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 };
  46
  47 struct block_data {
  48    /**
  49     * Which entries in the fs_copy_prop_dataflow acp table are live at the
  50     * start of this block.  This is the useful output of the analysis, since
  51     * it lets us plug those into the local copy propagation on the second
  52     * pass.
  53     */
  54    BITSET_WORD *livein;
  55
  56    /**
  57     * Which entries in the fs_copy_prop_dataflow acp table are live at the end
  58     * of this block.  This is done in initial setup from the per-block acps
  59     * returned by the first local copy prop pass.
  60     */
  61    BITSET_WORD *liveout;
  62
  63    /**
  64     * Which entries in the fs_copy_prop_dataflow acp table are generated by
  65     * instructions in this block which reach the end of the block without
  66     * being killed.
  67     */
  68    BITSET_WORD *copy;
  69
  70    /**
  71     * Which entries in the fs_copy_prop_dataflow acp table are killed over the
  72     * course of this block.
  73     */
  74    BITSET_WORD *kill;
  75 };
  76
  77 class fs_copy_prop_dataflow
  78 {
  79 public:
  80    fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
  81                          exec_list *out_acp[ACP_HASH_SIZE]);
  82
  83    void setup_initial_values();
  84    void run();
  85
  86    void dump_block_data() const;
  87
  88    void *mem_ctx;
  89    cfg_t *cfg;
  90
  91    acp_entry **acp;
  92    int num_acp;
  93    int bitset_words;
  94
  95   struct block_data *bd;
  96 };
  97 } /* anonymous namespace */
  98
  99 fs_copy_prop_dataflow::fs_copy_prop_dataflow(void *mem_ctx, cfg_t *cfg,
 100                                              exec_list *out_acp[ACP_HASH_SIZE])
 101    : mem_ctx(mem_ctx), cfg(cfg)
 102 {
 103    bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
 104
 105    num_acp = 0;
 106    for (int b = 0; b < cfg->num_blocks; b++) {
 107       for (int i = 0; i < ACP_HASH_SIZE; i++) {
 108          foreach_list(entry_node, &out_acp[b][i]) {
 109             num_acp++;
 110          }
 111       }
 112    }
 113
 114    acp = rzalloc_array(mem_ctx, struct acp_entry *, num_acp);
 115
 116    bitset_words = BITSET_WORDS(num_acp);
 117
 118    int next_acp = 0;
 119    for (int b = 0; b < cfg->num_blocks; b++) {
 120       bd[b].livein = rzalloc_array(bd, BITSET_WORD, bitset_words);
 121       bd[b].liveout = rzalloc_array(bd, BITSET_WORD, bitset_words);
 122       bd[b].copy = rzalloc_array(bd, BITSET_WORD, bitset_words);
 123       bd[b].kill = rzalloc_array(bd, BITSET_WORD, bitset_words);
 124
 125       for (int i = 0; i < ACP_HASH_SIZE; i++) {
 126          foreach_list(entry_node, &out_acp[b][i]) {
 127             acp_entry *entry = (acp_entry *)entry_node;
 128
 129             acp[next_acp] = entry;
 130
 131             /* opt_copy_propagate_local populates out_acp with copies created
 132              * in a block which are still live at the end of the block.  This
 133              * is exactly what we want in the COPY set.
 134              */
 135             BITSET_SET(bd[b].copy, next_acp);
 136
 137             next_acp++;
 138          }
 139       }
 140    }
 141
 142    assert(next_acp == num_acp);
 143
 144    setup_initial_values();
 145    run();
 146 }
 147
 148 /**
 149  * Set up initial values for each of the data flow sets, prior to running
 150  * the fixed-point algorithm.
 151  */
 152 void
 153 fs_copy_prop_dataflow::setup_initial_values()
 154 {
 155    /* Initialize the COPY and KILL sets. */
 156    for (int b = 0; b < cfg->num_blocks; b++) {
 157       bblock_t *block = cfg->blocks[b];
 158
 159       for (fs_inst *inst = (fs_inst *)block->start;
 160            inst != block->end->next;
 161            inst = (fs_inst *)inst->next) {
 162          if (inst->dst.file != GRF)
 163             continue;
 164
 165          /* Mark ACP entries which are killed by this instruction. */
 166          for (int i = 0; i < num_acp; i++) {
 167             if (inst->overwrites_reg(acp[i]->dst) ||
 168                 inst->overwrites_reg(acp[i]->src)) {
 169                BITSET_SET(bd[b].kill, i);
 170             }
 171          }
 172       }
 173    }
 174
 175    /* Populate the initial values for the livein and liveout sets.  For the
 176     * block at the start of the program, livein = 0 and liveout = copy.
 177     * For the others, set liveout to 0 (the empty set) and livein to ~0
 178     * (the universal set).
 179     */
 180    for (int b = 0; b < cfg->num_blocks; b++) {
 181       bblock_t *block = cfg->blocks[b];
 182       if (block->parents.is_empty()) {
 183          for (int i = 0; i < bitset_words; i++) {
 184             bd[b].livein[i] = 0u;
 185             bd[b].liveout[i] = bd[b].copy[i];
 186          }
 187       } else {
 188          for (int i = 0; i < bitset_words; i++) {
 189             bd[b].liveout[i] = 0u;
 190             bd[b].livein[i] = ~0u;
 191          }
 192       }
 193    }
 194 }
 195
 196 /**
 197  * Walk the set of instructions in the block, marking which entries in the acp
 198  * are killed by the block.
 199  */
 200 void
 201 fs_copy_prop_dataflow::run()
 202 {
 203    bool progress;
 204
 205    do {
 206       progress = false;
 207
 208       /* Update liveout for all blocks. */
 209       for (int b = 0; b < cfg->num_blocks; b++) {
 210          if (cfg->blocks[b]->parents.is_empty())
 211             continue;
 212
 213          for (int i = 0; i < bitset_words; i++) {
 214             const BITSET_WORD old_liveout = bd[b].liveout[i];
 215
 216             bd[b].liveout[i] =
 217                bd[b].copy[i] | (bd[b].livein[i] & ~bd[b].kill[i]);
 218
 219             if (old_liveout != bd[b].liveout[i])
 220                progress = true;
 221          }
 222       }
 223
 224       /* Update livein for all blocks.  If a copy is live out of all parent
 225        * blocks, it's live coming in to this block.
 226        */
 227       for (int b = 0; b < cfg->num_blocks; b++) {
 228          if (cfg->blocks[b]->parents.is_empty())
 229             continue;
 230
 231          for (int i = 0; i < bitset_words; i++) {
 232             const BITSET_WORD old_livein = bd[b].livein[i];
 233
 234             bd[b].livein[i] = ~0u;
 235             foreach_list(block_node, &cfg->blocks[b]->parents) {
 236                bblock_link *link = (bblock_link *)block_node;
 237                bblock_t *block = link->block;
 238                bd[b].livein[i] &= bd[block->block_num].liveout[i];
 239             }
 240
 241             if (old_livein != bd[b].livein[i])
 242                progress = true;
 243          }
 244       }
 245    } while (progress);
 246 }
 247
 248 void
 249 fs_copy_prop_dataflow::dump_block_data() const
 250 {
 251    for (int b = 0; b < cfg->num_blocks; b++) {
 252       bblock_t *block = cfg->blocks[b];
 253       fprintf(stderr, "Block %d [%d, %d] (parents ", block->block_num,
 254              block->start_ip, block->end_ip);
 255       foreach_list(block_node, &block->parents) {
 256          bblock_t *parent = ((bblock_link *) block_node)->block;
 257          fprintf(stderr, "%d ", parent->block_num);
 258       }
 259       fprintf(stderr, "):\n");
 260       fprintf(stderr, "       livein = 0x");
 261       for (int i = 0; i < bitset_words; i++)
 262          fprintf(stderr, "%08x", bd[b].livein[i]);
 263       fprintf(stderr, ", liveout = 0x");
 264       for (int i = 0; i < bitset_words; i++)
 265          fprintf(stderr, "%08x", bd[b].liveout[i]);
 266       fprintf(stderr, ",\n       copy   = 0x");
 267       for (int i = 0; i < bitset_words; i++)
 268          fprintf(stderr, "%08x", bd[b].copy[i]);
 269       fprintf(stderr, ", kill    = 0x");
 270       for (int i = 0; i < bitset_words; i++)
 271          fprintf(stderr, "%08x", bd[b].kill[i]);
 272       fprintf(stderr, "\n");
 273    }
 274 }
 275
 276 bool
 277 fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
 278 {
 279    if (entry->src.file == IMM)
 280       return false;
 281
 282    /* Bail if inst is reading more than entry is writing. */
 283    if ((inst->regs_read(this, arg) * inst->src[arg].stride *
 284         type_sz(inst->src[arg].type)) > type_sz(entry->dst.type))
 285       return false;
 286
 287    if (inst->src[arg].file != entry->dst.file ||
 288        inst->src[arg].reg != entry->dst.reg ||
 289        inst->src[arg].reg_offset != entry->dst.reg_offset ||
 290        inst->src[arg].subreg_offset != entry->dst.subreg_offset) {
 291       return false;
 292    }
 293
 294    /* See resolve_ud_negate() and comment in brw_fs_emit.cpp. */
 295    if (inst->conditional_mod &&
 296        inst->src[arg].type == BRW_REGISTER_TYPE_UD &&
 297        entry->src.negate)
 298       return false;
 299
 300    bool has_source_modifiers = entry->src.abs || entry->src.negate;
 301
 302    if ((has_source_modifiers || entry->src.file == UNIFORM ||
 303         !entry->src.is_contiguous()) &&
 304        !can_do_source_mods(inst))
 305       return false;
 306
 307    /* Bail if the result of composing both strides would exceed the
 308     * hardware limit.
 309     */
 310    if (entry->src.stride * inst->src[arg].stride > 4)
 311       return false;
 312
 313    /* Bail if the result of composing both strides cannot be expressed
 314     * as another stride. This avoids, for example, trying to transform
 315     * this:
 316     *
 317     *     MOV (8) rX<1>UD rY<0;1,0>UD
 318     *     FOO (8) ...     rX<8;8,1>UW
 319     *
 320     * into this:
 321     *
 322     *     FOO (8) ...     rY<0;1,0>UW
 323     *
 324     * Which would have different semantics.
 325     */
 326    if (entry->src.stride != 1 &&
 327        (inst->src[arg].stride *
 328         type_sz(inst->src[arg].type)) % type_sz(entry->src.type) != 0)
 329       return false;
 330
 331    if (has_source_modifiers && entry->dst.type != inst->src[arg].type)
 332       return false;
 333
 334    inst->src[arg].file = entry->src.file;
 335    inst->src[arg].reg = entry->src.reg;
 336    inst->src[arg].reg_offset = entry->src.reg_offset;
 337    inst->src[arg].subreg_offset = entry->src.subreg_offset;
 338    inst->src[arg].stride *= entry->src.stride;
 339
 340    if (!inst->src[arg].abs) {
 341       inst->src[arg].abs = entry->src.abs;
 342       inst->src[arg].negate ^= entry->src.negate;
 343    }
 344
 345    return true;
 346 }
 347
 348
 349 bool
 350 fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
 351 {
 352    bool progress = false;
 353
 354    if (entry->src.file != IMM)
 355       return false;
 356
 357    for (int i = 2; i >= 0; i--) {
 358       if (inst->src[i].file != entry->dst.file ||
 359           inst->src[i].reg != entry->dst.reg ||
 360           inst->src[i].reg_offset != entry->dst.reg_offset ||
 361           inst->src[i].subreg_offset != entry->dst.subreg_offset ||
 362           inst->src[i].type != entry->dst.type ||
 363           inst->src[i].stride > 1)
 364          continue;
 365
 366       /* Don't bother with cases that should have been taken care of by the
 367        * GLSL compiler's constant folding pass.
 368        */
 369       if (inst->src[i].negate || inst->src[i].abs)
 370          continue;
 371
 372       switch (inst->opcode) {
 373       case BRW_OPCODE_MOV:
 374          inst->src[i] = entry->src;
 375          progress = true;
 376          break;
 377
 378       case BRW_OPCODE_BFI1:
 379       case BRW_OPCODE_ASR:
 380       case BRW_OPCODE_SHL:
 381       case BRW_OPCODE_SHR:
 382       case BRW_OPCODE_SUBB:
 383          if (i == 1) {
 384             inst->src[i] = entry->src;
 385             progress = true;
 386          }
 387          break;
 388
 389       case BRW_OPCODE_MACH:
 390       case BRW_OPCODE_MUL:
 391       case BRW_OPCODE_ADD:
 392       case BRW_OPCODE_OR:
 393       case BRW_OPCODE_AND:
 394       case BRW_OPCODE_XOR:
 395       case BRW_OPCODE_ADDC:
 396          if (i == 1) {
 397             inst->src[i] = entry->src;
 398             progress = true;
 399          } else if (i == 0 && inst->src[1].file != IMM) {
 400             /* Fit this constant in by commuting the operands.
 401              * Exception: we can't do this for 32-bit integer MUL/MACH
 402              * because it's asymmetric.
 403              */
 404             if ((inst->opcode == BRW_OPCODE_MUL ||
 405                  inst->opcode == BRW_OPCODE_MACH) &&
 406                 (inst->src[1].type == BRW_REGISTER_TYPE_D ||
 407                  inst->src[1].type == BRW_REGISTER_TYPE_UD))
 408                break;
 409             inst->src[0] = inst->src[1];
 410             inst->src[1] = entry->src;
 411             progress = true;
 412          }
 413          break;
 414
 415       case BRW_OPCODE_CMP:
 416       case BRW_OPCODE_IF:
 417          if (i == 1) {
 418             inst->src[i] = entry->src;
 419             progress = true;
 420          } else if (i == 0 && inst->src[1].file != IMM) {
 421             uint32_t new_cmod;
 422
 423             new_cmod = brw_swap_cmod(inst->conditional_mod);
 424             if (new_cmod != ~0u) {
 425                /* Fit this constant in by swapping the operands and
 426                 * flipping the test
 427                 */
 428                inst->src[0] = inst->src[1];
 429                inst->src[1] = entry->src;
 430                inst->conditional_mod = new_cmod;
 431                progress = true;
 432             }
 433          }
 434          break;
 435
 436       case BRW_OPCODE_SEL:
 437          if (i == 1) {
 438             inst->src[i] = entry->src;
 439             progress = true;
 440          } else if (i == 0 && inst->src[1].file != IMM) {
 441             inst->src[0] = inst->src[1];
 442             inst->src[1] = entry->src;
 443
 444             /* If this was predicated, flipping operands means
 445              * we also need to flip the predicate.
 446              */
 447             if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
 448                inst->predicate_inverse =
 449                   !inst->predicate_inverse;
 450             }
 451             progress = true;
 452          }
 453          break;
 454
 455       case SHADER_OPCODE_RCP:
 456          /* The hardware doesn't do math on immediate values
 457           * (because why are you doing that, seriously?), but
 458           * the correct answer is to just constant fold it
 459           * anyway.
 460           */
 461          assert(i == 0);
 462          if (inst->src[0].imm.f != 0.0f) {
 463             inst->opcode = BRW_OPCODE_MOV;
 464             inst->src[0] = entry->src;
 465             inst->src[0].imm.f = 1.0f / inst->src[0].imm.f;
 466             progress = true;
 467          }
 468          break;
 469
 470       case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
 471          inst->src[i] = entry->src;
 472          progress = true;
 473          break;
 474
 475       default:
 476          break;
 477       }
 478    }
 479
 480    return progress;
 481 }
 482 /* Walks a basic block and does copy propagation on it using the acp
 483  * list.
 484  */
 485 bool
 486 fs_visitor::opt_copy_propagate_local(void *mem_ctx, bblock_t *block,
 487                                      exec_list *acp)
 488 {
 489    bool progress = false;
 490
 491    for (fs_inst *inst = (fs_inst *)block->start;
 492         inst != block->end->next;
 493         inst = (fs_inst *)inst->next) {
 494
 495       /* Try propagating into this instruction. */
 496       for (int i = 0; i < 3; i++) {
 497          if (inst->src[i].file != GRF)
 498             continue;
 499
 500          foreach_list(entry_node, &acp[inst->src[i].reg % ACP_HASH_SIZE]) {
 501             acp_entry *entry = (acp_entry *)entry_node;
 502
 503             if (try_constant_propagate(inst, entry))
 504                progress = true;
 505
 506             if (try_copy_propagate(inst, i, entry))
 507                progress = true;
 508          }
 509       }
 510
 511       /* kill the destination from the ACP */
 512       if (inst->dst.file == GRF) {
 513          foreach_list_safe(entry_node, &acp[inst->dst.reg % ACP_HASH_SIZE]) {
 514             acp_entry *entry = (acp_entry *)entry_node;
 515
 516             if (inst->overwrites_reg(entry->dst)) {
 517                entry->remove();
 518             }
 519          }
 520
 521          /* Oops, we only have the chaining hash based on the destination, not
 522           * the source, so walk across the entire table.
 523           */
 524          for (int i = 0; i < ACP_HASH_SIZE; i++) {
 525             foreach_list_safe(entry_node, &acp[i]) {
 526                acp_entry *entry = (acp_entry *)entry_node;
 527                if (inst->overwrites_reg(entry->src))
 528                   entry->remove();
 529             }
 530          }
 531       }
 532
 533       /* If this instruction's source could potentially be folded into the
 534        * operand of another instruction, add it to the ACP.
 535        */
 536       if (inst->opcode == BRW_OPCODE_MOV &&
 537           inst->dst.file == GRF &&
 538           ((inst->src[0].file == GRF &&
 539             (inst->src[0].reg != inst->dst.reg ||
 540              inst->src[0].reg_offset != inst->dst.reg_offset)) ||
 541            inst->src[0].file == UNIFORM ||
 542            inst->src[0].file == IMM) &&
 543           inst->src[0].type == inst->dst.type &&
 544           !inst->saturate &&
 545           !inst->is_partial_write()) {
 546          acp_entry *entry = ralloc(mem_ctx, acp_entry);
 547          entry->dst = inst->dst;
 548          entry->src = inst->src[0];
 549          acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 550       }
 551    }
 552
 553    return progress;
 554 }
 555
 556 bool
 557 fs_visitor::opt_copy_propagate()
 558 {
 559    bool progress = false;
 560    void *mem_ctx = ralloc_context(this->mem_ctx);
 561    cfg_t cfg(&instructions);
 562    exec_list *out_acp[cfg.num_blocks];
 563    for (int i = 0; i < cfg.num_blocks; i++)
 564       out_acp[i] = new exec_list [ACP_HASH_SIZE];
 565
 566    /* First, walk through each block doing local copy propagation and getting
 567     * the set of copies available at the end of the block.
 568     */
 569    for (int b = 0; b < cfg.num_blocks; b++) {
 570       bblock_t *block = cfg.blocks[b];
 571
 572       progress = opt_copy_propagate_local(mem_ctx, block,
 573                                           out_acp[b]) || progress;
 574    }
 575
 576    /* Do dataflow analysis for those available copies. */
 577    fs_copy_prop_dataflow dataflow(mem_ctx, &cfg, out_acp);
 578
 579    /* Next, re-run local copy propagation, this time with the set of copies
 580     * provided by the dataflow analysis available at the start of a block.
 581     */
 582    for (int b = 0; b < cfg.num_blocks; b++) {
 583       bblock_t *block = cfg.blocks[b];
 584       exec_list in_acp[ACP_HASH_SIZE];
 585
 586       for (int i = 0; i < dataflow.num_acp; i++) {
 587          if (BITSET_TEST(dataflow.bd[b].livein, i)) {
 588             struct acp_entry *entry = dataflow.acp[i];
 589             in_acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 590          }
 591       }
 592
 593       progress = opt_copy_propagate_local(mem_ctx, block, in_acp) || progress;
 594    }
 595
 596    for (int i = 0; i < cfg.num_blocks; i++)
 597       delete [] out_acp[i];
 598    ralloc_free(mem_ctx);
 599
 600    if (progress)
 601       invalidate_live_intervals();
 602
 603    return progress;
 604 }