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