false;
 293
 294    if (entry->src.file == IMM)
 295       return false;
 296    assert(entry->src.file == GRF || entry->src.file == UNIFORM ||
 297           entry->src.file == ATTR);
 298
 299    if (entry->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
 300        inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD)
 301       return false;
 302
 303    assert(entry->dst.file == GRF);
 304    if (inst->src[arg].reg != entry->dst.reg)
 305       return false;
 306
 307    /* Bail if inst is reading a range that isn't contained in the range
 308     * that entry is writing.
 309     */
 310    if (inst->src[arg].reg_offset < entry->dst.reg_offset ||
 311        (inst->src[arg].reg_offset * 32 + inst->src[arg].subreg_offset +
 312         inst->regs_read(arg) * inst->src[arg].stride * 32) >
 313        (entry->dst.reg_offset + entry->regs_written) * 32)
 314       return false;
 315
 316    /* we can't generally copy-propagate UD negations because we
 317     * can end up accessing the resulting values as signed integers
 318     * instead. See also resolve_ud_negate() and comment in
 319     * fs_generator::generate_code.
 320     */
 321    if (entry->src.type == BRW_REGISTER_TYPE_UD &&
 322        entry->src.negate)
 323       return false;
 324
 325    bool has_source_modifiers = entry->src.abs || entry->src.negate;
 326
 327    if ((has_source_modifiers || entry->src.file == UNIFORM ||
 328         !entry->src.is_contiguous()) &&
 329        !inst->can_do_source_mods(devinfo))
 330       return false;
 331
 332    if (has_source_modifiers &&
 333        inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
 334       return false;
 335
 336    /* Bail if the result of composing both strides would exceed the
 337     * hardware limit.
 338     */
 339    if (entry->src.stride * inst->src[arg].stride > 4)
 340       return false;
 341
 342    /* Bail if the result of composing both strides cannot be expressed
 343     * as another stride. This avoids, for example, trying to transform
 344     * this:
 345     *
 346     *     MOV (8) rX<1>UD rY<0;1,0>UD
 347     *     FOO (8) ...     rX<8;8,1>UW
 348     *
 349     * into this:
 350     *
 351     *     FOO (8) ...     rY<0;1,0>UW
 352     *
 353     * Which would have different semantics.
 354     */
 355    if (entry->src.stride != 1 &&
 356        (inst->src[arg].stride *
 357         type_sz(inst->src[arg].type)) % type_sz(entry->src.type) != 0)
 358       return false;
 359
 360    if (has_source_modifiers &&
 361        entry->dst.type != inst->src[arg].type &&
 362        !can_change_source_types(inst))
 363       return false;
 364
 365    if (devinfo->gen >= 8 && (entry->src.negate || entry->src.abs) &&
 366        is_logic_op(inst->opcode)) {
 367       return false;
 368    }
 369
 370    if (entry->saturate) {
 371       switch(inst->opcode) {
 372       case BRW_OPCODE_SEL:
 373          if (inst->src[1].file != IMM ||
 374              inst->src[1].fixed_hw_reg.dw1.f < 0.0 ||
 375              inst->src[1].fixed_hw_reg.dw1.f > 1.0) {
 376             return false;
 377          }
 378          break;
 379       default:
 380          return false;
 381       }
 382    }
 383
 384    inst->src[arg].file = entry->src.file;
 385    inst->src[arg].reg = entry->src.reg;
 386    inst->src[arg].stride *= entry->src.stride;
 387    inst->saturate = inst->saturate || entry->saturate;
 388
 389    switch (entry->src.file) {
 390    case UNIFORM:
 391       assert(entry->src.width == 1);
 392    case BAD_FILE:
 393    case HW_REG:
 394       inst->src[arg].width = entry->src.width;
 395       inst->src[arg].reg_offset = entry->src.reg_offset;
 396       inst->src[arg].subreg_offset = entry->src.subreg_offset;
 397       break;
 398    case ATTR:
 399    case GRF:
 400       {
 401          assert(entry->src.width % inst->src[arg].width == 0);
 402          /* In this case, we'll just leave the width alone.  The source
 403           * register could have different widths depending on how it is
 404           * being used.  For instance, if only half of the register was
 405           * used then we want to preserve that and continue to only use
 406           * half.
 407           *
 408           * Also, we have to deal with mapping parts of vgrfs to other
 409           * parts of vgrfs so we have to do some reg_offset magic.
 410           */
 411
 412          /* Compute the offset of inst->src[arg] relative to inst->dst */
 413          assert(entry->dst.subreg_offset == 0);
 414          int rel_offset = inst->src[arg].reg_offset - entry->dst.reg_offset;
 415          int rel_suboffset = inst->src[arg].subreg_offset;
 416
 417          /* Compute the final register offset (in bytes) */
 418          int offset = entry->src.reg_offset * 32 + entry->src.subreg_offset;
 419          offset += rel_offset * 32 + rel_suboffset;
 420          inst->src[arg].reg_offset = offset / 32;
 421          inst->src[arg].subreg_offset = offset % 32;
 422       }
 423       break;
 424    default:
 425       unreachable("Invalid register file");
 426       break;
 427    }
 428
 429    if (has_source_modifiers) {
 430       if (entry->dst.type != inst->src[arg].type) {
 431          /* We are propagating source modifiers from a MOV with a different
 432           * type.  If we got here, then we can just change the source and
 433           * destination types of the instruction and keep going.
 434           */
 435          assert(can_change_source_types(inst));
 436          for (int i = 0; i < inst->sources; i++) {
 437             inst->src[i].type = entry->dst.type;
 438          }
 439          inst->dst.type = entry->dst.type;
 440       }
 441
 442       if (!inst->src[arg].abs) {
 443          inst->src[arg].abs = entry->src.abs;
 444          inst->src[arg].negate ^= entry->src.negate;
 445       }
 446    }
 447
 448    return true;
 449 }
 450
 451
 452 bool
 453 fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
 454 {
 455    bool progress = false;
 456
 457    if (entry->src.file != IMM)
 458       return false;
 459    if (entry->saturate)
 460       return false;
 461
 462    for (int i = inst->sources - 1; i >= 0; i--) {
 463       if (inst->src[i].file != GRF)
 464          continue;
 465
 466       assert(entry->dst.file == GRF);
 467       if (inst->src[i].reg != entry->dst.reg)
 468          continue;
 469
 470       /* Bail if inst is reading a range that isn't contained in the range
 471        * that entry is writing.
 472        */
 473       if (inst->src[i].reg_offset < entry->dst.reg_offset ||
 474           (inst->src[i].reg_offset * 32 + inst->src[i].subreg_offset +
 475            inst->regs_read(i) * inst->src[i].stride * 32) >
 476           (entry->dst.reg_offset + entry->regs_written) * 32)
 477          continue;
 478
 479       fs_reg val = entry->src;
 480       val.type = inst->src[i].type;
 481
 482       if (inst->src[i].abs) {
 483          if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
 484              !brw_abs_immediate(val.type, &val.fixed_hw_reg)) {
 485             continue;
 486          }
 487       }
 488
 489       if (inst->src[i].negate) {
 490          if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
 491              !brw_negate_immediate(val.type, &val.fixed_hw_reg)) {
 492             continue;
 493          }
 494       }
 495
 496       switch (inst->opcode) {
 497       case BRW_OPCODE_MOV:
 498       case SHADER_OPCODE_LOAD_PAYLOAD:
 499          inst->src[i] = val;
 500          progress = true;
 501          break;
 502
 503       case SHADER_OPCODE_INT_QUOTIENT:
 504       case SHADER_OPCODE_INT_REMAINDER:
 505          /* FINISHME: Promote non-float constants and remove this. */
 506          if (devinfo->gen < 8)
 507             break;
 508          /* fallthrough */
 509       case SHADER_OPCODE_POW:
 510          /* Allow constant propagation into src1 (except on Gen 6), and let
 511           * constant combining promote the constant on Gen < 8.
 512           *
 513           * While Gen 6 MATH can take a scalar source, its source and
 514           * destination offsets must be equal and we cannot ensure that.
 515           */
 516          if (devinfo->gen == 6)
 517             break;
 518          /* fallthrough */
 519       case BRW_OPCODE_BFI1:
 520       case BRW_OPCODE_ASR:
 521       case BRW_OPCODE_SHL:
 522       case BRW_OPCODE_SHR:
 523       case BRW_OPCODE_SUBB:
 524          if (i == 1) {
 525             inst->src[i] = val;
 526             progress = true;
 527          }
 528          break;
 529
 530       case BRW_OPCODE_MACH:
 531       case BRW_OPCODE_MUL:
 532       case BRW_OPCODE_ADD:
 533       case BRW_OPCODE_OR:
 534       case BRW_OPCODE_AND:
 535       case BRW_OPCODE_XOR:
 536       case BRW_OPCODE_ADDC:
 537          if (i == 1) {
 538             inst->src[i] = val;
 539             progress = true;
 540          } else if (i == 0 && inst->src[1].file != IMM) {
 541             /* Fit this constant in by commuting the operands.
 542              * Exception: we can't do this for 32-bit integer MUL/MACH
 543              * because it's asymmetric.
 544              *
 545              * The BSpec says for Broadwell that
 546              *
 547              *    "When multiplying DW x DW, the dst cannot be accumulator."
 548              *
 549              * Integer MUL with a non-accumulator destination will be lowered
 550              * by lower_integer_multiplication(), so don't restrict it.
 551              */
 552             if (((inst->opcode == BRW_OPCODE_MUL &&
 553                   inst->dst.is_accumulator()) ||
 554                  inst->opcode == BRW_OPCODE_MACH) &&
 555                 (inst->src[1].type == BRW_REGISTER_TYPE_D ||
 556                  inst->src[1].type == BRW_REGISTER_TYPE_UD))
 557                break;
 558             inst->src[0] = inst->src[1];
 559             inst->src[1] = val;
 560             progress = true;
 561          }
 562          break;
 563
 564       case BRW_OPCODE_CMP:
 565       case BRW_OPCODE_IF:
 566          if (i == 1) {
 567             inst->src[i] = val;
 568             progress = true;
 569          } else if (i == 0 && inst->src[1].file != IMM) {
 570             enum brw_conditional_mod new_cmod;
 571
 572             new_cmod = brw_swap_cmod(inst->conditional_mod);
 573             if (new_cmod != BRW_CONDITIONAL_NONE) {
 574                /* Fit this constant in by swapping the operands and
 575                 * flipping the test
 576                 */
 577                inst->src[0] = inst->src[1];
 578                inst->src[1] = val;
 579                inst->conditional_mod = new_cmod;
 580                progress = true;
 581             }
 582          }
 583          break;
 584
 585       case BRW_OPCODE_SEL:
 586          if (i == 1) {
 587             inst->src[i] = val;
 588             progress = true;
 589          } else if (i == 0 && inst->src[1].file != IMM) {
 590             inst->src[0] = inst->src[1];
 591             inst->src[1] = val;
 592
 593             /* If this was predicated, flipping operands means
 594              * we also need to flip the predicate.
 595              */
 596             if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
 597                inst->predicate_inverse =
 598                   !inst->predicate_inverse;
 599             }
 600             progress = true;
 601          }
 602          break;
 603
 604       case SHADER_OPCODE_RCP:
 605          /* The hardware doesn't do math on immediate values
 606           * (because why are you doing that, seriously?), but
 607           * the correct answer is to just constant fold it
 608           * anyway.
 609           */
 610          assert(i == 0);
 611          if (inst->src[0].fixed_hw_reg.dw1.f != 0.0f) {
 612             inst->opcode = BRW_OPCODE_MOV;
 613             inst->src[0] = val;
 614             inst->src[0].fixed_hw_reg.dw1.f = 1.0f / inst->src[0].fixed_hw_reg.dw1.f;
 615             progress = true;
 616          }
 617          break;
 618
 619       case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
 620       case SHADER_OPCODE_BROADCAST:
 621          inst->src[i] = val;
 622          progress = true;
 623          break;
 624
 625       case BRW_OPCODE_MAD:
 626       case BRW_OPCODE_LRP:
 627          inst->src[i] = val;
 628          progress = true;
 629          break;
 630
 631       default:
 632          break;
 633       }
 634    }
 635
 636    return progress;
 637 }
 638
 639 static bool
 640 can_propagate_from(fs_inst *inst)
 641 {
 642    return (inst->opcode == BRW_OPCODE_MOV &&
 643            inst->dst.file == GRF &&
 644            ((inst->src[0].file == GRF &&
 645              (inst->src[0].reg != inst->dst.reg ||
 646               inst->src[0].reg_offset != inst->dst.reg_offset)) ||
 647             inst->src[0].file == ATTR ||
 648             inst->src[0].file == UNIFORM ||
 649             inst->src[0].file == IMM) &&
 650            inst->src[0].type == inst->dst.type &&
 651            !inst->is_partial_write());
 652 }
 653
 654 /* Walks a basic block and does copy propagation on it using the acp
 655  * list.
 656  */
 657 bool
 658 fs_visitor::opt_copy_propagate_local(void *copy_prop_ctx, bblock_t *block,
 659                                      exec_list *acp)
 660 {
 661    bool progress = false;
 662
 663    foreach_inst_in_block(fs_inst, inst, block) {
 664       /* Try propagating into this instruction. */
 665       for (int i = 0; i < inst->sources; i++) {
 666          if (inst->src[i].file != GRF)
 667             continue;
 668
 669          foreach_in_list(acp_entry, entry, &acp[inst->src[i].reg % ACP_HASH_SIZE]) {
 670             if (try_constant_propagate(inst, entry))
 671                progress = true;
 672
 673             if (try_copy_propagate(inst, i, entry))
 674                progress = true;
 675          }
 676       }
 677
 678       /* kill the destination from the ACP */
 679       if (inst->dst.file == GRF) {
 680          foreach_in_list_safe(acp_entry, entry, &acp[inst->dst.reg % ACP_HASH_SIZE]) {
 681             if (inst->overwrites_reg(entry->dst)) {
 682                entry->remove();
 683             }
 684          }
 685
 686          /* Oops, we only have the chaining hash based on the destination, not
 687           * the source, so walk across the entire table.
 688           */
 689          for (int i = 0; i < ACP_HASH_SIZE; i++) {
 690             foreach_in_list_safe(acp_entry, entry, &acp[i]) {
 691                if (inst->overwrites_reg(entry->src))
 692                   entry->remove();
 693             }
 694          }
 695       }
 696
 697       /* If this instruction's source could potentially be folded into the
 698        * operand of another instruction, add it to the ACP.
 699        */
 700       if (can_propagate_from(inst)) {
 701          acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
 702          entry->dst = inst->dst;
 703          entry->src = inst->src[0];
 704          entry->regs_written = inst->regs_written;
 705          entry->opcode = inst->opcode;
 706          entry->saturate = inst->saturate;
 707          acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 708       } else if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
 709                  inst->dst.file == GRF) {
 710          int offset = 0;
 711          for (int i = 0; i < inst->sources; i++) {
 712             int effective_width = i < inst->header_size ? 8 : inst->exec_size;
 713             int regs_written = effective_width / 8;
 714             if (inst->src[i].file == GRF) {
 715                acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
 716                entry->dst = inst->dst;
 717                entry->dst.reg_offset = offset;
 718                entry->dst.width = effective_width;
 719                entry->src = inst->src[i];
 720                entry->regs_written = regs_written;
 721                entry->opcode = inst->opcode;
 722                if (!entry->dst.equals(inst->src[i])) {
 723                   acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 724                } else {
 725                   ralloc_free(entry);
 726                }
 727             }
 728             offset += regs_written;
 729          }
 730       }
 731    }
 732
 733    return progress;
 734 }
 735
 736 bool
 737 fs_visitor::opt_copy_propagate()
 738 {
 739    bool progress = false;
 740    void *copy_prop_ctx = ralloc_context(NULL);
 741    exec_list *out_acp[cfg->num_blocks];
 742
 743    for (int i = 0; i < cfg->num_blocks; i++)
 744       out_acp[i] = new exec_list [ACP_HASH_SIZE];
 745
 746    /* First, walk through each block doing local copy propagation and getting
 747     * the set of copies available at the end of the block.
 748     */
 749    foreach_block (block, cfg) {
 750       progress = opt_copy_propagate_local(copy_prop_ctx, block,
 751                                           out_acp[block->num]) || progress;
 752    }
 753
 754    /* Do dataflow analysis for those available copies. */
 755    fs_copy_prop_dataflow dataflow(copy_prop_ctx, cfg, out_acp);
 756
 757    /* Next, re-run local copy propagation, this time with the set of copies
 758     * provided by the dataflow analysis available at the start of a block.
 759     */
 760    foreach_block (block, cfg) {
 761       exec_list in_acp[ACP_HASH_SIZE];
 762
 763       for (int i = 0; i < dataflow.num_acp; i++) {
 764          if (BITSET_TEST(dataflow.bd[block->num].livein, i)) {
 765             struct acp_entry *entry = dataflow.acp[i];
 766             in_acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 767          }
 768       }
 769
 770       progress = opt_copy_propagate_local(copy_prop_ctx, block, in_acp) || progress;
 771    }
 772
 773    for (int i = 0; i < cfg->num_blocks; i++)
 774       delete [] out_acp[i];
 775    ralloc_free(copy_prop_ctx);
 776
 777    if (progress)
 778       invalidate_live_intervals();
 779
 780    return progress;
 781 }