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