src/gallium/drivers/nvc0/nvc0_pc.c

   1 /*
   2  * Copyright 2010 Christoph Bumiller
   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 shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #define NOUVEAU_DEBUG 1
  24
  25 #include "nvc0_pc.h"
  26 #include "nvc0_program.h"
  27
  28 boolean
  29 nvc0_insn_can_load(struct nv_instruction *nvi, int s,
  30                    struct nv_instruction *ld)
  31 {
  32    int i;
  33
  34    if (ld->opcode == NV_OP_MOV && ld->src[0]->value->reg.file == NV_FILE_IMM) {
  35       if (s > 1 || !(nvc0_op_info_table[nvi->opcode].immediate & (1 << s)))
  36          return FALSE;
  37       if (!(nvc0_op_info_table[nvi->opcode].immediate & 4))
  38          if (ld->src[0]->value->reg.imm.u32 & 0xfff)
  39             return FALSE;
  40    } else
  41    if (!(nvc0_op_info_table[nvi->opcode].memory & (1 << s)))
  42       return FALSE;
  43
  44    if (ld->indirect >= 0)
  45       return FALSE;
  46
  47    for (i = 0; i < 3 && nvi->src[i]; ++i)
  48       if (nvi->src[i]->value->reg.file == NV_FILE_IMM)
  49          return FALSE;
  50
  51    return TRUE;
  52 }
  53
  54 /* Return whether this instruction can be executed conditionally. */
  55 boolean
  56 nvc0_insn_is_predicateable(struct nv_instruction *nvi)
  57 {
  58    int s;
  59
  60    if (!nv_op_predicateable(nvi->opcode))
  61       return FALSE;
  62    if (nvi->predicate >= 0)
  63       return FALSE;
  64    for (s = 0; s < 4 && nvi->src[s]; ++s)
  65       if (nvi->src[s]->value->reg.file == NV_FILE_IMM)
  66          return FALSE;
  67    return TRUE;
  68 }
  69
  70 int
  71 nvc0_insn_refcount(struct nv_instruction *nvi)
  72 {
  73    int rc = 0;
  74    int i;
  75    for (i = 0; i < 5 && nvi->def[i]; ++i) {
  76       if (!nvi->def[i])
  77          return rc;
  78       rc += nvi->def[i]->refc;
  79    }
  80    return rc;
  81 }
  82
  83 int
  84 nvc0_pc_replace_value(struct nv_pc *pc,
  85                       struct nv_value *old_val,
  86                       struct nv_value *new_val)
  87 {
  88    int i, n, s;
  89
  90    if (old_val == new_val)
  91       return old_val->refc;
  92
  93    for (i = 0, n = 0; i < pc->num_refs; ++i) {
  94       if (pc->refs[i]->value == old_val) {
  95          ++n;
  96          for (s = 0; s < 6 && pc->refs[i]->insn->src[s]; ++s)
  97             if (pc->refs[i]->insn->src[s] == pc->refs[i])
  98                break;
  99          assert(s < 6);
 100          nv_reference(pc, pc->refs[i]->insn, s, new_val);
 101       }
 102    }
 103    return n;
 104 }
 105
 106 struct nv_value *
 107 nvc0_pc_find_constant(struct nv_ref *ref)
 108 {
 109    struct nv_value *src;
 110
 111    if (!ref)
 112       return NULL;
 113
 114    src = ref->value;
 115    while (src->insn && src->insn->opcode == NV_OP_MOV) {
 116       assert(!src->insn->src[0]->mod);
 117       src = src->insn->src[0]->value;
 118    }
 119    if ((src->reg.file == NV_FILE_IMM) ||
 120        (src->insn &&
 121         src->insn->opcode == NV_OP_LD &&
 122         src->insn->src[0]->value->reg.file >= NV_FILE_MEM_C(0) &&
 123         src->insn->src[0]->value->reg.file <= NV_FILE_MEM_C(15)))
 124       return src;
 125    return NULL;
 126 }
 127
 128 struct nv_value *
 129 nvc0_pc_find_immediate(struct nv_ref *ref)
 130 {
 131    struct nv_value *src = nvc0_pc_find_constant(ref);
 132
 133    return (src && src->reg.file == NV_FILE_IMM) ? src : NULL;
 134 }
 135
 136 static void
 137 nv_pc_free_refs(struct nv_pc *pc)
 138 {
 139    int i;
 140    for (i = 0; i < pc->num_refs; i += 64)
 141       FREE(pc->refs[i]);
 142    FREE(pc->refs);
 143 }
 144
 145 static const char *
 146 edge_name(ubyte type)
 147 {
 148    switch (type) {
 149    case CFG_EDGE_FORWARD: return "forward";
 150    case CFG_EDGE_BACK: return "back";
 151    case CFG_EDGE_LOOP_ENTER: return "loop";
 152    case CFG_EDGE_LOOP_LEAVE: return "break";
 153    case CFG_EDGE_FAKE: return "fake";
 154    default:
 155       return "?";
 156    }
 157 }
 158
 159 void
 160 nvc0_pc_pass_in_order(struct nv_basic_block *root, nv_pc_pass_func f,
 161                       void *priv)
 162 {
 163    struct nv_basic_block *bb[64], *bbb[16], *b;
 164    int j, p, pp;
 165
 166    bb[0] = root;
 167    p = 1;
 168    pp = 0;
 169
 170    while (p > 0) {
 171       b = bb[--p];
 172       b->priv = 0;
 173
 174       for (j = 1; j >= 0; --j) {
 175          if (!b->out[j])
 176             continue;
 177
 178          switch (b->out_kind[j]) {
 179          case CFG_EDGE_BACK:
 180             continue;
 181          case CFG_EDGE_FORWARD:
 182          case CFG_EDGE_FAKE:
 183             if (++b->out[j]->priv == b->out[j]->num_in)
 184                bb[p++] = b->out[j];
 185             break;
 186          case CFG_EDGE_LOOP_ENTER:
 187             bb[p++] = b->out[j];
 188             break;
 189          case CFG_EDGE_LOOP_LEAVE:
 190             bbb[pp++] = b->out[j];
 191             break;
 192          default:
 193             assert(0);
 194             break;
 195          }
 196       }
 197
 198       f(priv, b);
 199
 200       if (!p) {
 201          p = pp;
 202          for (; pp > 0; --pp)
 203             bb[pp - 1] = bbb[pp - 1];
 204       }
 205    }
 206 }
 207
 208 static void
 209 nv_do_print_function(void *priv, struct nv_basic_block *b)
 210 {
 211    struct nv_instruction *i;
 212
 213    debug_printf("=== BB %i ", b->id);
 214    if (b->out[0])
 215       debug_printf("[%s -> %i] ", edge_name(b->out_kind[0]), b->out[0]->id);
 216    if (b->out[1])
 217       debug_printf("[%s -> %i] ", edge_name(b->out_kind[1]), b->out[1]->id);
 218    debug_printf("===\n");
 219
 220    i = b->phi;
 221    if (!i)
 222       i = b->entry;
 223    for (; i; i = i->next)
 224       nvc0_print_instruction(i);
 225 }
 226
 227 void
 228 nvc0_print_function(struct nv_basic_block *root)
 229 {
 230    if (root->subroutine)
 231       debug_printf("SUBROUTINE %i\n", root->subroutine);
 232    else
 233       debug_printf("MAIN\n");
 234
 235    nvc0_pc_pass_in_order(root, nv_do_print_function, root);
 236 }
 237
 238 void
 239 nvc0_print_program(struct nv_pc *pc)
 240 {
 241    int i;
 242    for (i = 0; i < pc->num_subroutines + 1; ++i)
 243       if (pc->root[i])
 244          nvc0_print_function(pc->root[i]);
 245 }
 246
 247 #if NOUVEAU_DEBUG > 1
 248 static void
 249 nv_do_print_cfgraph(struct nv_pc *pc, FILE *f, struct nv_basic_block *b)
 250 {
 251    int i;
 252
 253    b->pass_seq = pc->pass_seq;
 254
 255    fprintf(f, "\t%i [shape=box]\n", b->id);
 256
 257    for (i = 0; i < 2; ++i) {
 258       if (!b->out[i])
 259          continue;
 260       switch (b->out_kind[i]) {
 261       case CFG_EDGE_FORWARD:
 262          fprintf(f, "\t%i -> %i;\n", b->id, b->out[i]->id);
 263          break;
 264       case CFG_EDGE_LOOP_ENTER:
 265          fprintf(f, "\t%i -> %i [color=green];\n", b->id, b->out[i]->id);
 266          break;
 267       case CFG_EDGE_LOOP_LEAVE:
 268          fprintf(f, "\t%i -> %i [color=red];\n", b->id, b->out[i]->id);
 269          break;
 270       case CFG_EDGE_BACK:
 271          fprintf(f, "\t%i -> %i;\n", b->id, b->out[i]->id);
 272          continue;
 273       case CFG_EDGE_FAKE:
 274          fprintf(f, "\t%i -> %i [style=dotted];\n", b->id, b->out[i]->id);
 275          break;
 276       default:
 277          assert(0);
 278          break;
 279       }
 280       if (b->out[i]->pass_seq < pc->pass_seq)
 281          nv_do_print_cfgraph(pc, f, b->out[i]);
 282    }
 283 }
 284
 285 /* Print the control flow graph of subroutine @subr (0 == MAIN) to a file. */
 286 static void
 287 nv_print_cfgraph(struct nv_pc *pc, const char *filepath, int subr)
 288 {
 289    FILE *f;
 290
 291    f = fopen(filepath, "a");
 292    if (!f)
 293       return;
 294
 295    fprintf(f, "digraph G {\n");
 296
 297    ++pc->pass_seq;
 298
 299    nv_do_print_cfgraph(pc, f, pc->root[subr]);
 300
 301    fprintf(f, "}\n");
 302
 303    fclose(f);
 304 }
 305 #endif
 306
 307 static INLINE void
 308 nvc0_pc_print_binary(struct nv_pc *pc)
 309 {
 310    unsigned i;
 311
 312    NOUVEAU_DBG("nvc0_pc_print_binary(%u ops)\n", pc->emit_size / 8);
 313
 314    for (i = 0; i < pc->emit_size / 4; i += 2) {
 315       debug_printf("0x%08x ", pc->emit[i + 0]);
 316       debug_printf("0x%08x ", pc->emit[i + 1]);
 317       if ((i % 16) == 15)
 318          debug_printf("\n");
 319    }
 320    debug_printf("\n");
 321 }
 322
 323 static int
 324 nvc0_emit_program(struct nv_pc *pc)
 325 {
 326    uint32_t *code = pc->emit;
 327    int n;
 328
 329    NOUVEAU_DBG("emitting program: size = %u\n", pc->emit_size);
 330
 331    for (n = 0; n < pc->num_blocks; ++n) {
 332       struct nv_instruction *i;
 333       struct nv_basic_block *b = pc->bb_list[n];
 334
 335       for (i = b->entry; i; i = i->next) {
 336          nvc0_emit_instruction(pc, i);
 337          pc->emit += 2;
 338          pc->emit_pos += 2;
 339       }
 340    }
 341    assert(pc->emit == &code[pc->emit_size / 4]);
 342
 343    pc->emit[0] = 0x00001de7;
 344    pc->emit[1] = 0x80000000;
 345    pc->emit_size += 8;
 346
 347    pc->emit = code;
 348
 349 #ifdef NOUVEAU_DEBUG
 350    nvc0_pc_print_binary(pc);
 351 #else
 352    debug_printf("not printing binary\n");
 353 #endif
 354    return 0;
 355 }
 356
 357 int
 358 nvc0_generate_code(struct nvc0_translation_info *ti)
 359 {
 360    struct nv_pc *pc;
 361    int ret;
 362    int i;
 363
 364    pc = CALLOC_STRUCT(nv_pc);
 365    if (!pc)
 366       return 1;
 367
 368    pc->is_fragprog = ti->prog->type == PIPE_SHADER_FRAGMENT;
 369
 370    pc->root = CALLOC(ti->num_subrs + 1, sizeof(pc->root[0]));
 371    if (!pc->root) {
 372       FREE(pc);
 373       return 1;
 374    }
 375    pc->num_subroutines = ti->num_subrs;
 376
 377    ret = nvc0_tgsi_to_nc(pc, ti);
 378    if (ret)
 379       goto out;
 380 #if NOUVEAU_DEBUG > 1
 381    nvc0_print_program(pc);
 382 #endif
 383
 384    pc->opt_reload_elim = ti->require_stores ? FALSE : TRUE;
 385
 386    /* optimization */
 387    ret = nvc0_pc_exec_pass0(pc);
 388    if (ret)
 389       goto out;
 390 #ifdef NOUVEAU_DEBUG
 391    nvc0_print_program(pc);
 392 #endif
 393
 394    /* register allocation */
 395    ret = nvc0_pc_exec_pass1(pc);
 396    if (ret)
 397       goto out;
 398 #if NOUVEAU_DEBUG > 1
 399    nv_print_program(pc);
 400    nv_print_cfgraph(pc, "nvc0_shader_cfgraph.dot", 0);
 401 #endif
 402
 403    /* prepare for emission */
 404    ret = nvc0_pc_exec_pass2(pc);
 405    if (ret)
 406       goto out;
 407    assert(!(pc->emit_size % 8));
 408
 409    pc->emit = CALLOC(pc->emit_size / 4 + 2, 4);
 410    if (!pc->emit) {
 411       ret = 3;
 412       goto out;
 413    }
 414    ret = nvc0_emit_program(pc);
 415    if (ret)
 416       goto out;
 417
 418    ti->prog->code = pc->emit;
 419    ti->prog->code_base = 0;
 420    ti->prog->code_size = pc->emit_size;
 421    ti->prog->parm_size = 0;
 422
 423    ti->prog->max_gpr = MAX2(4, pc->max_reg[NV_FILE_GPR] + 1);
 424
 425    ti->prog->relocs = pc->reloc_entries;
 426    ti->prog->num_relocs = pc->num_relocs;
 427
 428    NOUVEAU_DBG("SHADER TRANSLATION - %s\n", ret ? "failure" : "success");
 429
 430 out:
 431    nv_pc_free_refs(pc);
 432
 433    for (i = 0; i < pc->num_blocks; ++i)
 434       FREE(pc->bb_list[i]);
 435    if (pc->root)
 436       FREE(pc->root);
 437    if (ret) {
 438       /* on success, these will be referenced by struct nvc0_program */
 439       if (pc->emit)
 440          FREE(pc->emit);
 441       if (pc->immd_buf)
 442          FREE(pc->immd_buf);
 443       if (pc->reloc_entries)
 444          FREE(pc->reloc_entries);
 445    }
 446    FREE(pc);
 447    return ret;
 448 }
 449
 450 static void
 451 nvbb_insert_phi(struct nv_basic_block *b, struct nv_instruction *i)
 452 {
 453    if (!b->phi) {
 454       i->prev = NULL;
 455       b->phi = i;
 456       i->next = b->entry;
 457       if (b->entry) {
 458          assert(!b->entry->prev && b->exit);
 459          b->entry->prev = i;
 460       } else {
 461          b->entry = i;
 462          b->exit = i;
 463       }
 464    } else {
 465       assert(b->entry);
 466       if (b->entry->opcode == NV_OP_PHI) { /* insert after entry */
 467          assert(b->entry == b->exit);
 468          b->entry->next = i;
 469          i->prev = b->entry;
 470          b->entry = i;
 471          b->exit = i;
 472       } else { /* insert before entry */
 473          assert(b->entry->prev && b->exit);
 474          i->next = b->entry;
 475          i->prev = b->entry->prev;
 476          b->entry->prev = i;
 477          i->prev->next = i;
 478       }
 479    }
 480 }
 481
 482 void
 483 nvc0_insn_append(struct nv_basic_block *b, struct nv_instruction *i)
 484 {
 485    if (i->opcode == NV_OP_PHI) {
 486       nvbb_insert_phi(b, i);
 487    } else {
 488       i->prev = b->exit;
 489       if (b->exit)
 490          b->exit->next = i;
 491       b->exit = i;
 492       if (!b->entry)
 493          b->entry = i;
 494       else
 495       if (i->prev && i->prev->opcode == NV_OP_PHI)
 496          b->entry = i;
 497    }
 498
 499    i->bb = b;
 500    b->num_instructions++;
 501 }
 502
 503 void
 504 nvc0_insn_insert_after(struct nv_instruction *at, struct nv_instruction *ni)
 505 {
 506    if (!at->next) {
 507       nvc0_insn_append(at->bb, ni);
 508       return;
 509    }
 510    ni->next = at->next;
 511    ni->prev = at;
 512    ni->next->prev = ni;
 513    ni->prev->next = ni;
 514 }
 515
 516 void
 517 nvc0_insn_delete(struct nv_instruction *nvi)
 518 {
 519    struct nv_basic_block *b = nvi->bb;
 520    int s;
 521
 522    /* debug_printf("REM: "); nv_print_instruction(nvi); */
 523
 524    for (s = 0; s < 6 && nvi->src[s]; ++s)
 525       nv_reference(NULL, nvi, s, NULL);
 526
 527    if (nvi->next)
 528       nvi->next->prev = nvi->prev;
 529    else {
 530       assert(nvi == b->exit);
 531       b->exit = nvi->prev;
 532    }
 533
 534    if (nvi->prev)
 535       nvi->prev->next = nvi->next;
 536
 537    if (nvi == b->entry) {
 538       /* PHIs don't get hooked to b->entry */
 539       b->entry = nvi->next;
 540       assert(!nvi->prev || nvi->prev->opcode == NV_OP_PHI);
 541    }
 542
 543    if (nvi == b->phi) {
 544       if (nvi->opcode != NV_OP_PHI)
 545          NOUVEAU_DBG("NOTE: b->phi points to non-PHI instruction\n");
 546
 547       assert(!nvi->prev);
 548       if (!nvi->next || nvi->next->opcode != NV_OP_PHI)
 549          b->phi = NULL;
 550       else
 551          b->phi = nvi->next;
 552    }
 553 }
 554
 555 void
 556 nvc0_insns_permute(struct nv_instruction *i1, struct nv_instruction *i2)
 557 {
 558    struct nv_basic_block *b = i1->bb;
 559
 560    assert(i1->opcode != NV_OP_PHI &&
 561           i2->opcode != NV_OP_PHI);
 562    assert(i1->next == i2);
 563
 564    if (b->exit == i2)
 565       b->exit = i1;
 566
 567    if (b->entry == i1)
 568       b->entry = i2;
 569
 570    i2->prev = i1->prev;
 571    i1->next = i2->next;
 572    i2->next = i1;
 573    i1->prev = i2;
 574
 575    if (i2->prev)
 576       i2->prev->next = i2;
 577    if (i1->next)
 578       i1->next->prev = i1;
 579 }
 580
 581 void
 582 nvc0_bblock_attach(struct nv_basic_block *parent,
 583                    struct nv_basic_block *b, ubyte edge_kind)
 584 {
 585    assert(b->num_in < 8);
 586
 587    if (parent->out[0]) {
 588       assert(!parent->out[1]);
 589       parent->out[1] = b;
 590       parent->out_kind[1] = edge_kind;
 591    } else {
 592       parent->out[0] = b;
 593       parent->out_kind[0] = edge_kind;
 594    }
 595
 596    b->in[b->num_in] = parent;
 597    b->in_kind[b->num_in++] = edge_kind;
 598 }
 599
 600 /* NOTE: all BRKs are treated as conditional, so there are 2 outgoing BBs */
 601
 602 boolean
 603 nvc0_bblock_dominated_by(struct nv_basic_block *b, struct nv_basic_block *d)
 604 {
 605    int j;
 606
 607    if (b == d)
 608       return TRUE;
 609
 610    for (j = 0; j < b->num_in; ++j)
 611       if ((b->in_kind[j] != CFG_EDGE_BACK) &&
 612           !nvc0_bblock_dominated_by(b->in[j], d))
 613          return FALSE;
 614
 615    return j ? TRUE : FALSE;
 616 }
 617
 618 /* check if @bf (future) can be reached from @bp (past), stop at @bt */
 619 boolean
 620 nvc0_bblock_reachable_by(struct nv_basic_block *bf, struct nv_basic_block *bp,
 621                          struct nv_basic_block *bt)
 622 {
 623    struct nv_basic_block *q[NV_PC_MAX_BASIC_BLOCKS], *b;
 624    int i, p, n;
 625
 626    p = 0;
 627    n = 1;
 628    q[0] = bp;
 629
 630    while (p < n) {
 631       b = q[p++];
 632
 633       if (b == bf)
 634          break;
 635       if (b == bt)
 636          continue;
 637       assert(n <= (1024 - 2));
 638
 639       for (i = 0; i < 2; ++i) {
 640          if (b->out[i] && !IS_WALL_EDGE(b->out_kind[i]) && !b->out[i]->priv) {
 641             q[n] = b->out[i];
 642             q[n++]->priv = 1;
 643          }
 644       }
 645    }
 646    for (--n; n >= 0; --n)
 647       q[n]->priv = 0;
 648
 649    return (b == bf);
 650 }
 651
 652 static struct nv_basic_block *
 653 nvbb_find_dom_frontier(struct nv_basic_block *b, struct nv_basic_block *df)
 654 {
 655    struct nv_basic_block *out;
 656    int i;
 657
 658    if (!nvc0_bblock_dominated_by(df, b)) {
 659       for (i = 0; i < df->num_in; ++i) {
 660          if (df->in_kind[i] == CFG_EDGE_BACK)
 661             continue;
 662          if (nvc0_bblock_dominated_by(df->in[i], b))
 663             return df;
 664       }
 665    }
 666    for (i = 0; i < 2 && df->out[i]; ++i) {
 667       if (df->out_kind[i] == CFG_EDGE_BACK)
 668          continue;
 669       if ((out = nvbb_find_dom_frontier(b, df->out[i])))
 670          return out;
 671    }
 672    return NULL;
 673 }
 674
 675 struct nv_basic_block *
 676 nvc0_bblock_dom_frontier(struct nv_basic_block *b)
 677 {
 678    struct nv_basic_block *df;
 679    int i;
 680
 681    for (i = 0; i < 2 && b->out[i]; ++i)
 682       if ((df = nvbb_find_dom_frontier(b, b->out[i])))
 683          return df;
 684    return NULL;
 685 }