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       return 0;
  61    default:
  62       return 1;
  63    }
  64 }
  65
  66 boolean
  67 nv50_nvi_can_use_imm(struct nv_instruction *nvi, int s)
  68 {
  69    if (nvi->flags_src || nvi->flags_def)
  70       return FALSE;
  71
  72    switch (nvi->opcode) {
  73    case NV_OP_ADD:
  74    case NV_OP_MUL:
  75    case NV_OP_AND:
  76    case NV_OP_OR:
  77    case NV_OP_XOR:
  78    case NV_OP_SHL:
  79    case NV_OP_SHR:
  80       return (s == 1) && (nvi->src[0]->value->reg.file == NV_FILE_GPR) &&
  81          (nvi->def[0]->reg.file == NV_FILE_GPR);
  82    case NV_OP_MOV:
  83       assert(s == 0);
  84       return (nvi->def[0]->reg.file == NV_FILE_GPR);
  85    default:
  86       return FALSE;
  87    }
  88 }
  89
  90 boolean
  91 nv50_nvi_can_load(struct nv_instruction *nvi, int s, struct nv_value *value)
  92 {
  93    int i;
  94
  95    for (i = 0; i < 3 && nvi->src[i]; ++i)
  96       if (nvi->src[i]->value->reg.file == NV_FILE_IMM)
  97          return FALSE;
  98
  99    switch (nvi->opcode) {
 100    case NV_OP_ABS:
 101    case NV_OP_ADD:
 102    case NV_OP_CEIL:
 103    case NV_OP_FLOOR:
 104    case NV_OP_TRUNC:
 105    case NV_OP_CVT:
 106    case NV_OP_MAD:
 107    case NV_OP_MUL:
 108    case NV_OP_SAT:
 109    case NV_OP_SUB:
 110    case NV_OP_MAX:
 111    case NV_OP_MIN:
 112       if (s == 0 && (value->reg.file == NV_FILE_MEM_S ||
 113                      value->reg.file == NV_FILE_MEM_P))
 114          return TRUE;
 115       if (s == 1 &&
 116           value->reg.file >= NV_FILE_MEM_C(0) &&
 117           value->reg.file <= NV_FILE_MEM_C(15))
 118          return TRUE;
 119       if (s == 2 && nvi->src[1]->value->reg.file == NV_FILE_GPR)
 120          return TRUE;
 121       return FALSE;
 122    case NV_OP_MOV:
 123       assert(s == 0);
 124       return TRUE;
 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 int
 175 nv_nvi_refcount(struct nv_instruction *nvi)
 176 {
 177    int i, rc;
 178
 179    rc = nvi->flags_def ? nvi->flags_def->refc : 0;
 180
 181    for (i = 0; i < 4; ++i) {
 182       if (!nvi->def[i])
 183          return rc;
 184       rc += nvi->def[i]->refc;
 185    }
 186    return rc;
 187 }
 188
 189 int
 190 nvcg_replace_value(struct nv_pc *pc, struct nv_value *old_val,
 191                    struct nv_value *new_val)
 192 {
 193    int i, n;
 194
 195    if (old_val == new_val)
 196       return old_val->refc;
 197
 198    for (i = 0, n = 0; i < pc->num_refs; ++i) {
 199       if (pc->refs[i]->value == old_val) {
 200          ++n;
 201          nv_reference(pc, &pc->refs[i], new_val);
 202       }
 203    }
 204    return n;
 205 }
 206
 207 static void
 208 nv_pc_free_refs(struct nv_pc *pc)
 209 {
 210    int i;
 211    for (i = 0; i < pc->num_refs; i += 64)
 212       FREE(pc->refs[i]);
 213 }
 214
 215 static const char *
 216 edge_name(ubyte type)
 217 {
 218    switch (type) {
 219    case CFG_EDGE_FORWARD: return "forward";
 220    case CFG_EDGE_BACK: return "back";
 221    case CFG_EDGE_LOOP_ENTER: return "loop";
 222    case CFG_EDGE_LOOP_LEAVE: return "break";
 223    case CFG_EDGE_FAKE: return "fake";
 224    default:
 225       return "?";
 226    }
 227 }
 228
 229 void
 230 nv_pc_pass_in_order(struct nv_basic_block *root, nv_pc_pass_func f, void *priv)
 231 {
 232    struct nv_basic_block *bb[64], *bbb[16], *b;
 233    int j, p, pp;
 234
 235    bb[0] = root;
 236    p = 1;
 237    pp = 0;
 238
 239    while (p > 0) {
 240       b = bb[--p];
 241       b->priv = 0;
 242
 243       for (j = 1; j >= 0; --j) {
 244          if (!b->out[j])
 245             continue;
 246
 247          switch (b->out_kind[j]) {
 248          case CFG_EDGE_BACK:
 249             continue;
 250          case CFG_EDGE_FORWARD:
 251          case CFG_EDGE_FAKE:
 252             if (++b->out[j]->priv == b->out[j]->num_in)
 253                bb[p++] = b->out[j];
 254             break;
 255          case CFG_EDGE_LOOP_ENTER:
 256             bb[p++] = b->out[j];
 257             break;
 258          case CFG_EDGE_LOOP_LEAVE:
 259             bbb[pp++] = b->out[j];
 260             break;
 261          default:
 262             assert(0);
 263             break;
 264          }
 265       }
 266
 267       f(priv, b);
 268
 269       if (!p) {
 270          p = pp;
 271          for (; pp > 0; --pp)
 272             bb[pp - 1] = bbb[pp - 1];
 273       }
 274    }
 275 }
 276
 277 static void
 278 nv_do_print_program(void *priv, struct nv_basic_block *b)
 279 {
 280    struct nv_instruction *i = b->phi;
 281
 282    debug_printf("=== BB %i ", b->id);
 283    if (b->out[0])
 284       debug_printf("[%s -> %i] ", edge_name(b->out_kind[0]), b->out[0]->id);
 285    if (b->out[1])
 286       debug_printf("[%s -> %i] ", edge_name(b->out_kind[1]), b->out[1]->id);
 287    debug_printf("===\n");
 288
 289    i = b->phi;
 290    if (!i)
 291       i = b->entry;
 292    for (; i; i = i->next)
 293       nv_print_instruction(i);
 294 }
 295
 296 void
 297 nv_print_program(struct nv_basic_block *root)
 298 {
 299    nv_pc_pass_in_order(root, nv_do_print_program, root);
 300
 301    debug_printf("END\n\n");
 302 }
 303
 304 static INLINE void
 305 nvcg_show_bincode(struct nv_pc *pc)
 306 {
 307    int i;
 308
 309    for (i = 0; i < pc->bin_size / 4; ++i)
 310       debug_printf("0x%08x ", pc->emit[i]);
 311    debug_printf("\n");
 312 }
 313
 314 static int
 315 nv50_emit_program(struct nv_pc *pc)
 316 {
 317    uint32_t *code = pc->emit;
 318    int n;
 319
 320    NV50_DBGMSG("emitting program: size = %u\n", pc->bin_size);
 321
 322    for (n = 0; n < pc->num_blocks; ++n) {
 323       struct nv_instruction *i;
 324       struct nv_basic_block *b = pc->bb_list[n];
 325
 326       for (i = b->entry; i; i = i->next) {
 327          nv50_emit_instruction(pc, i);
 328
 329          pc->bin_pos += 1 + (pc->emit[0] & 1);
 330          pc->emit += 1 + (pc->emit[0] & 1);
 331       }
 332    }
 333    assert(pc->emit == &code[pc->bin_size / 4]);
 334
 335    /* XXX: we can do better than this ... */
 336    if (!(pc->emit[-2] & 1) || (pc->emit[-2] & 2) || (pc->emit[-1] & 3)) {
 337       pc->emit[0] = 0xf0000001;
 338       pc->emit[1] = 0xe0000000;
 339       pc->bin_size += 8;
 340    }
 341
 342    pc->emit = code;
 343    code[pc->bin_size / 4 - 1] |= 1;
 344
 345 #ifdef NV50PC_DEBUG
 346    nvcg_show_bincode(pc);
 347 #endif
 348
 349    return 0;
 350 }
 351
 352 int
 353 nv50_generate_code(struct nv50_translation_info *ti)
 354 {
 355    struct nv_pc *pc;
 356    int ret;
 357
 358    pc = CALLOC_STRUCT(nv_pc);
 359    if (!pc)
 360       return 1;
 361
 362    ret = nv50_tgsi_to_nc(pc, ti);
 363    if (ret)
 364       goto out;
 365 #ifdef NV50PC_DEBUG
 366    nv_print_program(pc->root);
 367 #endif
 368
 369    /* optimization */
 370    ret = nv_pc_exec_pass0(pc);
 371    if (ret)
 372       goto out;
 373 #ifdef NV50PC_DEBUG
 374    nv_print_program(pc->root);
 375 #endif
 376
 377    /* register allocation */
 378    ret = nv_pc_exec_pass1(pc);
 379    if (ret)
 380       goto out;
 381 #ifdef NV50PC_DEBUG
 382    nv_print_program(pc->root);
 383 #endif
 384
 385    /* prepare for emission */
 386    ret = nv_pc_exec_pass2(pc);
 387    if (ret)
 388       goto out;
 389
 390    pc->emit = CALLOC(pc->bin_size / 4 + 2, 4);
 391    if (!pc->emit) {
 392       ret = 3;
 393       goto out;
 394    }
 395    ret = nv50_emit_program(pc);
 396    if (ret)
 397       goto out;
 398
 399    ti->p->code_size = pc->bin_size;
 400    ti->p->code = pc->emit;
 401
 402    ti->p->immd_size = pc->immd_count * 4;
 403    ti->p->immd = pc->immd_buf;
 404
 405    /* highest 16 bit reg to num of 32 bit regs */
 406    ti->p->max_gpr = (pc->max_reg[NV_FILE_GPR] >> 1) + 1;
 407
 408    ti->p->fixups = pc->fixups;
 409    ti->p->num_fixups = pc->num_fixups;
 410
 411    NV50_DBGMSG("SHADER TRANSLATION - %s\n", ret ? "failure" : "success");
 412
 413 out:
 414    nv_pc_free_refs(pc);
 415    if (ret) {
 416       if (pc->emit)
 417          free(pc->emit);
 418       if (pc->immd_buf)
 419          free(pc->immd_buf);
 420       if (pc->fixups)
 421          free(pc->fixups);
 422    }
 423    free(pc);
 424
 425    return ret;
 426 }
 427
 428 static void
 429 nvbb_insert_phi(struct nv_basic_block *b, struct nv_instruction *i)
 430 {
 431    if (!b->phi) {
 432       i->prev = NULL;
 433       b->phi = i;
 434       i->next = b->entry;
 435       if (b->entry) {
 436          assert(!b->entry->prev && b->exit);
 437          b->entry->prev = i;
 438       } else {
 439          b->entry = i;
 440          b->exit = i;
 441       }
 442    } else {
 443       assert(b->entry);
 444       if (b->entry->opcode == NV_OP_PHI) { /* insert after entry */
 445          assert(b->entry == b->exit);
 446          b->entry->next = i;
 447          i->prev = b->entry;
 448          b->entry = i;
 449          b->exit = i;
 450       } else { /* insert before entry */
 451          assert(b->entry->prev && b->exit);
 452          i->next = b->entry;
 453          i->prev = b->entry->prev;
 454          b->entry->prev = i;
 455          i->prev->next = i;
 456       }
 457    }
 458 }
 459
 460 void
 461 nvbb_insert_tail(struct nv_basic_block *b, struct nv_instruction *i)
 462 {
 463    if (i->opcode == NV_OP_PHI) {
 464       nvbb_insert_phi(b, i);
 465    } else {
 466       i->prev = b->exit;
 467       if (b->exit)
 468          b->exit->next = i;
 469       b->exit = i;
 470       if (!b->entry)
 471          b->entry = i;
 472       else
 473       if (i->prev && i->prev->opcode == NV_OP_PHI)
 474          b->entry = i;
 475    }
 476
 477    i->bb = b;
 478    b->num_instructions++;
 479 }
 480
 481 void
 482 nv_nvi_delete(struct nv_instruction *nvi)
 483 {
 484    struct nv_basic_block *b = nvi->bb;
 485    int j;
 486
 487    /* debug_printf("REM: "); nv_print_instruction(nvi); */
 488
 489    for (j = 0; j < 5; ++j)
 490       nv_reference(NULL, &nvi->src[j], NULL);
 491    nv_reference(NULL, &nvi->flags_src, NULL);
 492
 493    if (nvi->next)
 494       nvi->next->prev = nvi->prev;
 495    else {
 496       assert(nvi == b->exit);
 497       b->exit = nvi->prev;
 498    }
 499
 500    if (nvi->prev)
 501       nvi->prev->next = nvi->next;
 502
 503    if (nvi == b->entry) {
 504       /* PHIs don't get hooked to b->entry */
 505       b->entry = nvi->next;
 506       assert(!nvi->prev || nvi->prev->opcode == NV_OP_PHI);
 507    }
 508
 509    if (nvi == b->phi) {
 510       if (nvi->opcode != NV_OP_PHI)
 511          NV50_DBGMSG("NOTE: b->phi points to non-PHI instruction\n");
 512
 513       assert(!nvi->prev);
 514       if (!nvi->next || nvi->next->opcode != NV_OP_PHI)
 515          b->phi = NULL;
 516       else
 517          b->phi = nvi->next;
 518    }
 519 }
 520
 521 void
 522 nv_nvi_permute(struct nv_instruction *i1, struct nv_instruction *i2)
 523 {
 524    struct nv_basic_block *b = i1->bb;
 525
 526    assert(i1->opcode != NV_OP_PHI &&
 527           i2->opcode != NV_OP_PHI);
 528    assert(i1->next == i2);
 529
 530    if (b->exit == i2)
 531       b->exit = i1;
 532
 533    if (b->entry == i1)
 534       b->entry = i2;
 535
 536    i2->prev = i1->prev;
 537    i1->next = i2->next;
 538    i2->next = i1;
 539    i1->prev = i2;
 540
 541    if (i2->prev)
 542       i2->prev->next = i2;
 543    if (i1->next)
 544       i1->next->prev = i1;
 545 }
 546
 547 void
 548 nvbb_attach_block(struct nv_basic_block *parent,
 549                   struct nv_basic_block *b, ubyte edge_kind)
 550 {
 551    assert(b->num_in < 8);
 552
 553    if (parent->out[0]) {
 554       assert(!parent->out[1]);
 555       parent->out[1] = b;
 556       parent->out_kind[1] = edge_kind;
 557    } else {
 558       parent->out[0] = b;
 559       parent->out_kind[0] = edge_kind;
 560    }
 561
 562    b->in[b->num_in] = parent;
 563    b->in_kind[b->num_in++] = edge_kind;
 564 }
 565
 566 /* NOTE: all BRKs are treated as conditional, so there are 2 outgoing BBs */
 567
 568 boolean
 569 nvbb_dominated_by(struct nv_basic_block *b, struct nv_basic_block *d)
 570 {
 571    int j;
 572
 573    if (b == d)
 574       return TRUE;
 575
 576    for (j = 0; j < b->num_in; ++j)
 577       if ((b->in_kind[j] != CFG_EDGE_BACK) && !nvbb_dominated_by(b->in[j], d))
 578          return FALSE;
 579
 580    return j ? TRUE : FALSE;
 581 }
 582
 583 /* check if bf (future) can be reached from bp (past) */
 584 boolean
 585 nvbb_reachable_by(struct nv_basic_block *bf, struct nv_basic_block *bp,
 586                   struct nv_basic_block *bt)
 587 {
 588    if (bf == bp)
 589       return TRUE;
 590    if (bp == bt)
 591       return FALSE;
 592
 593    if (bp->out[0] && !IS_WALL_EDGE(bp->out_kind[0]) &&
 594        nvbb_reachable_by(bf, bp->out[0], bt))
 595       return TRUE;
 596    if (bp->out[1] && !IS_WALL_EDGE(bp->out_kind[1]) &&
 597        nvbb_reachable_by(bf, bp->out[1], bt))
 598       return TRUE;
 599    return FALSE;
 600 }
 601
 602 static struct nv_basic_block *
 603 nvbb_find_dom_frontier(struct nv_basic_block *b, struct nv_basic_block *df)
 604 {
 605    int i;
 606
 607    if (!nvbb_dominated_by(df, b)) {
 608       for (i = 0; i < df->num_in; ++i) {
 609          if (df->in_kind[i] == CFG_EDGE_BACK)
 610             continue;
 611          if (nvbb_dominated_by(df->in[i], b))
 612             return df;
 613       }
 614    }
 615    for (i = 0; i < 2 && b->out[i]; ++i) {
 616       if (b->out_kind[i] == CFG_EDGE_BACK)
 617          continue;
 618       if ((df = nvbb_find_dom_frontier(b, b->out[i])))
 619          return df;
 620    }
 621    return NULL;
 622 }
 623
 624 struct nv_basic_block *
 625 nvbb_dom_frontier(struct nv_basic_block *b)
 626 {
 627    struct nv_basic_block *df;
 628    int i;
 629
 630    for (i = 0; i < 2 && b->out[i]; ++i)
 631       if ((df = nvbb_find_dom_frontier(b, b->out[i])))
 632          return df;
 633    return NULL;
 634 }