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       printf("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          printf("%d ", parent->block_num);
 258       }
 259       printf("):\n");
 260       printf("       livein = 0x");
 261       for (int i = 0; i < bitset_words; i++)
 262          printf("%08x", bd[b].livein[i]);
 263       printf(", liveout = 0x");
 264       for (int i = 0; i < bitset_words; i++)
 265          printf("%08x", bd[b].liveout[i]);
 266       printf(",\n       copy   = 0x");
 267       for (int i = 0; i < bitset_words; i++)
 268          printf("%08x", bd[b].copy[i]);
 269       printf(", kill    = 0x");
 270       for (int i = 0; i < bitset_words; i++)
 271          printf("%08x", bd[b].kill[i]);
 272       printf("\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    if (inst->regs_read(this, arg) > 1)
 283       return false;
 284
 285    if (inst->src[arg].file != entry->dst.file ||
 286        inst->src[arg].reg != entry->dst.reg ||
 287        inst->src[arg].reg_offset != entry->dst.reg_offset) {
 288       return false;
 289    }
 290
 291    /* See resolve_ud_negate() and comment in brw_fs_emit.cpp. */
 292    if (inst->conditional_mod &&
 293        inst->src[arg].type == BRW_REGISTER_TYPE_UD &&
 294        entry->src.negate)
 295       return false;
 296
 297    bool has_source_modifiers = entry->src.abs || entry->src.negate;
 298
 299    if ((has_source_modifiers || entry->src.file == UNIFORM ||
 300         entry->src.smear != -1) && !can_do_source_mods(inst))
 301       return false;
 302
 303    if (has_source_modifiers && entry->dst.type != inst->src[arg].type)
 304       return false;
 305
 306    inst->src[arg].file = entry->src.file;
 307    inst->src[arg].reg = entry->src.reg;
 308    inst->src[arg].reg_offset = entry->src.reg_offset;
 309    if (entry->src.smear != -1)
 310       inst->src[arg].smear = entry->src.smear;
 311
 312    if (!inst->src[arg].abs) {
 313       inst->src[arg].abs = entry->src.abs;
 314       inst->src[arg].negate ^= entry->src.negate;
 315    }
 316
 317    return true;
 318 }
 319
 320
 321 bool
 322 fs_visitor::try_constant_propagate(fs_inst *inst, acp_entry *entry)
 323 {
 324    bool progress = false;
 325
 326    if (entry->src.file != IMM)
 327       return false;
 328
 329    for (int i = 2; i >= 0; i--) {
 330       if (inst->src[i].file != entry->dst.file ||
 331           inst->src[i].reg != entry->dst.reg ||
 332           inst->src[i].reg_offset != entry->dst.reg_offset)
 333          continue;
 334
 335       /* Don't bother with cases that should have been taken care of by the
 336        * GLSL compiler's constant folding pass.
 337        */
 338       if (inst->src[i].negate || inst->src[i].abs)
 339          continue;
 340
 341       switch (inst->opcode) {
 342       case BRW_OPCODE_MOV:
 343          inst->src[i] = entry->src;
 344          progress = true;
 345          break;
 346
 347       case BRW_OPCODE_SHL:
 348       case BRW_OPCODE_SHR:
 349       case BRW_OPCODE_ADDC:
 350       case BRW_OPCODE_SUBB:
 351          if (i == 1) {
 352             inst->src[i] = entry->src;
 353             progress = true;
 354          }
 355          break;
 356
 357       case BRW_OPCODE_MACH:
 358       case BRW_OPCODE_MUL:
 359       case BRW_OPCODE_ADD:
 360       case BRW_OPCODE_OR:
 361       case BRW_OPCODE_AND:
 362       case BRW_OPCODE_XOR:
 363          if (i == 1) {
 364             inst->src[i] = entry->src;
 365             progress = true;
 366          } else if (i == 0 && inst->src[1].file != IMM) {
 367             /* Fit this constant in by commuting the operands.
 368              * Exception: we can't do this for 32-bit integer MUL/MACH
 369              * because it's asymmetric.
 370              */
 371             if ((inst->opcode == BRW_OPCODE_MUL ||
 372                  inst->opcode == BRW_OPCODE_MACH) &&
 373                 (inst->src[1].type == BRW_REGISTER_TYPE_D ||
 374                  inst->src[1].type == BRW_REGISTER_TYPE_UD))
 375                break;
 376             inst->src[0] = inst->src[1];
 377             inst->src[1] = entry->src;
 378             progress = true;
 379          }
 380          break;
 381
 382       case BRW_OPCODE_CMP:
 383       case BRW_OPCODE_IF:
 384          if (i == 1) {
 385             inst->src[i] = entry->src;
 386             progress = true;
 387          } else if (i == 0 && inst->src[1].file != IMM) {
 388             uint32_t new_cmod;
 389
 390             new_cmod = brw_swap_cmod(inst->conditional_mod);
 391             if (new_cmod != ~0u) {
 392                /* Fit this constant in by swapping the operands and
 393                 * flipping the test
 394                 */
 395                inst->src[0] = inst->src[1];
 396                inst->src[1] = entry->src;
 397                inst->conditional_mod = new_cmod;
 398                progress = true;
 399             }
 400          }
 401          break;
 402
 403       case BRW_OPCODE_SEL:
 404          if (i == 1) {
 405             inst->src[i] = entry->src;
 406             progress = true;
 407          } else if (i == 0 && inst->src[1].file != IMM) {
 408             inst->src[0] = inst->src[1];
 409             inst->src[1] = entry->src;
 410
 411             /* If this was predicated, flipping operands means
 412              * we also need to flip the predicate.
 413              */
 414             if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
 415                inst->predicate_inverse =
 416                   !inst->predicate_inverse;
 417             }
 418             progress = true;
 419          }
 420          break;
 421
 422       case SHADER_OPCODE_RCP:
 423          /* The hardware doesn't do math on immediate values
 424           * (because why are you doing that, seriously?), but
 425           * the correct answer is to just constant fold it
 426           * anyway.
 427           */
 428          assert(i == 0);
 429          if (inst->src[0].imm.f != 0.0f) {
 430             inst->opcode = BRW_OPCODE_MOV;
 431             inst->src[0] = entry->src;
 432             inst->src[0].imm.f = 1.0f / inst->src[0].imm.f;
 433             progress = true;
 434          }
 435          break;
 436
 437       case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
 438          inst->src[i] = entry->src;
 439          progress = true;
 440          break;
 441
 442       default:
 443          break;
 444       }
 445    }
 446
 447    return progress;
 448 }
 449 /* Walks a basic block and does copy propagation on it using the acp
 450  * list.
 451  */
 452 bool
 453 fs_visitor::opt_copy_propagate_local(void *mem_ctx, bblock_t *block,
 454                                      exec_list *acp)
 455 {
 456    bool progress = false;
 457
 458    for (fs_inst *inst = (fs_inst *)block->start;
 459         inst != block->end->next;
 460         inst = (fs_inst *)inst->next) {
 461
 462       /* Try propagating into this instruction. */
 463       for (int i = 0; i < 3; i++) {
 464          if (inst->src[i].file != GRF)
 465             continue;
 466
 467          foreach_list(entry_node, &acp[inst->src[i].reg % ACP_HASH_SIZE]) {
 468             acp_entry *entry = (acp_entry *)entry_node;
 469
 470             if (try_constant_propagate(inst, entry))
 471                progress = true;
 472
 473             if (try_copy_propagate(inst, i, entry))
 474                progress = true;
 475          }
 476       }
 477
 478       /* kill the destination from the ACP */
 479       if (inst->dst.file == GRF) {
 480          foreach_list_safe(entry_node, &acp[inst->dst.reg % ACP_HASH_SIZE]) {
 481             acp_entry *entry = (acp_entry *)entry_node;
 482
 483             if (inst->overwrites_reg(entry->dst)) {
 484                entry->remove();
 485             }
 486          }
 487
 488          /* Oops, we only have the chaining hash based on the destination, not
 489           * the source, so walk across the entire table.
 490           */
 491          for (int i = 0; i < ACP_HASH_SIZE; i++) {
 492             foreach_list_safe(entry_node, &acp[i]) {
 493                acp_entry *entry = (acp_entry *)entry_node;
 494                if (inst->overwrites_reg(entry->src))
 495                   entry->remove();
 496             }
 497          }
 498       }
 499
 500       /* If this instruction's source could potentially be folded into the
 501        * operand of another instruction, add it to the ACP.
 502        */
 503       if (inst->opcode == BRW_OPCODE_MOV &&
 504           inst->dst.file == GRF &&
 505           ((inst->src[0].file == GRF &&
 506             (inst->src[0].reg != inst->dst.reg ||
 507              inst->src[0].reg_offset != inst->dst.reg_offset)) ||
 508            inst->src[0].file == UNIFORM ||
 509            inst->src[0].file == IMM) &&
 510           inst->src[0].type == inst->dst.type &&
 511           !inst->saturate &&
 512           !inst->is_partial_write()) {
 513          acp_entry *entry = ralloc(mem_ctx, acp_entry);
 514          entry->dst = inst->dst;
 515          entry->src = inst->src[0];
 516          acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 517       }
 518    }
 519
 520    return progress;
 521 }
 522
 523 bool
 524 fs_visitor::opt_copy_propagate()
 525 {
 526    bool progress = false;
 527    void *mem_ctx = ralloc_context(this->mem_ctx);
 528    cfg_t cfg(this);
 529    exec_list *out_acp[cfg.num_blocks];
 530    for (int i = 0; i < cfg.num_blocks; i++)
 531       out_acp[i] = new exec_list [ACP_HASH_SIZE];
 532
 533    /* First, walk through each block doing local copy propagation and getting
 534     * the set of copies available at the end of the block.
 535     */
 536    for (int b = 0; b < cfg.num_blocks; b++) {
 537       bblock_t *block = cfg.blocks[b];
 538
 539       progress = opt_copy_propagate_local(mem_ctx, block,
 540                                           out_acp[b]) || progress;
 541    }
 542
 543    /* Do dataflow analysis for those available copies. */
 544    fs_copy_prop_dataflow dataflow(mem_ctx, &cfg, out_acp);
 545
 546    /* Next, re-run local copy propagation, this time with the set of copies
 547     * provided by the dataflow analysis available at the start of a block.
 548     */
 549    for (int b = 0; b < cfg.num_blocks; b++) {
 550       bblock_t *block = cfg.blocks[b];
 551       exec_list in_acp[ACP_HASH_SIZE];
 552
 553       for (int i = 0; i < dataflow.num_acp; i++) {
 554          if (BITSET_TEST(dataflow.bd[b].livein, i)) {
 555             struct acp_entry *entry = dataflow.acp[i];
 556             in_acp[entry->dst.reg % ACP_HASH_SIZE].push_tail(entry);
 557          }
 558       }
 559
 560       progress = opt_copy_propagate_local(mem_ctx, block, in_acp) || progress;
 561    }
 562
 563    for (int i = 0; i < cfg.num_blocks; i++)
 564       delete [] out_acp[i];
 565    ralloc_free(mem_ctx);
 566
 567    if (progress)
 568       invalidate_live_intervals();
 569
 570    return progress;
 571 }