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    if (nvi->predicate >= 0) /* already predicated */
  59       return FALSE;
  60    if (!nvc0_op_info_table[nvi->opcode].predicate &&
  61        !nvc0_op_info_table[nvi->opcode].pseudo)
  62       return FALSE;
  63    return TRUE;
  64 }
  65
  66 int
  67 nvc0_insn_refcount(struct nv_instruction *nvi)
  68 {
  69    int rc = 0;
  70    int i;
  71    for (i = 0; i < 5 && nvi->def[i]; ++i) {
  72       if (!nvi->def[i])
  73          return rc;
  74       rc += nvi->def[i]->refc;
  75    }
  76    return rc;
  77 }
  78
  79 int
  80 nvc0_pc_replace_value(struct nv_pc *pc,
  81                       struct nv_value *old_val,
  82                       struct nv_value *new_val)
  83 {
  84    int i, n, s;
  85
  86    if (old_val == new_val)
  87       return old_val->refc;
  88
  89    for (i = 0, n = 0; i < pc->num_refs; ++i) {
  90       if (pc->refs[i]->value == old_val) {
  91          ++n;
  92          for (s = 0; s < 6 && pc->refs[i]->insn->src[s]; ++s)
  93             if (pc->refs[i]->insn->src[s] == pc->refs[i])
  94                break;
  95          assert(s < 6);
  96          nv_reference(pc, pc->refs[i]->insn, s, new_val);
  97       }
  98    }
  99    return n;
 100 }
 101
 102 static INLINE boolean
 103 is_gpr63(struct nv_value *val)
 104 {
 105    return (val->reg.file == NV_FILE_GPR && val->reg.id == 63);
 106 }
 107
 108 struct nv_value *
 109 nvc0_pc_find_constant(struct nv_ref *ref)
 110 {
 111    struct nv_value *src;
 112
 113    if (!ref)
 114       return NULL;
 115
 116    src = ref->value;
 117    while (src->insn && src->insn->opcode == NV_OP_MOV) {
 118       assert(!src->insn->src[0]->mod);
 119       src = src->insn->src[0]->value;
 120    }
 121    if ((src->reg.file == NV_FILE_IMM) || is_gpr63(src) ||
 122        (src->insn &&
 123         src->insn->opcode == NV_OP_LD &&
 124         src->insn->src[0]->value->reg.file >= NV_FILE_MEM_C(0) &&
 125         src->insn->src[0]->value->reg.file <= NV_FILE_MEM_C(15)))
 126       return src;
 127    return NULL;
 128 }
 129
 130 struct nv_value *
 131 nvc0_pc_find_immediate(struct nv_ref *ref)
 132 {
 133    struct nv_value *src = nvc0_pc_find_constant(ref);
 134
 135    return (src && (src->reg.file == NV_FILE_IMM || is_gpr63(src))) ? src : NULL;
 136 }
 137
 138 static void
 139 nv_pc_free_refs(struct nv_pc *pc)
 140 {
 141    int i;
 142    for (i = 0; i < pc->num_refs; i += 64)
 143       FREE(pc->refs[i]);
 144    FREE(pc->refs);
 145 }
 146
 147 static const char *
 148 edge_name(ubyte type)
 149 {
 150    switch (type) {
 151    case CFG_EDGE_FORWARD: return "forward";
 152    case CFG_EDGE_BACK: return "back";
 153    case CFG_EDGE_LOOP_ENTER: return "loop";
 154    case CFG_EDGE_LOOP_LEAVE: return "break";
 155    case CFG_EDGE_FAKE: return "fake";
 156    default:
 157       return "?";
 158    }
 159 }
 160
 161 void
 162 nvc0_pc_pass_in_order(struct nv_basic_block *root, nv_pc_pass_func f,
 163                       void *priv)
 164 {
 165    struct nv_basic_block *bb[64], *bbb[16], *b;
 166    int j, p, pp;
 167
 168    bb[0] = root;
 169    p = 1;
 170    pp = 0;
 171
 172    while (p > 0) {
 173       b = bb[--p];
 174       b->priv = 0;
 175
 176       for (j = 1; j >= 0; --j) {
 177          if (!b->out[j])
 178             continue;
 179
 180          switch (b->out_kind[j]) {
 181          case CFG_EDGE_BACK:
 182             continue;
 183          case CFG_EDGE_FORWARD:
 184          case CFG_EDGE_FAKE:
 185             if (++b->out[j]->priv == b->out[j]->num_in)
 186                bb[p++] = b->out[j];
 187             break;
 188          case CFG_EDGE_LOOP_ENTER:
 189             bb[p++] = b->out[j];
 190             break;
 191          case CFG_EDGE_LOOP_LEAVE:
 192             bbb[pp++] = b->out[j];
 193             break;
 194          default:
 195             assert(0);
 196             break;
 197          }
 198       }
 199
 200       f(priv, b);
 201
 202       if (!p) {
 203          p = pp;
 204          for (; pp > 0; --pp)
 205             bb[pp - 1] = bbb[pp - 1];
 206       }
 207    }
 208 }
 209
 210 static void
 211 nv_do_print_function(void *priv, struct nv_basic_block *b)
 212 {
 213    struct nv_instruction *i;
 214
 215    debug_printf("=== BB %i ", b->id);
 216    if (b->out[0])
 217       debug_printf("[%s -> %i] ", edge_name(b->out_kind[0]), b->out[0]->id);
 218    if (b->out[1])
 219       debug_printf("[%s -> %i] ", edge_name(b->out_kind[1]), b->out[1]->id);
 220    debug_printf("===\n");
 221
 222    i = b->phi;
 223    if (!i)
 224       i = b->entry;
 225    for (; i; i = i->next)
 226       nvc0_print_instruction(i);
 227 }
 228
 229 void
 230 nvc0_print_function(struct nv_basic_block *root)
 231 {
 232    if (root->subroutine)
 233       debug_printf("SUBROUTINE %i\n", root->subroutine);
 234    else
 235       debug_printf("MAIN\n");
 236
 237    nvc0_pc_pass_in_order(root, nv_do_print_function, root);
 238 }
 239
 240 void
 241 nvc0_print_program(struct nv_pc *pc)
 242 {
 243    int i;
 244    for (i = 0; i < pc->num_subroutines + 1; ++i)
 245       if (pc->root[i])
 246          nvc0_print_function(pc->root[i]);
 247 }
 248
 249 #if NOUVEAU_DEBUG > 1
 250 static void
 251 nv_do_print_cfgraph(struct nv_pc *pc, FILE *f, struct nv_basic_block *b)
 252 {
 253    int i;
 254
 255    b->pass_seq = pc->pass_seq;
 256
 257    fprintf(f, "\t%i [shape=box]\n", b->id);
 258
 259    for (i = 0; i < 2; ++i) {
 260       if (!b->out[i])
 261          continue;
 262       switch (b->out_kind[i]) {
 263       case CFG_EDGE_FORWARD:
 264          fprintf(f, "\t%i -> %i;\n", b->id, b->out[i]->id);
 265          break;
 266       case CFG_EDGE_LOOP_ENTER:
 267          fprintf(f, "\t%i -> %i [color=green];\n", b->id, b->out[i]->id);
 268          break;
 269       case CFG_EDGE_LOOP_LEAVE:
 270          fprintf(f, "\t%i -> %i [color=red];\n", b->id, b->out[i]->id);
 271          break;
 272       case CFG_EDGE_BACK:
 273          fprintf(f, "\t%i -> %i;\n", b->id, b->out[i]->id);
 274          continue;
 275       case CFG_EDGE_FAKE:
 276          fprintf(f, "\t%i -> %i [style=dotted];\n", b->id, b->out[i]->id);
 277          break;
 278       default:
 279          assert(0);
 280          break;
 281       }
 282       if (b->out[i]->pass_seq < pc->pass_seq)
 283          nv_do_print_cfgraph(pc, f, b->out[i]);
 284    }
 285 }
 286
 287 /* Print the control flow graph of subroutine @subr (0 == MAIN) to a file. */
 288 static void
 289 nv_print_cfgraph(struct nv_pc *pc, const char *filepath, int subr)
 290 {
 291    FILE *f;
 292
 293    f = fopen(filepath, "a");
 294    if (!f)
 295       return;
 296
 297    fprintf(f, "digraph G {\n");
 298
 299    ++pc->pass_seq;
 300
 301    nv_do_print_cfgraph(pc, f, pc->root[subr]);
 302
 303    fprintf(f, "}\n");
 304
 305    fclose(f);
 306 }
 307 #endif
 308
 309 static INLINE void
 310 nvc0_pc_print_binary(struct nv_pc *pc)
 311 {
 312    unsigned i;
 313
 314    NOUVEAU_DBG("nvc0_pc_print_binary(%u ops)\n", pc->emit_size / 8);
 315
 316    for (i = 0; i < pc->emit_size / 4; i += 2) {
 317       debug_printf("0x%08x ", pc->emit[i + 0]);
 318       debug_printf("0x%08x ", pc->emit[i + 1]);
 319       if ((i % 16) == 15)
 320          debug_printf("\n");
 321    }
 322    debug_printf("\n");
 323 }
 324
 325 static int
 326 nvc0_emit_program(struct nv_pc *pc)
 327 {
 328    uint32_t *code = pc->emit;
 329    int n;
 330
 331    NOUVEAU_DBG("emitting program: size = %u\n", pc->emit_size);
 332
 333    pc->emit_pos = 0;
 334    for (n = 0; n < pc->num_blocks; ++n) {
 335       struct nv_instruction *i;
 336       struct nv_basic_block *b = pc->bb_list[n];
 337
 338       for (i = b->entry; i; i = i->next) {
 339          nvc0_emit_instruction(pc, i);
 340          pc->emit += 2;
 341          pc->emit_pos += 8;
 342       }
 343    }
 344    assert(pc->emit == &code[pc->emit_size / 4]);
 345
 346    pc->emit[0] = 0x00001de7;
 347    pc->emit[1] = 0x80000000;
 348    pc->emit_size += 8;
 349
 350    pc->emit = code;
 351
 352 #ifdef NOUVEAU_DEBUG
 353    nvc0_pc_print_binary(pc);
 354 #else
 355    debug_printf("not printing binary\n");
 356 #endif
 357    return 0;
 358 }
 359
 360 int
 361 nvc0_generate_code(struct nvc0_translation_info *ti)
 362 {
 363    struct nv_pc *pc;
 364    int ret;
 365    int i;
 366
 367    pc = CALLOC_STRUCT(nv_pc);
 368    if (!pc)
 369       return 1;
 370
 371    pc->is_fragprog = ti->prog->type == PIPE_SHADER_FRAGMENT;
 372
 373    pc->root = CALLOC(ti->num_subrs + 1, sizeof(pc->root[0]));
 374    if (!pc->root) {
 375       FREE(pc);
 376       return 1;
 377    }
 378    pc->num_subroutines = ti->num_subrs;
 379
 380    ret = nvc0_tgsi_to_nc(pc, ti);
 381    if (ret)
 382       goto out;
 383 #if NOUVEAU_DEBUG > 1
 384    nvc0_print_program(pc);
 385 #endif
 386
 387    pc->opt_reload_elim = ti->require_stores ? FALSE : TRUE;
 388
 389    /* optimization */
 390    ret = nvc0_pc_exec_pass0(pc);
 391    if (ret)
 392       goto out;
 393 #ifdef NOUVEAU_DEBUG
 394    nvc0_print_program(pc);
 395 #endif
 396
 397    /* register allocation */
 398    ret = nvc0_pc_exec_pass1(pc);
 399    if (ret)
 400       goto out;
 401 #if NOUVEAU_DEBUG > 1
 402    nvc0_print_program(pc);
 403    nv_print_cfgraph(pc, "nvc0_shader_cfgraph.dot", 0);
 404 #endif
 405
 406    /* prepare for emission */
 407    ret = nvc0_pc_exec_pass2(pc);
 408    if (ret)
 409       goto out;
 410    assert(!(pc->emit_size % 8));
 411
 412    pc->emit = CALLOC(pc->emit_size / 4 + 2, 4);
 413    if (!pc->emit) {
 414       ret = 3;
 415       goto out;
 416    }
 417    ret = nvc0_emit_program(pc);
 418    if (ret)
 419       goto out;
 420
 421    ti->prog->code = pc->emit;
 422    ti->prog->code_base = 0;
 423    ti->prog->code_size = pc->emit_size;
 424    ti->prog->parm_size = 0;
 425
 426    ti->prog->max_gpr = MAX2(4, pc->max_reg[NV_FILE_GPR] + 1);
 427
 428    ti->prog->relocs = pc->reloc_entries;
 429    ti->prog->num_relocs = pc->num_relocs;
 430
 431    NOUVEAU_DBG("SHADER TRANSLATION - %s\n", ret ? "failure" : "success");
 432
 433 out:
 434    nv_pc_free_refs(pc);
 435
 436    for (i = 0; i < pc->num_blocks; ++i)
 437       FREE(pc->bb_list[i]);
 438    if (pc->root)
 439       FREE(pc->root);
 440    if (ret) {
 441       /* on success, these will be referenced by struct nvc0_program */
 442       if (pc->emit)
 443          FREE(pc->emit);
 444       if (pc->immd_buf)
 445          FREE(pc->immd_buf);
 446       if (pc->reloc_entries)
 447          FREE(pc->reloc_entries);
 448    }
 449    FREE(pc);
 450    return ret;
 451 }
 452
 453 static void
 454 nvbb_insert_phi(struct nv_basic_block *b, struct nv_instruction *i)
 455 {
 456    if (!b->phi) {
 457       i->prev = NULL;
 458       b->phi = i;
 459       i->next = b->entry;
 460       if (b->entry) {
 461          assert(!b->entry->prev && b->exit);
 462          b->entry->prev = i;
 463       } else {
 464          b->entry = i;
 465          b->exit = i;
 466       }
 467    } else {
 468       assert(b->entry);
 469       if (b->entry->opcode == NV_OP_PHI) { /* insert after entry */
 470          assert(b->entry == b->exit);
 471          b->entry->next = i;
 472          i->prev = b->entry;
 473          b->entry = i;
 474          b->exit = i;
 475       } else { /* insert before entry */
 476          assert(b->entry->prev && b->exit);
 477          i->next = b->entry;
 478          i->prev = b->entry->prev;
 479          b->entry->prev = i;
 480          i->prev->next = i;
 481       }
 482    }
 483 }
 484
 485 void
 486 nvc0_insn_append(struct nv_basic_block *b, struct nv_instruction *i)
 487 {
 488    if (i->opcode == NV_OP_PHI) {
 489       nvbb_insert_phi(b, i);
 490    } else {
 491       i->prev = b->exit;
 492       if (b->exit)
 493          b->exit->next = i;
 494       b->exit = i;
 495       if (!b->entry)
 496          b->entry = i;
 497       else
 498       if (i->prev && i->prev->opcode == NV_OP_PHI)
 499          b->entry = i;
 500    }
 501
 502    i->bb = b;
 503    b->num_instructions++;
 504
 505    if (i->prev && i->prev->terminator)
 506       nvc0_insns_permute(i->prev, i);
 507 }
 508
 509 void
 510 nvc0_insn_insert_after(struct nv_instruction *at, struct nv_instruction *ni)
 511 {
 512    if (!at->next) {
 513       nvc0_insn_append(at->bb, ni);
 514       return;
 515    }
 516    ni->next = at->next;
 517    ni->prev = at;
 518    ni->next->prev = ni;
 519    ni->prev->next = ni;
 520    ni->bb = at->bb;
 521    ni->bb->num_instructions++;
 522 }
 523
 524 void
 525 nvc0_insn_insert_before(struct nv_instruction *at, struct nv_instruction *ni)
 526 {
 527    nvc0_insn_insert_after(at, ni);
 528    nvc0_insns_permute(at, ni);
 529 }
 530
 531 void
 532 nvc0_insn_delete(struct nv_instruction *nvi)
 533 {
 534    struct nv_basic_block *b = nvi->bb;
 535    int s;
 536
 537    /* debug_printf("REM: "); nv_print_instruction(nvi); */
 538
 539    for (s = 0; s < 6 && nvi->src[s]; ++s)
 540       nv_reference(NULL, nvi, s, NULL);
 541
 542    if (nvi->next)
 543       nvi->next->prev = nvi->prev;
 544    else {
 545       assert(nvi == b->exit);
 546       b->exit = nvi->prev;
 547    }
 548
 549    if (nvi->prev)
 550       nvi->prev->next = nvi->next;
 551
 552    if (nvi == b->entry) {
 553       /* PHIs don't get hooked to b->entry */
 554       b->entry = nvi->next;
 555       assert(!nvi->prev || nvi->prev->opcode == NV_OP_PHI);
 556    }
 557
 558    if (nvi == b->phi) {
 559       if (nvi->opcode != NV_OP_PHI)
 560          NOUVEAU_DBG("NOTE: b->phi points to non-PHI instruction\n");
 561
 562       assert(!nvi->prev);
 563       if (!nvi->next || nvi->next->opcode != NV_OP_PHI)
 564          b->phi = NULL;
 565       else
 566          b->phi = nvi->next;
 567    }
 568 }
 569
 570 void
 571 nvc0_insns_permute(struct nv_instruction *i1, struct nv_instruction *i2)
 572 {
 573    struct nv_basic_block *b = i1->bb;
 574
 575    assert(i1->opcode != NV_OP_PHI &&
 576           i2->opcode != NV_OP_PHI);
 577    assert(i1->next == i2);
 578
 579    if (b->exit == i2)
 580       b->exit = i1;
 581
 582    if (b->entry == i1)
 583       b->entry = i2;
 584
 585    i2->prev = i1->prev;
 586    i1->next = i2->next;
 587    i2->next = i1;
 588    i1->prev = i2;
 589
 590    if (i2->prev)
 591       i2->prev->next = i2;
 592    if (i1->next)
 593       i1->next->prev = i1;
 594 }
 595
 596 void
 597 nvc0_bblock_attach(struct nv_basic_block *parent,
 598                    struct nv_basic_block *b, ubyte edge_kind)
 599 {
 600    assert(b->num_in < 8);
 601
 602    if (parent->out[0]) {
 603       assert(!parent->out[1]);
 604       parent->out[1] = b;
 605       parent->out_kind[1] = edge_kind;
 606    } else {
 607       parent->out[0] = b;
 608       parent->out_kind[0] = edge_kind;
 609    }
 610
 611    b->in[b->num_in] = parent;
 612    b->in_kind[b->num_in++] = edge_kind;
 613 }
 614
 615 /* NOTE: all BRKs are treated as conditional, so there are 2 outgoing BBs */
 616
 617 boolean
 618 nvc0_bblock_dominated_by(struct nv_basic_block *b, struct nv_basic_block *d)
 619 {
 620    int j;
 621
 622    if (b == d)
 623       return TRUE;
 624
 625    for (j = 0; j < b->num_in; ++j)
 626       if ((b->in_kind[j] != CFG_EDGE_BACK) &&
 627           !nvc0_bblock_dominated_by(b->in[j], d))
 628          return FALSE;
 629
 630    return j ? TRUE : FALSE;
 631 }
 632
 633 /* check if @bf (future) can be reached from @bp (past), stop at @bt */
 634 boolean
 635 nvc0_bblock_reachable_by(struct nv_basic_block *bf, struct nv_basic_block *bp,
 636                          struct nv_basic_block *bt)
 637 {
 638    struct nv_basic_block *q[NV_PC_MAX_BASIC_BLOCKS], *b;
 639    int i, p, n;
 640
 641    p = 0;
 642    n = 1;
 643    q[0] = bp;
 644
 645    while (p < n) {
 646       b = q[p++];
 647
 648       if (b == bf)
 649          break;
 650       if (b == bt)
 651          continue;
 652       assert(n <= (1024 - 2));
 653
 654       for (i = 0; i < 2; ++i) {
 655          if (b->out[i] && !IS_WALL_EDGE(b->out_kind[i]) && !b->out[i]->priv) {
 656             q[n] = b->out[i];
 657             q[n++]->priv = 1;
 658          }
 659       }
 660    }
 661    for (--n; n >= 0; --n)
 662       q[n]->priv = 0;
 663
 664    return (b == bf);
 665 }
 666
 667 static struct nv_basic_block *
 668 nvbb_find_dom_frontier(struct nv_basic_block *b, struct nv_basic_block *df)
 669 {
 670    struct nv_basic_block *out;
 671    int i;
 672
 673    if (!nvc0_bblock_dominated_by(df, b)) {
 674       for (i = 0; i < df->num_in; ++i) {
 675          if (df->in_kind[i] == CFG_EDGE_BACK)
 676             continue;
 677          if (nvc0_bblock_dominated_by(df->in[i], b))
 678             return df;
 679       }
 680    }
 681    for (i = 0; i < 2 && df->out[i]; ++i) {
 682       if (df->out_kind[i] == CFG_EDGE_BACK)
 683          continue;
 684       if ((out = nvbb_find_dom_frontier(b, df->out[i])))
 685          return out;
 686    }
 687    return NULL;
 688 }
 689
 690 struct nv_basic_block *
 691 nvc0_bblock_dom_frontier(struct nv_basic_block *b)
 692 {
 693    struct nv_basic_block *df;
 694    int i;
 695
 696    for (i = 0; i < 2 && b->out[i]; ++i)
 697       if ((df = nvbb_find_dom_frontier(b, b->out[i])))
 698          return df;
 699    return NULL;
 700 }