src/gallium/drivers/r600/sb/sb_bc_parser.cpp

   1 /*
   2  * Copyright 2013 Vadim Girlin <vadimgirlin@gmail.com>
   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  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *      Vadim Girlin
  25  */
  26
  27 #define BCP_DEBUG 0
  28
  29 #if BCP_DEBUG
  30 #define BCP_DUMP(q) do { q } while (0)
  31 #else
  32 #define BCP_DUMP(q)
  33 #endif
  34
  35 extern "C" {
  36 #include "r600_pipe.h"
  37 #include "r600_shader.h"
  38 }
  39
  40 #include <stack>
  41
  42 #include "sb_bc.h"
  43 #include "sb_shader.h"
  44 #include "sb_pass.h"
  45
  46 namespace r600_sb {
  47
  48 int bc_parser::decode() {
  49
  50         dw = bc->bytecode;
  51         bc_ndw = bc->ndw;
  52         max_cf = 0;
  53
  54         dec = new bc_decoder(ctx, dw, bc_ndw);
  55
  56         shader_target t = TARGET_UNKNOWN;
  57
  58         if (pshader) {
  59                 switch (bc->type) {
  60                 case TGSI_PROCESSOR_FRAGMENT: t = TARGET_PS; break;
  61                 case TGSI_PROCESSOR_VERTEX: t = TARGET_VS; break;
  62                 case TGSI_PROCESSOR_COMPUTE: t = TARGET_COMPUTE; break;
  63                 default: assert(!"unknown shader target"); return -1; break;
  64                 }
  65         } else {
  66                 if (bc->type == TGSI_PROCESSOR_COMPUTE)
  67                         t = TARGET_COMPUTE;
  68                 else
  69                         t = TARGET_FETCH;
  70         }
  71
  72         sh = new shader(ctx, t, bc->debug_id);
  73         int r = decode_shader();
  74
  75         delete dec;
  76
  77         sh->ngpr = bc->ngpr;
  78         sh->nstack = bc->nstack;
  79
  80         return r;
  81 }
  82
  83 int bc_parser::decode_shader() {
  84         int r = 0;
  85         unsigned i = 0;
  86         bool eop = false;
  87
  88         sh->init();
  89
  90         do {
  91                 eop = false;
  92                 if ((r = decode_cf(i, eop)))
  93                         return r;
  94
  95         } while (!eop || (i >> 1) <= max_cf);
  96
  97         return 0;
  98 }
  99
 100 int bc_parser::prepare() {
 101         int r = 0;
 102         if ((r = parse_decls()))
 103                 return r;
 104         if ((r = prepare_ir()))
 105                 return r;
 106         return 0;
 107 }
 108
 109 int bc_parser::parse_decls() {
 110
 111         if (!pshader) {
 112                 if (gpr_reladdr)
 113                         sh->add_gpr_array(0, bc->ngpr, 0x0F);
 114
 115                 // compute shaders have some values preloaded in R0, R1
 116                 sh->add_input(0 /* GPR */, true /* preloaded */, 0x0F /* mask */);
 117                 sh->add_input(1 /* GPR */, true /* preloaded */, 0x0F /* mask */);
 118                 return 0;
 119         }
 120
 121         if (pshader->indirect_files & ~(1 << TGSI_FILE_CONSTANT)) {
 122
 123                 assert(pshader->num_arrays);
 124
 125                 if (pshader->num_arrays) {
 126                         for (unsigned i = 0; i < pshader->num_arrays; ++i) {
 127                                 r600_shader_array &a = pshader->arrays[i];
 128                                 sh->add_gpr_array(a.gpr_start, a.gpr_count, a.comp_mask);
 129                         }
 130                 } else {
 131                         sh->add_gpr_array(0, pshader->bc.ngpr, 0x0F);
 132                 }
 133         }
 134
 135         if (sh->target == TARGET_VS)
 136                 sh->add_input(0, 1, 0x0F);
 137
 138         bool ps_interp = ctx.hw_class >= HW_CLASS_EVERGREEN
 139                         && sh->target == TARGET_PS;
 140
 141         unsigned linear = 0, persp = 0, centroid = 1;
 142
 143         for (unsigned i = 0; i < pshader->ninput; ++i) {
 144                 r600_shader_io & in = pshader->input[i];
 145                 bool preloaded = sh->target == TARGET_PS && !(ps_interp && in.spi_sid);
 146                 sh->add_input(in.gpr, preloaded, /*in.write_mask*/ 0x0F);
 147                 if (ps_interp && in.spi_sid) {
 148                         if (in.interpolate == TGSI_INTERPOLATE_LINEAR ||
 149                                         in.interpolate == TGSI_INTERPOLATE_COLOR)
 150                                 linear = 1;
 151                         else if (in.interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
 152                                 persp = 1;
 153                         if (in.centroid)
 154                                 centroid = 2;
 155                 }
 156         }
 157
 158         if (ps_interp) {
 159                 unsigned mask = (1 << (2 * (linear + persp) * centroid)) - 1;
 160                 unsigned gpr = 0;
 161
 162                 while (mask) {
 163                         sh->add_input(gpr, true, mask & 0x0F);
 164                         ++gpr;
 165                         mask >>= 4;
 166                 }
 167         }
 168
 169         return 0;
 170 }
 171
 172 int bc_parser::decode_cf(unsigned &i, bool &eop) {
 173
 174         int r;
 175
 176         cf_node *cf = sh->create_cf();
 177         sh->root->push_back(cf);
 178
 179         unsigned id = i >> 1;
 180
 181         cf->bc.id = id;
 182
 183         if (cf_map.size() < id + 1)
 184                 cf_map.resize(id + 1);
 185
 186         cf_map[id] = cf;
 187
 188         if ((r = dec->decode_cf(i, cf->bc)))
 189                 return r;
 190
 191         cf_op_flags flags = (cf_op_flags)cf->bc.op_ptr->flags;
 192
 193         if (flags & CF_ALU) {
 194                 if ((r = decode_alu_clause(cf)))
 195                         return r;
 196         } else if (flags & CF_FETCH) {
 197                 if ((r = decode_fetch_clause(cf)))
 198                         return r;;
 199         } else if (flags & CF_EXP) {
 200                 if (cf->bc.rw_rel)
 201                         gpr_reladdr = true;
 202                 assert(!cf->bc.rw_rel);
 203         } else if (flags & (CF_STRM | CF_RAT)) {
 204                 if (cf->bc.rw_rel)
 205                         gpr_reladdr = true;
 206                 assert(!cf->bc.rw_rel);
 207         } else if (flags & CF_BRANCH) {
 208                 if (cf->bc.addr > max_cf)
 209                         max_cf = cf->bc.addr;
 210         }
 211
 212         eop = cf->bc.end_of_program || cf->bc.op == CF_OP_CF_END ||
 213                         cf->bc.op == CF_OP_RET;
 214         return 0;
 215 }
 216
 217 int bc_parser::decode_alu_clause(cf_node* cf) {
 218         unsigned i = cf->bc.addr << 1, cnt = cf->bc.count + 1, gcnt;
 219
 220         cf->subtype = NST_ALU_CLAUSE;
 221
 222         cgroup = 0;
 223         memset(slots[0], 0, 5*sizeof(slots[0][0]));
 224
 225         unsigned ng = 0;
 226
 227         do {
 228                 decode_alu_group(cf, i, gcnt);
 229                 assert(gcnt <= cnt);
 230                 cnt -= gcnt;
 231                 ng++;
 232         } while (cnt);
 233
 234         return 0;
 235 }
 236
 237 int bc_parser::decode_alu_group(cf_node* cf, unsigned &i, unsigned &gcnt) {
 238         int r;
 239         alu_node *n;
 240         alu_group_node *g = sh->create_alu_group();
 241
 242         cgroup = !cgroup;
 243         memset(slots[cgroup], 0, 5*sizeof(slots[0][0]));
 244         gcnt = 0;
 245
 246         unsigned literal_mask = 0;
 247
 248         do {
 249                 n = sh->create_alu();
 250                 g->push_back(n);
 251
 252                 if ((r = dec->decode_alu(i, n->bc)))
 253                         return r;
 254
 255                 if (!sh->assign_slot(n, slots[cgroup])) {
 256                         assert(!"alu slot assignment failed");
 257                         return -1;
 258                 }
 259
 260                 gcnt++;
 261
 262         } while (gcnt <= 5 && !n->bc.last);
 263
 264         assert(n->bc.last);
 265
 266         for (node_iterator I = g->begin(), E = g->end(); I != E; ++I) {
 267                 n = static_cast<alu_node*>(*I);
 268
 269                 if (n->bc.dst_rel)
 270                         gpr_reladdr = true;
 271
 272                 for (int k = 0; k < n->bc.op_ptr->src_count; ++k) {
 273                         bc_alu_src &src = n->bc.src[k];
 274                         if (src.rel)
 275                                 gpr_reladdr = true;
 276                         if (src.sel == ALU_SRC_LITERAL) {
 277                                 literal_mask |= (1 << src.chan);
 278                                 src.value.u = dw[i + src.chan];
 279                         }
 280                 }
 281         }
 282
 283         unsigned literal_ndw = 0;
 284         while (literal_mask) {
 285                 g->literals.push_back(dw[i + literal_ndw]);
 286                 literal_ndw += 1;
 287                 literal_mask >>= 1;
 288         }
 289
 290         literal_ndw = (literal_ndw + 1) & ~1u;
 291
 292         i += literal_ndw;
 293         gcnt += literal_ndw >> 1;
 294
 295         cf->push_back(g);
 296         return 0;
 297 }
 298
 299 int bc_parser::prepare_alu_clause(cf_node* cf) {
 300
 301         // loop over alu groups
 302         for (node_iterator I = cf->begin(), E = cf->end(); I != E; ++I) {
 303                 assert(I->subtype == NST_ALU_GROUP);
 304                 alu_group_node *g = static_cast<alu_group_node*>(*I);
 305                 prepare_alu_group(cf, g);
 306         }
 307
 308         return 0;
 309 }
 310
 311 int bc_parser::prepare_alu_group(cf_node* cf, alu_group_node *g) {
 312
 313         alu_node *n;
 314
 315         cgroup = !cgroup;
 316         memset(slots[cgroup], 0, 5*sizeof(slots[0][0]));
 317
 318         for (node_iterator I = g->begin(), E = g->end();
 319                         I != E; ++I) {
 320                 n = static_cast<alu_node*>(*I);
 321
 322                 if (!sh->assign_slot(n, slots[cgroup])) {
 323                         assert(!"alu slot assignment failed");
 324                         return -1;
 325                 }
 326
 327                 unsigned src_count = n->bc.op_ptr->src_count;
 328
 329                 if (ctx.alu_slots(n->bc.op) & AF_4SLOT)
 330                         n->flags |= NF_ALU_4SLOT;
 331
 332                 n->src.resize(src_count);
 333
 334                 unsigned flags = n->bc.op_ptr->flags;
 335
 336                 if (flags & AF_PRED) {
 337                         n->dst.resize(3);
 338                         if (n->bc.update_pred)
 339                                 n->dst[1] = sh->get_special_value(SV_ALU_PRED);
 340                         if (n->bc.update_exec_mask)
 341                                 n->dst[2] = sh->get_special_value(SV_EXEC_MASK);
 342
 343                         n->flags |= NF_DONT_HOIST;
 344
 345                 } else if (flags & AF_KILL) {
 346
 347                         n->dst.resize(2);
 348                         n->dst[1] = sh->get_special_value(SV_VALID_MASK);
 349                         sh->set_uses_kill();
 350
 351                         n->flags |= NF_DONT_HOIST | NF_DONT_MOVE |
 352                                         NF_DONT_KILL | NF_SCHEDULE_EARLY;
 353
 354                 } else {
 355                         n->dst.resize(1);
 356                 }
 357
 358                 if (flags & AF_MOVA) {
 359
 360                         n->dst[0] = sh->get_special_value(SV_AR_INDEX);
 361
 362                         n->flags |= NF_DONT_HOIST;
 363
 364                 } else if (n->bc.op_ptr->src_count == 3 || n->bc.write_mask) {
 365                         assert(!n->bc.dst_rel || n->bc.index_mode == INDEX_AR_X);
 366
 367                         value *v = sh->get_gpr_value(false, n->bc.dst_gpr, n->bc.dst_chan,
 368                                         n->bc.dst_rel);
 369
 370                         n->dst[0] = v;
 371                 }
 372
 373                 if (n->bc.pred_sel) {
 374                         sh->has_alu_predication = true;
 375                         n->pred = sh->get_special_value(SV_ALU_PRED);
 376                 }
 377
 378                 for (unsigned s = 0; s < src_count; ++s) {
 379                         bc_alu_src &src = n->bc.src[s];
 380
 381                         if (src.sel == ALU_SRC_LITERAL) {
 382                                 n->src[s] = sh->get_const_value(src.value);
 383                         } else if (src.sel == ALU_SRC_PS || src.sel == ALU_SRC_PV) {
 384                                 unsigned pgroup = !cgroup, prev_slot = src.sel == ALU_SRC_PS ?
 385                                                 SLOT_TRANS : src.chan;
 386                                 alu_node *prev_alu = slots[pgroup][prev_slot];
 387
 388                                 assert(prev_alu);
 389
 390                                 if (!prev_alu->dst[0]) {
 391                                         value * t = sh->create_temp_value();
 392                                         prev_alu->dst[0] = t;
 393                                 }
 394
 395                                 value *d = prev_alu->dst[0];
 396
 397                                 if (d->is_rel()) {
 398                                         d = sh->get_gpr_value(true, prev_alu->bc.dst_gpr,
 399                                                               prev_alu->bc.dst_chan,
 400                                                               prev_alu->bc.dst_rel);
 401                                 }
 402
 403                                 n->src[s] = d;
 404                         } else if (ctx.is_kcache_sel(src.sel)) {
 405                                 unsigned sel = src.sel, kc_addr;
 406                                 unsigned kc_set = ((sel >> 7) & 2) + ((sel >> 5) & 1);
 407
 408                                 bc_kcache &kc = cf->bc.kc[kc_set];
 409                                 kc_addr = (kc.addr << 4) + (sel & 0x1F);
 410                                 n->src[s] = sh->get_kcache_value(kc.bank, kc_addr, src.chan);
 411                         } else if (src.sel < MAX_GPR) {
 412                                 value *v = sh->get_gpr_value(true, src.sel, src.chan, src.rel);
 413
 414                                 n->src[s] = v;
 415
 416                         } else if (src.sel >= ALU_SRC_PARAM_OFFSET) {
 417                                 // using slot for value channel because in fact the slot
 418                                 // determines the channel that is loaded by INTERP_LOAD_P0
 419                                 // (and maybe some others).
 420                                 // otherwise GVN will consider INTERP_LOAD_P0s with the same
 421                                 // param index as equal instructions and leave only one of them
 422                                 n->src[s] = sh->get_special_ro_value(sel_chan(src.sel,
 423                                                                               n->bc.slot));
 424                         } else {
 425                                 switch (src.sel) {
 426                                 case ALU_SRC_0:
 427                                         n->src[s] = sh->get_const_value(0);
 428                                         break;
 429                                 case ALU_SRC_0_5:
 430                                         n->src[s] = sh->get_const_value(0.5f);
 431                                         break;
 432                                 case ALU_SRC_1:
 433                                         n->src[s] = sh->get_const_value(1.0f);
 434                                         break;
 435                                 case ALU_SRC_1_INT:
 436                                         n->src[s] = sh->get_const_value(1);
 437                                         break;
 438                                 case ALU_SRC_M_1_INT:
 439                                         n->src[s] = sh->get_const_value(-1);
 440                                         break;
 441                                 default:
 442                                         n->src[s] = sh->get_special_ro_value(src.sel);
 443                                         break;
 444                                 }
 445                         }
 446                 }
 447         }
 448
 449         // pack multislot instructions into alu_packed_node
 450
 451         alu_packed_node *p = NULL;
 452         for (node_iterator N, I = g->begin(), E = g->end(); I != E; I = N) {
 453                 N = I + 1;
 454                 alu_node *a = static_cast<alu_node*>(*I);
 455                 unsigned sflags = a->bc.slot_flags;
 456
 457                 if (sflags == AF_4V || (ctx.is_cayman() && sflags == AF_S)) {
 458                         if (!p)
 459                                 p = sh->create_alu_packed();
 460
 461                         a->remove();
 462                         p->push_back(a);
 463                 }
 464         }
 465
 466         if (p) {
 467                 g->push_front(p);
 468
 469                 if (p->count() == 3 && ctx.is_cayman()) {
 470                         // cayman's scalar instruction that can use 3 or 4 slots
 471
 472                         // FIXME for simplicity we'll always add 4th slot,
 473                         // but probably we might want to always remove 4th slot and make
 474                         // sure that regalloc won't choose 'w' component for dst
 475
 476                         alu_node *f = static_cast<alu_node*>(p->first);
 477                         alu_node *a = sh->create_alu();
 478                         a->src = f->src;
 479                         a->dst.resize(f->dst.size());
 480                         a->bc = f->bc;
 481                         a->bc.slot = SLOT_W;
 482                         p->push_back(a);
 483                 }
 484         }
 485
 486         return 0;
 487 }
 488
 489 int bc_parser::decode_fetch_clause(cf_node* cf) {
 490         int r;
 491         unsigned i = cf->bc.addr << 1, cnt = cf->bc.count + 1;
 492
 493         cf->subtype = NST_TEX_CLAUSE;
 494
 495         while (cnt--) {
 496                 fetch_node *n = sh->create_fetch();
 497                 cf->push_back(n);
 498                 if ((r = dec->decode_fetch(i, n->bc)))
 499                         return r;
 500                 if (n->bc.src_rel || n->bc.dst_rel)
 501                         gpr_reladdr = true;
 502
 503         }
 504         return 0;
 505 }
 506
 507 int bc_parser::prepare_fetch_clause(cf_node *cf) {
 508
 509         vvec grad_v, grad_h;
 510
 511         for (node_iterator I = cf->begin(), E = cf->end(); I != E; ++I) {
 512
 513                 fetch_node *n = static_cast<fetch_node*>(*I);
 514                 assert(n->is_valid());
 515
 516                 unsigned flags = n->bc.op_ptr->flags;
 517
 518                 unsigned vtx = flags & FF_VTX;
 519                 unsigned num_src = vtx ? ctx.vtx_src_num : 4;
 520
 521                 n->dst.resize(4);
 522
 523                 if (flags & (FF_SETGRAD | FF_USEGRAD | FF_GETGRAD)) {
 524                         sh->uses_gradients = true;
 525                 }
 526
 527                 if (flags & FF_SETGRAD) {
 528
 529                         vvec *grad = NULL;
 530
 531                         switch (n->bc.op) {
 532                                 case FETCH_OP_SET_GRADIENTS_V:
 533                                         grad = &grad_v;
 534                                         break;
 535                                 case FETCH_OP_SET_GRADIENTS_H:
 536                                         grad = &grad_h;
 537                                         break;
 538                                 default:
 539                                         assert(!"unexpected SET_GRAD instruction");
 540                                         return -1;
 541                         }
 542
 543                         if (grad->empty())
 544                                 grad->resize(4);
 545
 546                         for(unsigned s = 0; s < 4; ++s) {
 547                                 unsigned sw = n->bc.src_sel[s];
 548                                 if (sw <= SEL_W)
 549                                         (*grad)[s] = sh->get_gpr_value(true, n->bc.src_gpr,
 550                                                                        sw, false);
 551                                 else if (sw == SEL_0)
 552                                         (*grad)[s] = sh->get_const_value(0.0f);
 553                                 else if (sw == SEL_1)
 554                                         (*grad)[s] = sh->get_const_value(1.0f);
 555                         }
 556                 } else {
 557
 558                         if (flags & FF_USEGRAD) {
 559                                 n->src.resize(12);
 560                                 std::copy(grad_v.begin(), grad_v.end(), n->src.begin() + 4);
 561                                 std::copy(grad_h.begin(), grad_h.end(), n->src.begin() + 8);
 562                         } else {
 563                                 n->src.resize(4);
 564                         }
 565
 566                         for(int s = 0; s < 4; ++s) {
 567                                 if (n->bc.dst_sel[s] != SEL_MASK)
 568                                         n->dst[s] = sh->get_gpr_value(false, n->bc.dst_gpr, s, false);
 569                                 // NOTE: it doesn't matter here which components of the result we
 570                                 // are using, but original n->bc.dst_sel should be taken into
 571                                 // account when building the bytecode
 572                         }
 573                         for(unsigned s = 0; s < num_src; ++s) {
 574                                 if (n->bc.src_sel[s] <= SEL_W)
 575                                         n->src[s] = sh->get_gpr_value(true, n->bc.src_gpr,
 576                                                                       n->bc.src_sel[s], false);
 577                         }
 578
 579                 }
 580         }
 581
 582         return 0;
 583 }
 584
 585 int bc_parser::prepare_ir() {
 586
 587         for(id_cf_map::iterator I = cf_map.begin(), E = cf_map.end(); I != E; ++I) {
 588                 cf_node *c = *I;
 589
 590                 if (!c)
 591                         continue;
 592
 593                 unsigned flags = c->bc.op_ptr->flags;
 594
 595                 if (flags & CF_ALU) {
 596                         prepare_alu_clause(c);
 597                 } else if (flags & CF_FETCH) {
 598                         prepare_fetch_clause(c);
 599                 } else if (c->bc.op == CF_OP_CALL_FS) {
 600                         sh->init_call_fs(c);
 601                         c->flags |= NF_SCHEDULE_EARLY | NF_DONT_MOVE;
 602                 } else if (flags & CF_LOOP_START) {
 603                         prepare_loop(c);
 604                 } else if (c->bc.op == CF_OP_JUMP) {
 605                         prepare_if(c);
 606                 } else if (c->bc.op == CF_OP_LOOP_END) {
 607                         loop_stack.pop();
 608                 } else if (c->bc.op == CF_OP_LOOP_CONTINUE) {
 609                         assert(!loop_stack.empty());
 610                         repeat_node *rep = sh->create_repeat(loop_stack.top());
 611                         if (c->parent->first != c)
 612                                 rep->move(c->parent->first, c);
 613                         c->replace_with(rep);
 614                         sh->simplify_dep_rep(rep);
 615                 } else if (c->bc.op == CF_OP_LOOP_BREAK) {
 616                         assert(!loop_stack.empty());
 617                         depart_node *dep = sh->create_depart(loop_stack.top());
 618                         if (c->parent->first != c)
 619                                 dep->move(c->parent->first, c);
 620                         c->replace_with(dep);
 621                         sh->simplify_dep_rep(dep);
 622                 } else if (flags & CF_EXP) {
 623
 624                         // unroll burst exports
 625
 626                         assert(c->bc.op == CF_OP_EXPORT || c->bc.op == CF_OP_EXPORT_DONE);
 627
 628                         c->bc.set_op(CF_OP_EXPORT);
 629
 630                         unsigned burst_count = c->bc.burst_count;
 631                         unsigned eop = c->bc.end_of_program;
 632
 633                         c->bc.end_of_program = 0;
 634                         c->bc.burst_count = 0;
 635
 636                         do {
 637                                 c->src.resize(4);
 638
 639                                 for(int s = 0; s < 4; ++s) {
 640                                         switch (c->bc.sel[s]) {
 641                                         case SEL_0:
 642                                                 c->src[s] = sh->get_const_value(0.0f);
 643                                                 break;
 644                                         case SEL_1:
 645                                                 c->src[s] = sh->get_const_value(1.0f);
 646                                                 break;
 647                                         case SEL_MASK:
 648                                                 break;
 649                                         default:
 650                                                 if (c->bc.sel[s] <= SEL_W)
 651                                                         c->src[s] = sh->get_gpr_value(true, c->bc.rw_gpr,
 652                                                                         c->bc.sel[s], false);
 653                                                 else
 654                                                         assert(!"invalid src_sel for export");
 655                                         }
 656                                 }
 657
 658                                 if (!burst_count--)
 659                                         break;
 660
 661                                 cf_node *cf_next = sh->create_cf();
 662                                 cf_next->bc = c->bc;
 663                                 ++cf_next->bc.rw_gpr;
 664                                 ++cf_next->bc.array_base;
 665
 666                                 c->insert_after(cf_next);
 667                                 c = cf_next;
 668
 669                         } while (1);
 670
 671                         c->bc.end_of_program = eop;
 672                 } else if (flags & (CF_STRM | CF_RAT)) {
 673
 674                         unsigned burst_count = c->bc.burst_count;
 675                         unsigned eop = c->bc.end_of_program;
 676
 677                         c->bc.end_of_program = 0;
 678                         c->bc.burst_count = 0;
 679
 680                         do {
 681
 682                                 c->src.resize(4);
 683
 684                                 for(int s = 0; s < 4; ++s) {
 685                                         if (c->bc.comp_mask & (1 << s))
 686                                                 c->src[s] =
 687                                                                 sh->get_gpr_value(true, c->bc.rw_gpr, s, false);
 688                                 }
 689
 690                                 if ((flags & CF_RAT) && (c->bc.type & 1)) { // indexed write
 691                                         c->src.resize(8);
 692                                         for(int s = 0; s < 3; ++s) {
 693                                                 c->src[4 + s] =
 694                                                         sh->get_gpr_value(true, c->bc.index_gpr, s, false);
 695                                         }
 696
 697                                         // FIXME probably we can relax it a bit
 698                                         c->flags |= NF_DONT_HOIST | NF_DONT_MOVE;
 699                                 }
 700
 701                                 if (!burst_count--)
 702                                         break;
 703
 704                                 cf_node *cf_next = sh->create_cf();
 705                                 cf_next->bc = c->bc;
 706                                 ++cf_next->bc.rw_gpr;
 707
 708                                 // FIXME is it correct?
 709                                 cf_next->bc.array_base += cf_next->bc.elem_size + 1;
 710
 711                                 c->insert_after(cf_next);
 712                                 c = cf_next;
 713                         } while (1);
 714
 715                         c->bc.end_of_program = eop;
 716
 717                 }
 718         }
 719
 720         assert(loop_stack.empty());
 721         return 0;
 722 }
 723
 724 int bc_parser::prepare_loop(cf_node* c) {
 725
 726         cf_node *end = cf_map[c->bc.addr - 1];
 727         assert(end->bc.op == CF_OP_LOOP_END);
 728         assert(c->parent == end->parent);
 729
 730         region_node *reg = sh->create_region();
 731         repeat_node *rep = sh->create_repeat(reg);
 732
 733         reg->push_back(rep);
 734         c->insert_before(reg);
 735         rep->move(c, end->next);
 736
 737         loop_stack.push(reg);
 738         return 0;
 739 }
 740
 741 int bc_parser::prepare_if(cf_node* c) {
 742         cf_node *c_else = NULL, *end = cf_map[c->bc.addr];
 743
 744         BCP_DUMP(
 745                 sblog << "parsing JUMP @" << c->bc.id;
 746                 sblog << "\n";
 747         );
 748
 749         if (end->bc.op == CF_OP_ELSE) {
 750                 BCP_DUMP(
 751                         sblog << "  found ELSE : ";
 752                         dump::dump_op(end);
 753                         sblog << "\n";
 754                 );
 755
 756                 c_else = end;
 757                 end = cf_map[c_else->bc.addr];
 758         } else {
 759                 BCP_DUMP(
 760                         sblog << "  no else\n";
 761                 );
 762
 763                 c_else = end;
 764         }
 765
 766         if (c_else->parent != c->parent)
 767                 c_else = NULL;
 768
 769         if (end->parent != c->parent)
 770                 end = NULL;
 771
 772         region_node *reg = sh->create_region();
 773
 774         depart_node *dep2 = sh->create_depart(reg);
 775         depart_node *dep = sh->create_depart(reg);
 776         if_node *n_if = sh->create_if();
 777
 778         c->insert_before(reg);
 779
 780         if (c_else != end)
 781                 dep->move(c_else, end);
 782         dep2->move(c, end);
 783
 784         reg->push_back(dep);
 785         dep->push_front(n_if);
 786         n_if->push_back(dep2);
 787
 788         n_if->cond = sh->get_special_value(SV_EXEC_MASK);
 789
 790         return 0;
 791 }
 792
 793
 794 } // namespace r600_sb