src/gallium/drivers/nv50/nv50_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 NV50PC_DEBUG */
  24
  25 #include "nv50_pc.h"
  26 #include "nv50_program.h"
  27
  28 #include <stdio.h>
  29
  30 /* returns TRUE if operands 0 and 1 can be swapped */
  31 boolean
  32 nv_op_commutative(uint opcode)
  33 {
  34    switch (opcode) {
  35    case NV_OP_ADD:
  36    case NV_OP_MUL:
  37    case NV_OP_MAD:
  38    case NV_OP_AND:
  39    case NV_OP_OR:
  40    case NV_OP_XOR:
  41    case NV_OP_MIN:
  42    case NV_OP_MAX:
  43    case NV_OP_SAD:
  44      return TRUE;
  45    default:
  46      return FALSE;
  47    }
  48 }
  49
  50 /* return operand to which the address register applies */
  51 int
  52 nv50_indirect_opnd(struct nv_instruction *i)
  53 {
  54    if (!i->src[4])
  55       return -1;
  56
  57    switch (i->opcode) {
  58    case NV_OP_MOV:
  59    case NV_OP_LDA:
  60    case NV_OP_STA:
  61       return 0;
  62    default:
  63       return 1;
  64    }
  65 }
  66
  67 boolean
  68 nv50_nvi_can_use_imm(struct nv_instruction *nvi, int s)
  69 {
  70    if (nvi->flags_src || nvi->flags_def)
  71       return FALSE;
  72
  73    switch (nvi->opcode) {
  74    case NV_OP_ADD:
  75    case NV_OP_MUL:
  76    case NV_OP_AND:
  77    case NV_OP_OR:
  78    case NV_OP_XOR:
  79    case NV_OP_SHL:
  80    case NV_OP_SHR:
  81       return (s == 1) && (nvi->src[0]->value->reg.file == NV_FILE_GPR) &&
  82          (nvi->def[0]->reg.file == NV_FILE_GPR);
  83    case NV_OP_MOV:
  84       assert(s == 0);
  85       return (nvi->def[0]->reg.file == NV_FILE_GPR);
  86    default:
  87       return FALSE;
  88    }
  89 }
  90
  91 boolean
  92 nv50_nvi_can_load(struct nv_instruction *nvi, int s, struct nv_value *value)
  93 {
  94    int i;
  95
  96    for (i = 0; i < 3 && nvi->src[i]; ++i)
  97       if (nvi->src[i]->value->reg.file == NV_FILE_IMM)
  98          return FALSE;
  99
 100    switch (nvi->opcode) {
 101    case NV_OP_ABS:
 102    case NV_OP_ADD:
 103    case NV_OP_CEIL:
 104    case NV_OP_FLOOR:
 105    case NV_OP_TRUNC:
 106    case NV_OP_CVT:
 107    case NV_OP_NEG:
 108    case NV_OP_MAD:
 109    case NV_OP_MUL:
 110    case NV_OP_SAT:
 111    case NV_OP_SUB:
 112    case NV_OP_MAX:
 113    case NV_OP_MIN:
 114       if (s == 0 && (value->reg.file == NV_FILE_MEM_S ||
 115                      value->reg.file == NV_FILE_MEM_P))
 116          return TRUE;
 117       if (value->reg.file < NV_FILE_MEM_C(0) ||
 118           value->reg.file > NV_FILE_MEM_C(15))
 119          return FALSE;
 120       return (s == 1) ||
 121          ((s == 2) && (nvi->src[1]->value->reg.file == NV_FILE_GPR));
 122    case NV_OP_MOV:
 123       assert(s == 0);
 124       return /* TRUE */ FALSE; /* don't turn MOVs into loads */
 125    default:
 126       return FALSE;
 127    }
 128 }
 129
 130 /* Return whether this instruction can be executed conditionally. */
 131 boolean
 132 nv50_nvi_can_predicate(struct nv_instruction *nvi)
 133 {
 134    int i;
 135
 136    if (nvi->flags_src)
 137       return FALSE;
 138    for (i = 0; i < 4 && nvi->src[i]; ++i)
 139       if (nvi->src[i]->value->reg.file == NV_FILE_IMM)
 140          return FALSE;
 141    return TRUE;
 142 }
 143
 144 ubyte
 145 nv50_supported_src_mods(uint opcode, int s)
 146 {
 147    switch (opcode) {
 148    case NV_OP_ABS:
 149       return NV_MOD_NEG | NV_MOD_ABS; /* obviously */
 150    case NV_OP_ADD:
 151    case NV_OP_MUL:
 152    case NV_OP_MAD:
 153       return NV_MOD_NEG;
 154    case NV_OP_DFDX:
 155    case NV_OP_DFDY:
 156       assert(s == 0);
 157       return NV_MOD_NEG;
 158    case NV_OP_MAX:
 159    case NV_OP_MIN:
 160       return NV_MOD_ABS;
 161    case NV_OP_CVT:
 162    case NV_OP_LG2:
 163    case NV_OP_NEG:
 164    case NV_OP_PREEX2:
 165    case NV_OP_PRESIN:
 166    case NV_OP_RCP:
 167    case NV_OP_RSQ:
 168       return NV_MOD_ABS | NV_MOD_NEG;
 169    default:
 170       return 0;
 171    }
 172 }
 173
 174 /* We may want an opcode table. */
 175 boolean
 176 nv50_op_can_write_flags(uint opcode)
 177 {
 178    if (nv_is_vector_op(opcode))
 179       return FALSE;
 180    switch (opcode) { /* obvious ones like KIL, CALL, etc. not included */
 181    case NV_OP_PHI:
 182    case NV_OP_MOV:
 183    case NV_OP_LINTERP:
 184    case NV_OP_PINTERP:
 185    case NV_OP_LDA:
 186       return FALSE;
 187    default:
 188       break;
 189    }
 190    if (opcode >= NV_OP_RCP && opcode <= NV_OP_PREEX2)
 191       return FALSE;
 192    return TRUE;
 193 }
 194
 195 int
 196 nv_nvi_refcount(struct nv_instruction *nvi)
 197 {
 198    int i, rc;
 199
 200    rc = nvi->flags_def ? nvi->flags_def->refc : 0;
 201
 202    for (i = 0; i < 4; ++i) {
 203       if (!nvi->def[i])
 204          return rc;
 205       rc += nvi->def[i]->refc;
 206    }
 207    return rc;
 208 }
 209
 210 int
 211 nvcg_replace_value(struct nv_pc *pc, struct nv_value *old_val,
 212                    struct nv_value *new_val)
 213 {
 214    int i, n;
 215
 216    if (old_val == new_val)
 217       return old_val->refc;
 218
 219    for (i = 0, n = 0; i < pc->num_refs; ++i) {
 220       if (pc->refs[i]->value == old_val) {
 221          ++n;
 222          nv_reference(pc, &pc->refs[i], new_val);
 223       }
 224    }
 225    return n;
 226 }
 227
 228 struct nv_value *
 229 nvcg_find_constant(struct nv_ref *ref)
 230 {
 231    struct nv_value *src;
 232
 233    if (!ref)
 234       return NULL;
 235
 236    src = ref->value;
 237    while (src->insn && src->insn->opcode == NV_OP_MOV) {
 238       assert(!src->insn->src[0]->mod);
 239       src = src->insn->src[0]->value;
 240    }
 241    if ((src->reg.file == NV_FILE_IMM) ||
 242        (src->insn && src->insn->opcode == NV_OP_LDA &&
 243         src->insn->src[0]->value->reg.file >= NV_FILE_MEM_C(0) &&
 244         src->insn->src[0]->value->reg.file <= NV_FILE_MEM_C(15)))
 245       return src;
 246    return NULL;
 247 }
 248
 249 struct nv_value *
 250 nvcg_find_immediate(struct nv_ref *ref)
 251 {
 252    struct nv_value *src = nvcg_find_constant(ref);
 253
 254    return (src && src->reg.file == NV_FILE_IMM) ? src : NULL;
 255 }
 256
 257 static void
 258 nv_pc_free_refs(struct nv_pc *pc)
 259 {
 260    int i;
 261    for (i = 0; i < pc->num_refs; i += 64)
 262       FREE(pc->refs[i]);
 263    FREE(pc->refs);
 264 }
 265
 266 static const char *
 267 edge_name(ubyte type)
 268 {
 269    switch (type) {
 270    case CFG_EDGE_FORWARD: return "forward";
 271    case CFG_EDGE_BACK: return "back";
 272    case CFG_EDGE_LOOP_ENTER: return "loop";
 273    case CFG_EDGE_LOOP_LEAVE: return "break";
 274    case CFG_EDGE_FAKE: return "fake";
 275    default:
 276       return "?";
 277    }
 278 }
 279
 280 void
 281 nv_pc_pass_in_order(struct nv_basic_block *root, nv_pc_pass_func f, void *priv)
 282 {
 283    struct nv_basic_block *bb[64], *bbb[16], *b;
 284    int j, p, pp;
 285
 286    bb[0] = root;
 287    p = 1;
 288    pp = 0;
 289
 290    while (p > 0) {
 291       b = bb[--p];
 292       b->priv = 0;
 293
 294       for (j = 1; j >= 0; --j) {
 295          if (!b->out[j])
 296             continue;
 297
 298          switch (b->out_kind[j]) {
 299          case CFG_EDGE_BACK:
 300             continue;
 301          case CFG_EDGE_FORWARD:
 302          case CFG_EDGE_FAKE:
 303             if (++b->out[j]->priv == b->out[j]->num_in)
 304                bb[p++] = b->out[j];
 305             break;
 306          case CFG_EDGE_LOOP_ENTER:
 307             bb[p++] = b->out[j];
 308             break;
 309          case CFG_EDGE_LOOP_LEAVE:
 310             bbb[pp++] = b->out[j];
 311             break;
 312          default:
 313             assert(0);
 314             break;
 315          }
 316       }
 317
 318       f(priv, b);
 319
 320       if (!p) {
 321          p = pp;
 322          for (; pp > 0; --pp)
 323             bb[pp - 1] = bbb[pp - 1];
 324       }
 325    }
 326 }
 327
 328 static void
 329 nv_do_print_function(void *priv, struct nv_basic_block *b)
 330 {
 331    struct nv_instruction *i = b->phi;
 332
 333    debug_printf("=== BB %i ", b->id);
 334    if (b->out[0])
 335       debug_printf("[%s -> %i] ", edge_name(b->out_kind[0]), b->out[0]->id);
 336    if (b->out[1])
 337       debug_printf("[%s -> %i] ", edge_name(b->out_kind[1]), b->out[1]->id);
 338    debug_printf("===\n");
 339
 340    i = b->phi;
 341    if (!i)
 342       i = b->entry;
 343    for (; i; i = i->next)
 344       nv_print_instruction(i);
 345 }
 346
 347 void
 348 nv_print_function(struct nv_basic_block *root)
 349 {
 350    if (root->subroutine)
 351       debug_printf("SUBROUTINE %i\n", root->subroutine);
 352    else
 353       debug_printf("MAIN\n");
 354
 355    nv_pc_pass_in_order(root, nv_do_print_function, root);
 356 }
 357
 358 void
 359 nv_print_program(struct nv_pc *pc)
 360 {
 361    int i;
 362    for (i = 0; i < pc->num_subroutines + 1; ++i)
 363       if (pc->root[i])
 364          nv_print_function(pc->root[i]);
 365 }
 366
 367 #ifdef NV50PC_DEBUG
 368 static void
 369 nv_do_print_cfgraph(struct nv_pc *pc, FILE *f, struct nv_basic_block *b)
 370 {
 371    int i;
 372
 373    b->pass_seq = pc->pass_seq;
 374
 375    fprintf(f, "\t%i [shape=box]\n", b->id);
 376
 377    for (i = 0; i < 2; ++i) {
 378       if (!b->out[i])
 379          continue;
 380       switch (b->out_kind[i]) {
 381       case CFG_EDGE_FORWARD:
 382          fprintf(f, "\t%i -> %i;\n", b->id, b->out[i]->id);
 383          break;
 384       case CFG_EDGE_LOOP_ENTER:
 385          fprintf(f, "\t%i -> %i [color=green];\n", b->id, b->out[i]->id);
 386          break;
 387       case CFG_EDGE_LOOP_LEAVE:
 388          fprintf(f, "\t%i -> %i [color=red];\n", b->id, b->out[i]->id);
 389          break;
 390       case CFG_EDGE_BACK:
 391          fprintf(f, "\t%i -> %i;\n", b->id, b->out[i]->id);
 392          continue;
 393       case CFG_EDGE_FAKE:
 394          fprintf(f, "\t%i -> %i [style=dotted];\n", b->id, b->out[i]->id);
 395          break;
 396       default:
 397          assert(0);
 398          break;
 399       }
 400       if (b->out[i]->pass_seq < pc->pass_seq)
 401          nv_do_print_cfgraph(pc, f, b->out[i]);
 402    }
 403 }
 404
 405 /* Print the control flow graph of subroutine @subr (0 == MAIN) to a file. */
 406 static void
 407 nv_print_cfgraph(struct nv_pc *pc, const char *filepath, int subr)
 408 {
 409    FILE *f;
 410
 411    f = fopen(filepath, "a");
 412    if (!f)
 413       return;
 414
 415    fprintf(f, "digraph G {\n");
 416
 417    ++pc->pass_seq;
 418
 419    nv_do_print_cfgraph(pc, f, pc->root[subr]);
 420
 421    fprintf(f, "}\n");
 422
 423    fclose(f);
 424 }
 425 #endif
 426
 427 static INLINE void
 428 nvcg_show_bincode(struct nv_pc *pc)
 429 {
 430    unsigned i;
 431
 432    for (i = 0; i < pc->bin_size / 4; ++i) {
 433       debug_printf("0x%08x ", pc->emit[i]);
 434       if ((i % 16) == 15)
 435          debug_printf("\n");
 436    }
 437    debug_printf("\n");
 438 }
 439
 440 static int
 441 nv50_emit_program(struct nv_pc *pc)
 442 {
 443    uint32_t *code = pc->emit;
 444    int n;
 445
 446    NV50_DBGMSG("emitting program: size = %u\n", pc->bin_size);
 447
 448    for (n = 0; n < pc->num_blocks; ++n) {
 449       struct nv_instruction *i;
 450       struct nv_basic_block *b = pc->bb_list[n];
 451
 452       for (i = b->entry; i; i = i->next) {
 453          nv50_emit_instruction(pc, i);
 454
 455          pc->bin_pos += 1 + (pc->emit[0] & 1);
 456          pc->emit += 1 + (pc->emit[0] & 1);
 457       }
 458    }
 459    assert(pc->emit == &code[pc->bin_size / 4]);
 460
 461    /* XXX: we can do better than this ... */
 462    if (!pc->bin_size ||
 463        !(pc->emit[-2] & 1) || (pc->emit[-2] & 2) || (pc->emit[-1] & 3)) {
 464       pc->emit[0] = 0xf0000001;
 465       pc->emit[1] = 0xe0000000;
 466       pc->bin_size += 8;
 467    }
 468
 469    pc->emit = code;
 470    code[pc->bin_size / 4 - 1] |= 1;
 471
 472 #ifdef NV50PC_DEBUG
 473    nvcg_show_bincode(pc);
 474 #endif
 475
 476    return 0;
 477 }
 478
 479 int
 480 nv50_generate_code(struct nv50_translation_info *ti)
 481 {
 482    struct nv_pc *pc;
 483    int ret;
 484    int i;
 485
 486    pc = CALLOC_STRUCT(nv_pc);
 487    if (!pc)
 488       return 1;
 489
 490    pc->root = CALLOC(ti->subr_nr + 1, sizeof(pc->root[0]));
 491    if (!pc->root) {
 492       FREE(pc);
 493       return 1;
 494    }
 495    pc->num_subroutines = ti->subr_nr;
 496
 497    ret = nv50_tgsi_to_nc(pc, ti);
 498    if (ret)
 499       goto out;
 500 #ifdef NV50PC_DEBUG
 501    nv_print_program(pc);
 502 #endif
 503
 504    pc->opt_reload_elim = ti->store_to_memory ? FALSE : TRUE;
 505
 506    /* optimization */
 507    ret = nv_pc_exec_pass0(pc);
 508    if (ret)
 509       goto out;
 510 #ifdef NV50PC_DEBUG
 511    nv_print_program(pc);
 512 #endif
 513
 514    /* register allocation */
 515    ret = nv_pc_exec_pass1(pc);
 516    if (ret)
 517       goto out;
 518 #ifdef NV50PC_DEBUG
 519    nv_print_program(pc);
 520    nv_print_cfgraph(pc, "nv50_shader_cfgraph.dot", 0);
 521 #endif
 522
 523    /* prepare for emission */
 524    ret = nv_pc_exec_pass2(pc);
 525    if (ret)
 526       goto out;
 527    assert(!(pc->bin_size % 8));
 528
 529    pc->emit = CALLOC(pc->bin_size / 4 + 2, 4);
 530    if (!pc->emit) {
 531       ret = 3;
 532       goto out;
 533    }
 534    ret = nv50_emit_program(pc);
 535    if (ret)
 536       goto out;
 537
 538    ti->p->code_size = pc->bin_size;
 539    ti->p->code = pc->emit;
 540
 541    ti->p->immd_size = pc->immd_count * 4;
 542    ti->p->immd = pc->immd_buf;
 543
 544    /* highest 16 bit reg to num of 32 bit regs, limit to >= 4 */
 545    ti->p->max_gpr = MAX2(4, (pc->max_reg[NV_FILE_GPR] >> 1) + 1);
 546
 547    ti->p->fixups = pc->fixups;
 548    ti->p->num_fixups = pc->num_fixups;
 549
 550    ti->p->uses_lmem = ti->store_to_memory;
 551
 552    NV50_DBGMSG("SHADER TRANSLATION - %s\n", ret ? "failure" : "success");
 553
 554 out:
 555    nv_pc_free_refs(pc);
 556
 557    for (i = 0; i < pc->num_blocks; ++i)
 558       FREE(pc->bb_list[i]);
 559    if (pc->root)
 560       FREE(pc->root);
 561    if (ret) { /* on success, these will be referenced by nv50_program */
 562       if (pc->emit)
 563          FREE(pc->emit);
 564       if (pc->immd_buf)
 565          FREE(pc->immd_buf);
 566       if (pc->fixups)
 567          FREE(pc->fixups);
 568    }
 569    FREE(pc);
 570    return ret;
 571 }
 572
 573 static void
 574 nvbb_insert_phi(struct nv_basic_block *b, struct nv_instruction *i)
 575 {
 576    if (!b->phi) {
 577       i->prev = NULL;
 578       b->phi = i;
 579       i->next = b->entry;
 580       if (b->entry) {
 581          assert(!b->entry->prev && b->exit);
 582          b->entry->prev = i;
 583       } else {
 584          b->entry = i;
 585          b->exit = i;
 586       }
 587    } else {
 588       assert(b->entry);
 589       if (b->entry->opcode == NV_OP_PHI) { /* insert after entry */
 590          assert(b->entry == b->exit);
 591          b->entry->next = i;
 592          i->prev = b->entry;
 593          b->entry = i;
 594          b->exit = i;
 595       } else { /* insert before entry */
 596          assert(b->entry->prev && b->exit);
 597          i->next = b->entry;
 598          i->prev = b->entry->prev;
 599          b->entry->prev = i;
 600          i->prev->next = i;
 601       }
 602    }
 603 }
 604
 605 void
 606 nvbb_insert_tail(struct nv_basic_block *b, struct nv_instruction *i)
 607 {
 608    if (i->opcode == NV_OP_PHI) {
 609       nvbb_insert_phi(b, i);
 610    } else {
 611       i->prev = b->exit;
 612       if (b->exit)
 613          b->exit->next = i;
 614       b->exit = i;
 615       if (!b->entry)
 616          b->entry = i;
 617       else
 618       if (i->prev && i->prev->opcode == NV_OP_PHI)
 619          b->entry = i;
 620    }
 621
 622    i->bb = b;
 623    b->num_instructions++;
 624 }
 625
 626 void
 627 nvi_insert_after(struct nv_instruction *at, struct nv_instruction *ni)
 628 {
 629    if (!at->next) {
 630       nvbb_insert_tail(at->bb, ni);
 631       return;
 632    }
 633    ni->next = at->next;
 634    ni->prev = at;
 635    ni->next->prev = ni;
 636    ni->prev->next = ni;
 637 }
 638
 639 void
 640 nv_nvi_delete(struct nv_instruction *nvi)
 641 {
 642    struct nv_basic_block *b = nvi->bb;
 643    int j;
 644
 645    /* debug_printf("REM: "); nv_print_instruction(nvi); */
 646
 647    for (j = 0; j < 5; ++j)
 648       nv_reference(NULL, &nvi->src[j], NULL);
 649    nv_reference(NULL, &nvi->flags_src, NULL);
 650
 651    if (nvi->next)
 652       nvi->next->prev = nvi->prev;
 653    else {
 654       assert(nvi == b->exit);
 655       b->exit = nvi->prev;
 656    }
 657
 658    if (nvi->prev)
 659       nvi->prev->next = nvi->next;
 660
 661    if (nvi == b->entry) {
 662       /* PHIs don't get hooked to b->entry */
 663       b->entry = nvi->next;
 664       assert(!nvi->prev || nvi->prev->opcode == NV_OP_PHI);
 665    }
 666
 667    if (nvi == b->phi) {
 668       if (nvi->opcode != NV_OP_PHI)
 669          NV50_DBGMSG("NOTE: b->phi points to non-PHI instruction\n");
 670
 671       assert(!nvi->prev);
 672       if (!nvi->next || nvi->next->opcode != NV_OP_PHI)
 673          b->phi = NULL;
 674       else
 675          b->phi = nvi->next;
 676    }
 677 }
 678
 679 void
 680 nv_nvi_permute(struct nv_instruction *i1, struct nv_instruction *i2)
 681 {
 682    struct nv_basic_block *b = i1->bb;
 683
 684    assert(i1->opcode != NV_OP_PHI &&
 685           i2->opcode != NV_OP_PHI);
 686    assert(i1->next == i2);
 687
 688    if (b->exit == i2)
 689       b->exit = i1;
 690
 691    if (b->entry == i1)
 692       b->entry = i2;
 693
 694    i2->prev = i1->prev;
 695    i1->next = i2->next;
 696    i2->next = i1;
 697    i1->prev = i2;
 698
 699    if (i2->prev)
 700       i2->prev->next = i2;
 701    if (i1->next)
 702       i1->next->prev = i1;
 703 }
 704
 705 void
 706 nvbb_attach_block(struct nv_basic_block *parent,
 707                   struct nv_basic_block *b, ubyte edge_kind)
 708 {
 709    assert(b->num_in < 8);
 710
 711    if (parent->out[0]) {
 712       assert(!parent->out[1]);
 713       parent->out[1] = b;
 714       parent->out_kind[1] = edge_kind;
 715    } else {
 716       parent->out[0] = b;
 717       parent->out_kind[0] = edge_kind;
 718    }
 719
 720    b->in[b->num_in] = parent;
 721    b->in_kind[b->num_in++] = edge_kind;
 722 }
 723
 724 /* NOTE: all BRKs are treated as conditional, so there are 2 outgoing BBs */
 725
 726 boolean
 727 nvbb_dominated_by(struct nv_basic_block *b, struct nv_basic_block *d)
 728 {
 729    int j;
 730
 731    if (b == d)
 732       return TRUE;
 733
 734    for (j = 0; j < b->num_in; ++j)
 735       if ((b->in_kind[j] != CFG_EDGE_BACK) && !nvbb_dominated_by(b->in[j], d))
 736          return FALSE;
 737
 738    return j ? TRUE : FALSE;
 739 }
 740
 741 /* check if @bf (future) can be reached from @bp (past), stop at @bt */
 742 boolean
 743 nvbb_reachable_by(struct nv_basic_block *bf, struct nv_basic_block *bp,
 744                   struct nv_basic_block *bt)
 745 {
 746    struct nv_basic_block *q[NV_PC_MAX_BASIC_BLOCKS], *b;
 747    int i, p, n;
 748
 749    p = 0;
 750    n = 1;
 751    q[0] = bp;
 752
 753    while (p < n) {
 754       b = q[p++];
 755
 756       if (b == bf)
 757          break;
 758       if (b == bt)
 759          continue;
 760       assert(n <= (1024 - 2));
 761
 762       for (i = 0; i < 2; ++i) {
 763          if (b->out[i] && !IS_WALL_EDGE(b->out_kind[i]) && !b->out[i]->priv) {
 764             q[n] = b->out[i];
 765             q[n++]->priv = 1;
 766          }
 767       }
 768    }
 769    for (--n; n >= 0; --n)
 770       q[n]->priv = 0;
 771
 772    return (b == bf);
 773 }
 774
 775 static struct nv_basic_block *
 776 nvbb_find_dom_frontier(struct nv_basic_block *b, struct nv_basic_block *df)
 777 {
 778    struct nv_basic_block *out;
 779    int i;
 780
 781    if (!nvbb_dominated_by(df, b)) {
 782       for (i = 0; i < df->num_in; ++i) {
 783          if (df->in_kind[i] == CFG_EDGE_BACK)
 784             continue;
 785          if (nvbb_dominated_by(df->in[i], b))
 786             return df;
 787       }
 788    }
 789    for (i = 0; i < 2 && df->out[i]; ++i) {
 790       if (df->out_kind[i] == CFG_EDGE_BACK)
 791          continue;
 792       if ((out = nvbb_find_dom_frontier(b, df->out[i])))
 793          return out;
 794    }
 795    return NULL;
 796 }
 797
 798 struct nv_basic_block *
 799 nvbb_dom_frontier(struct nv_basic_block *b)
 800 {
 801    struct nv_basic_block *df;
 802    int i;
 803
 804    for (i = 0; i < 2 && b->out[i]; ++i)
 805       if ((df = nvbb_find_dom_frontier(b, b->out[i])))
 806          return df;
 807    return NULL;
 808 }