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->dst.type == inst->src[0].type &&
 283           (inst->opcode == BRW_OPCODE_MOV ||
 284            (inst->opcode == BRW_OPCODE_SEL &&
 285             inst->predicate != BRW_PREDICATE_NONE &&
 286             !inst->src[1].abs && !inst->src[1].negate));
 287 }
 288
 289 bool
 290 fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
 291 {
 292    if (inst->src[arg].file != GRF)
 293       return false;
 294
 295    if (entry->src.file == IMM)
 296       return false;
 297    assert(entry->src.file == GRF || entry->src.file == UNIFORM ||
 298           entry->src.file == ATTR);
 299
 300    if (entry->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
 301        inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD)
 302       return false;
 303
 304    assert(entry->dst.file == GRF);
 305    if (inst->src[arg].reg != entry->dst.reg)
 306       return false;
 307
 308    /* Bail if inst is reading a range that isn't contained in the range
 309     * that entry is writing.
 310     */
 311    if (inst->src[arg].reg_offset < entry->dst.reg_offset ||
 312        (inst->src[arg].reg_offset * 32 + inst->src[arg].subreg_offset +
 313         inst->regs_read(arg) * inst->src[arg].stride * 32) >
 314        (entry->dst.reg_offset + entry->regs_written) * 32)
 315       return false;
 316
 317    /* we can't generally copy-propagate UD negations because we
 318     * can end up accessing the resulting values as signed integers
 319     * instead. See also resolve_ud_negate() and comment in
 320     * fs_generator::generate_code.
 321     */
 322    if (entry->src.type == BRW_REGISTER_TYPE_UD &&
 323        entry->src.negate)
 324       return false;
 325
 326    bool has_source_modifiers = entry->src.abs || entry->src.negate;
 327
 328    if ((has_source_modifiers || entry->src.file == UNIFORM ||
 329         !entry->src.is_contiguous()) &&
 330        !inst->can_do_source_mods(devinfo))
 331       return false;
 332
 333    if (has_source_modifiers &&
 334        inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
 335       return false;
 336
 337    /* Bail if the result of composing both strides would exceed the
 338     * hardware limit.
 339     */
 340    if (entry->src.stride * inst->src[arg].stride > 4)
 341       return false;
 342
 343    /* Bail if the instruction type is larger than the execution type of the
 344     * copy, what implies that each channel is reading multiple channels of the
 345     * destination of the copy, and simply replacing the sources would give a
 346     * program with different semantics.
 347     */
 348    if (type_sz(entry->dst.type) < type_sz(inst->src[arg].type))
 349       return false;
 350
 351    /* Bail if the result of composing both strides cannot be expressed
 352     * as another stride. This avoids, for example, trying to transform
 353     * this:
 354     *
 355     *     MOV (8) rX<1>UD rY<0;1,0>UD
 356     *     FOO (8) ...     rX<8;8,1>UW
 357     *
 358     * into this:
 359     *
 360     *     FOO (8) ...     rY<0;1,0>UW
 361     *
 362     * Which would have different semantics.
 363     */
 364    if (entry->src.stride != 1 &&
 365        (inst->src[arg].stride *
 366         type_sz(inst->src[arg].type)) % type_sz(entry->src.type) != 0)
 367       return false;
 368
 369    if (has_source_modifiers &&
 370        entry->dst.type != inst->src[arg].type &&
 371        !can_change_source_types(inst))
 372       return false;
 373
 374    if (devinfo->gen >= 8 && (entry->src.negate || entry->src.abs) &&
 375        is_logic_op(inst->opcode)) {
 376       return false;
 377    }
 378
 379    if (entry->saturate) {
 380       switch(inst->opcode) {
 381       case BRW_OPCODE_SEL:
 382          if (inst->src[1].file != IMM ||
 383              inst->src[1].fixed_hw_reg.dw1.f < 0.0 ||
 384              inst->src[1].fixed_hw_reg.dw1.f > 1.0) {
 385             return false;
 386          }
 387          break;
 388       default:
 389          return false;
 390       }
 391    }
 392
 393    inst->src[arg].file = entry->src.file;
 394    inst->src[arg].reg = entry->src.reg;
 395    inst->src[arg].stride *= entry->src.stride;
 396    inst->saturate = inst->saturate || entry->saturate;
 397
 398    switch (entry->src.file) {
 399    case UNIFORM:
 400    case BAD_FILE:
 401    case HW_REG:
 402       inst->src[arg].reg_offset = entry->src.reg_offset;
 403       inst->src[arg].subreg_offset = entry->src.subreg_offset;
 404       break;
 405    case ATTR:
 406    case GRF:
 407       {
 408          /* In this case, we'll just leave the width alone.  The source
 409           * register could have different widths depending on how it is
 410           * being used.  For instance, if only half of the register was
 411           * used then we want to preserve that and continue to only use
 412           * half.
 413           *
 414           * Also, we have to deal with mapping parts of vgrfs to other
 415           * parts of vgrfs so we have to do some reg_offset magic.
 416           */
 417
 418          /* Compute the offset of inst->src[arg] relative to inst->dst */
 419          assert(entry->dst.subreg_offset == 0);
 420          int rel_offset = inst->src[arg].reg_offset - entry->dst.reg_offset;
 421          int rel_suboffset = inst->src[arg].subreg_offset;
 422
 423          /* Compute the final register offset (in bytes) */
 424          int offset = entry->src.reg_offset * 32 + entry->src.subreg_offset;
 425          offset += rel_offset * 32 + rel_suboffset;
 426          inst->src[arg].reg_offset = offset / 32;
 427          inst->src[arg].subreg_offset = offset % 32;
 428       }
 429       break;
 430    default:
 431       unreachable("Invalid register file");
 432       break;
 433    }
 434
 435    if (has_source_modifiers) {
 436       if (entry->dst.type != inst->src[arg].type) {
 437          /* We are propagating source modifiers from a MOV with a different
 438           * type.  If we got here, then we can just change the source and
 439           * destination types of the instruction and keep going.
 440           */
 441          assert(can_change_source_types(inst));
 442          for (int i = 0; i < inst->sources; i++) {
 443             inst->src[i].type = entry->dst.type;
 444          }
 445          inst->dst.type = entry->dst.type;
 446       }
 447
 448       if (!inst->src[arg].abs) {
 449          inst->src[arg].abs = entry->src.abs;
 450          inst->src[arg].negate ^= entry->src.negate;
 451       }
 452    }
 453
 454    return true;
 455 }
 456
 457
 458 bool
 459 fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
 460 {
 461    bool progress = false;
 462
 463    if (entry->src.file != IMM)
 464       return false;
 465    if (entry->saturate)
 466       return false;
 467
 468    for (int i = inst->sources - 1; i >= 0; i--) {
 469       if (inst->src[i].file != GRF)
 470          continue;
 471
 472       assert(entry->dst.file == GRF);
 473       if (inst->src[i].reg != entry->dst.reg)
 474          continue;
 475
 476       /* Bail if inst is reading a range that isn't contained in the range
 477        * that entry is writing.
 478        */
 479       if (inst->src[i].reg_offset < entry->dst.reg_offset ||
 480           (inst->src[i].reg_offset * 32 + inst->src[i].subreg_offset +
 481            inst->regs_read(i) * inst->src[i].stride * 32) >
 482           (entry->dst.reg_offset + entry->regs_written) * 32)
 483          continue;
 484
 485       fs_reg val = entry->src;
 486       val.type = inst->src[i].type;
 487
 488       if (inst->src[i].abs) {
 489          if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
 490              !brw_abs_immediate(val.type, &val.fixed_hw_reg)) {
 491             continue;
 492          }
 493       }
 494
 495       if (inst->src[i].negate) {
 496          if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
 497              !brw_negate_immediate(val.type, &val.fixed_hw_reg)) {
 498             continue;
 499          }
 500       }
 501
 502       switch (inst->opcode) {
 503       case BRW_OPCODE_MOV:
 504       case SHADER_OPCODE_LOAD_PAYLOAD:
 505          inst->src[i] = val;
 506          progress = true;
 507          break;
 508
 509       case SHADER_OPCODE_INT_QUOTIENT:
 510       case SHADER_OPCODE_INT_REMAINDER:
 511          /* FINISHME: Promote non-float constants and remove this. */
 512          if (devinfo->gen < 8)
 513             break;
 514          /* fallthrough */
 515       case SHADER_OPCODE_POW:
 516          /* Allow constant propagation into src1 (except on Gen 6), and let
 517           * constant combining promote the constant on Gen < 8.
 518           *
 519           * While Gen 6 MATH can take a scalar source, its source and
 520           * destination offsets must be equal and we cannot ensure that.
 521           */
 522          if (devinfo->gen == 6)
 523             break;
 524          /* fallthrough */
 525       case BRW_OPCODE_BFI1:
 526       case BRW_OPCODE_ASR:
 527       case BRW_OPCODE_SHL:
 528       case BRW_OPCODE_SHR:
 529       case BRW_OPCODE_SUBB:
 530          if (i == 1) {
 531             inst->src[i] = val;
 532             progress = true;
 533          }
 534          break;
 535
 536       case BRW_OPCODE_MACH:
 537       case BRW_OPCODE_MUL:
 538       case SHADER_OPCODE_MULH:
 539       case BRW_OPCODE_ADD:
 540       case BRW_OPCODE_OR:
 541       case BRW_OPCODE_AND:
 542       case BRW_OPCODE_XOR:
 543       case BRW_OPCODE_ADDC:
 544          if (i == 1) {
 545             inst->src[i] = val;
 546             progress = true;
 547          } else if (i == 0 && inst->src[1].file != IMM) {
 548             /* Fit this constant in by commuting the operands.
 549              * Exception: we can't do this for 32-bit integer MUL/MACH
 550              * because it's asymmetric.
 551              *
 552              * The BSpec says for Broadwell that
 553              *
 554              *    "When multiplying DW x DW, the dst cannot be accumulator."
 555              *
 556              * Integer MUL with a non-accumulator destination will be lowered
 557              * by lower_integer_multiplication(), so don't restrict it.
 558              */
 559             if (((inst->opcode == BRW_OPCODE_MUL &&
 560                   inst->dst.is_accumulator()) ||
 561                  inst->opcode == BRW_OPCODE_MACH) &&
 562                 (inst->src[1].type == BRW_REGISTER_TYPE_D ||
 563                  inst->src[1].type == BRW_REGISTER_TYPE_UD))
 564                break;
 565             inst->src[0] = inst->src[1];
 566             inst->src[1] = val;
 567             progress = true;
 568          }
 569          break;
 570
 571       case BRW_OPCODE_CMP:
 572       case BRW_OPCODE_IF:
 573          if (i == 1) {
 574             inst->src[i] = val;
 575             progress = true;
 576          } else if (i == 0 && inst->src[1].file != IMM) {
 577             enum brw_conditional_mod new_cmod;
 578
 579             new_cmod = brw_swap_cmod(inst->conditional_mod);
 580             if (new_cmod != BRW_CONDITIONAL_NONE) {
 581                /* Fit this constant in by swapping the operands and
 582                 * flipping the test
 583                 */
 584                inst->src[0] = inst->src[1];
 585                inst->src[1] = val;
 586                inst->conditional_mod = new_cmod;
 587                progress = true;
 588             }
 589          }
 590          break;
 591
 592       case BRW_OPCODE_SEL:
 593          if (i == 1) {
 594             inst->src[i] = val;
 595             progress = true;
 596          } else if (i == 0 && inst->src[1].file != IMM) {
 597             inst->src[0] = inst->src[1];
 598             inst->src[1] = val;
 599
 600             /* If this was predicated, flipping operands means
 601              * we also need to flip the predicate.
 602              */
 603             if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
 604                inst->predicate_inverse =
 605                   !inst->predicate_inverse;
 606             }
 607             progress = true;
 608          }
 609          break;
 610
 611       case SHADER_OPCODE_RCP:
 612          /* The hardware doesn't do math on immediate values
 613           * (because why are you doing that, seriously?), but
 614           * the correct answer is to just constant fold it
 615           * anyway.
 616           */
 617          assert(i == 0);
 618          if (inst->src[0].fixed_hw_reg.dw1.f != 0.0f) {
 619             inst->opcode = BRW_OPCODE_MOV;
 620             inst->src[0] = val;
 621             inst->src[0].fixed_hw_reg.dw1.f = 1.0f / inst->src[0].fixed_hw_reg.dw1.f;
 622             progress = true;
 623          }
 624          break;
 625
 626       case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
 627       case SHADER_OPCODE_BROADCAST:
 628          inst->src[i] = val;
 629          progress = true;
 630          break;
 631
 632       case BRW_OPCODE_MAD:
 633       case BRW_OPCODE_LRP:
 634          inst->src[i] = val;
 635          progress = true;
 636          break;
 637
 638       default:
 639          break;
 640       }
 641    }
 642
 643    return progress;
 644 }
 645
 646 static bool
 647 can_propagate_from(fs_inst *inst)
 648 {
 649    return (inst->opcode == BRW_OPCODE_MOV &&
 650            inst->dst.file == GRF &&
 651            ((inst->src[0].file == GRF &&
 652              (inst->src[0].reg != inst->dst.reg ||
 653               inst->src[0].reg_offset != inst->dst.reg_offset)) ||
 654             inst->src[0].file == ATTR ||
 655             inst->src[0].file == UNIFORM ||
 656             inst->src[0].file == IMM) &&
 657            inst->src[0].type == inst->dst.type &&
 658            !inst->is_partial_write());
 659 }
 660
 661 /* Walks a basic block and does copy propagation on it using the acp
 662  * list.
 663  */
 664 bool
 665 fs_visitor::opt_copy_propagate_local(void *copy_prop_ctx, bblock_t *block,
 666                                      exec_list *acp)
 667 {
 668    bool progress = false;
 669
 670    foreach_inst_in_block(fs_inst, inst, block) {
 671       /* Try propagating into this instruction. */
 672       for (int i = 0; i < inst->sources; i++) {
 673          if (inst->src[i].file != GRF)
 674             continue;
 675
 676          foreach_in_list(acp_entry, entry, &acp[inst->src[i].reg % ACP_HASH_SIZE]) {
 677             if (try_constant_propagate(inst, entry))
 678                progress = true;
 679
 680             if (try_copy_propagate(inst, i, entry))
 681                progress = true;
 682          }
 683       }
 684
 685       /* kill the destination from the ACP */
 686       if (inst->dst.file == GRF) {
 687          foreach_in_list_safe(acp_entry, entry, &acp[inst->dst.reg % ACP_HASH_SIZE]) {
 688             if (inst->overwrites_reg(entry->dst)) {
 689                entry->remove();
 690             }
 691          }
 692
 693          /* Oops, we only have the chaining hash based on the destination, not
 694           * the source, so walk across the entire table.
 695           */
 696          for (int i = 0; i < ACP_HASH_SIZE; i++) {
 697             foreach_in_list_safe(acp_entry, entry, &acp[i]) {
 698                if (inst->overwrites_reg(entry->src))
 699                   entry->remove();
 700             }
 701          }
 702       }
 703
 704       /* If this instruction's source could potentially be folded into the
 705        * operand of another instruction, add it to the ACP.
 706        */
 707       if (can_propagate_from(inst)) {
 708          acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
 709          entry->dst = inst->dst;
 710          entry->src = inst->src[0];
 711          entry->regs_written = inst->regs_written;
 712          entry->opcode = inst->opcode;
 713          entry->saturate = inst->saturate;
 714          acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 715       } else if (inst->opcode == SHADER_OPCODE_LOAD_PAYLOAD &&
 716                  inst->dst.file == GRF) {
 717          int offset = 0;
 718          for (int i = 0; i < inst->sources; i++) {
 719             int effective_width = i < inst->header_size ? 8 : inst->exec_size;
 720             int regs_written = effective_width / 8;
 721             if (inst->src[i].file == GRF) {
 722                acp_entry *entry = ralloc(copy_prop_ctx, acp_entry);
 723                entry->dst = inst->dst;
 724                entry->dst.reg_offset = offset;
 725                entry->src = inst->src[i];
 726                entry->regs_written = regs_written;
 727                entry->opcode = inst->opcode;
 728                if (!entry->dst.equals(inst->src[i])) {
 729                   acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 730                } else {
 731                   ralloc_free(entry);
 732                }
 733             }
 734             offset += regs_written;
 735          }
 736       }
 737    }
 738
 739    return progress;
 740 }
 741
 742 bool
 743 fs_visitor::opt_copy_propagate()
 744 {
 745    bool progress = false;
 746    void *copy_prop_ctx = ralloc_context(NULL);
 747    exec_list *out_acp[cfg->num_blocks];
 748
 749    for (int i = 0; i < cfg->num_blocks; i++)
 750       out_acp[i] = new exec_list [ACP_HASH_SIZE];
 751
 752    /* First, walk through each block doing local copy propagation and getting
 753     * the set of copies available at the end of the block.
 754     */
 755    foreach_block (block, cfg) {
 756       progress = opt_copy_propagate_local(copy_prop_ctx, block,
 757                                           out_acp[block->num]) || progress;
 758    }
 759
 760    /* Do dataflow analysis for those available copies. */
 761    fs_copy_prop_dataflow dataflow(copy_prop_ctx, cfg, out_acp);
 762
 763    /* Next, re-run local copy propagation, this time with the set of copies
 764     * provided by the dataflow analysis available at the start of a block.
 765     */
 766    foreach_block (block, cfg) {
 767       exec_list in_acp[ACP_HASH_SIZE];
 768
 769       for (int i = 0; i < dataflow.num_acp; i++) {
 770          if (BITSET_TEST(dataflow.bd[block->num].livein, i)) {
 771             struct acp_entry *entry = dataflow.acp[i];
 772             in_acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 773          }
 774       }
 775
 776       progress = opt_copy_propagate_local(copy_prop_ctx, block, in_acp) || progress;
 777    }
 778
 779    for (int i = 0; i < cfg->num_blocks; i++)
 780       delete [] out_acp[i];
 781    ralloc_free(copy_prop_ctx);
 782
 783    if (progress)
 784       invalidate_live_intervals();
 785
 786    return progress;
 787 }