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         sh->safe_math = sb_context::safe_math || (t == TARGET_COMPUTE);
  74
  75         int r = decode_shader();
  76
  77         delete dec;
  78
  79         sh->ngpr = bc->ngpr;
  80         sh->nstack = bc->nstack;
  81
  82         return r;
  83 }
  84
  85 int bc_parser::decode_shader() {
  86         int r = 0;
  87         unsigned i = 0;
  88         bool eop = false;
  89
  90         sh->init();
  91
  92         do {
  93                 eop = false;
  94                 if ((r = decode_cf(i, eop)))
  95                         return r;
  96
  97         } while (!eop || (i >> 1) <= max_cf);
  98
  99         return 0;
 100 }
 101
 102 int bc_parser::prepare() {
 103         int r = 0;
 104         if ((r = parse_decls()))
 105                 return r;
 106         if ((r = prepare_ir()))
 107                 return r;
 108         return 0;
 109 }
 110
 111 int bc_parser::parse_decls() {
 112
 113         if (!pshader) {
 114                 if (gpr_reladdr)
 115                         sh->add_gpr_array(0, bc->ngpr, 0x0F);
 116
 117                 // compute shaders have some values preloaded in R0, R1
 118                 sh->add_input(0 /* GPR */, true /* preloaded */, 0x0F /* mask */);
 119                 sh->add_input(1 /* GPR */, true /* preloaded */, 0x0F /* mask */);
 120                 return 0;
 121         }
 122
 123         if (pshader->indirect_files & ~(1 << TGSI_FILE_CONSTANT)) {
 124
 125                 assert(pshader->num_arrays);
 126
 127                 if (pshader->num_arrays) {
 128                         for (unsigned i = 0; i < pshader->num_arrays; ++i) {
 129                                 r600_shader_array &a = pshader->arrays[i];
 130                                 sh->add_gpr_array(a.gpr_start, a.gpr_count, a.comp_mask);
 131                         }
 132                 } else {
 133                         sh->add_gpr_array(0, pshader->bc.ngpr, 0x0F);
 134                 }
 135         }
 136
 137         if (sh->target == TARGET_VS)
 138                 sh->add_input(0, 1, 0x0F);
 139
 140         bool ps_interp = ctx.hw_class >= HW_CLASS_EVERGREEN
 141                         && sh->target == TARGET_PS;
 142
 143         unsigned linear = 0, persp = 0, centroid = 1;
 144
 145         for (unsigned i = 0; i < pshader->ninput; ++i) {
 146                 r600_shader_io & in = pshader->input[i];
 147                 bool preloaded = sh->target == TARGET_PS && !(ps_interp && in.spi_sid);
 148                 sh->add_input(in.gpr, preloaded, /*in.write_mask*/ 0x0F);
 149                 if (ps_interp && in.spi_sid) {
 150                         if (in.interpolate == TGSI_INTERPOLATE_LINEAR ||
 151                                         in.interpolate == TGSI_INTERPOLATE_COLOR)
 152                                 linear = 1;
 153                         else if (in.interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
 154                                 persp = 1;
 155                         if (in.centroid)
 156                                 centroid = 2;
 157                 }
 158         }
 159
 160         if (ps_interp) {
 161                 unsigned mask = (1 << (2 * (linear + persp) * centroid)) - 1;
 162                 unsigned gpr = 0;
 163
 164                 while (mask) {
 165                         sh->add_input(gpr, true, mask & 0x0F);
 166                         ++gpr;
 167                         mask >>= 4;
 168                 }
 169         }
 170
 171         return 0;
 172 }
 173
 174 int bc_parser::decode_cf(unsigned &i, bool &eop) {
 175
 176         int r;
 177
 178         cf_node *cf = sh->create_cf();
 179         sh->root->push_back(cf);
 180
 181         unsigned id = i >> 1;
 182
 183         cf->bc.id = id;
 184
 185         if (cf_map.size() < id + 1)
 186                 cf_map.resize(id + 1);
 187
 188         cf_map[id] = cf;
 189
 190         if ((r = dec->decode_cf(i, cf->bc)))
 191                 return r;
 192
 193         cf_op_flags flags = (cf_op_flags)cf->bc.op_ptr->flags;
 194
 195         if (flags & CF_ALU) {
 196                 if ((r = decode_alu_clause(cf)))
 197                         return r;
 198         } else if (flags & CF_FETCH) {
 199                 if ((r = decode_fetch_clause(cf)))
 200                         return r;;
 201         } else if (flags & CF_EXP) {
 202                 if (cf->bc.rw_rel)
 203                         gpr_reladdr = true;
 204                 assert(!cf->bc.rw_rel);
 205         } else if (flags & (CF_STRM | CF_RAT)) {
 206                 if (cf->bc.rw_rel)
 207                         gpr_reladdr = true;
 208                 assert(!cf->bc.rw_rel);
 209         } else if (flags & CF_BRANCH) {
 210                 if (cf->bc.addr > max_cf)
 211                         max_cf = cf->bc.addr;
 212         }
 213
 214         eop = cf->bc.end_of_program || cf->bc.op == CF_OP_CF_END ||
 215                         cf->bc.op == CF_OP_RET;
 216         return 0;
 217 }
 218
 219 int bc_parser::decode_alu_clause(cf_node* cf) {
 220         unsigned i = cf->bc.addr << 1, cnt = cf->bc.count + 1, gcnt;
 221
 222         cf->subtype = NST_ALU_CLAUSE;
 223
 224         cgroup = 0;
 225         memset(slots[0], 0, 5*sizeof(slots[0][0]));
 226
 227         unsigned ng = 0;
 228
 229         do {
 230                 decode_alu_group(cf, i, gcnt);
 231                 assert(gcnt <= cnt);
 232                 cnt -= gcnt;
 233                 ng++;
 234         } while (cnt);
 235
 236         return 0;
 237 }
 238
 239 int bc_parser::decode_alu_group(cf_node* cf, unsigned &i, unsigned &gcnt) {
 240         int r;
 241         alu_node *n;
 242         alu_group_node *g = sh->create_alu_group();
 243
 244         cgroup = !cgroup;
 245         memset(slots[cgroup], 0, 5*sizeof(slots[0][0]));
 246         gcnt = 0;
 247
 248         unsigned literal_mask = 0;
 249
 250         do {
 251                 n = sh->create_alu();
 252                 g->push_back(n);
 253
 254                 if ((r = dec->decode_alu(i, n->bc)))
 255                         return r;
 256
 257                 if (!sh->assign_slot(n, slots[cgroup])) {
 258                         assert(!"alu slot assignment failed");
 259                         return -1;
 260                 }
 261
 262                 gcnt++;
 263
 264         } while (gcnt <= 5 && !n->bc.last);
 265
 266         assert(n->bc.last);
 267
 268         for (node_iterator I = g->begin(), E = g->end(); I != E; ++I) {
 269                 n = static_cast<alu_node*>(*I);
 270
 271                 if (n->bc.dst_rel)
 272                         gpr_reladdr = true;
 273
 274                 for (int k = 0; k < n->bc.op_ptr->src_count; ++k) {
 275                         bc_alu_src &src = n->bc.src[k];
 276                         if (src.rel)
 277                                 gpr_reladdr = true;
 278                         if (src.sel == ALU_SRC_LITERAL) {
 279                                 literal_mask |= (1 << src.chan);
 280                                 src.value.u = dw[i + src.chan];
 281                         }
 282                 }
 283         }
 284
 285         unsigned literal_ndw = 0;
 286         while (literal_mask) {
 287                 g->literals.push_back(dw[i + literal_ndw]);
 288                 literal_ndw += 1;
 289                 literal_mask >>= 1;
 290         }
 291
 292         literal_ndw = (literal_ndw + 1) & ~1u;
 293
 294         i += literal_ndw;
 295         gcnt += literal_ndw >> 1;
 296
 297         cf->push_back(g);
 298         return 0;
 299 }
 300
 301 int bc_parser::prepare_alu_clause(cf_node* cf) {
 302
 303         // loop over alu groups
 304         for (node_iterator I = cf->begin(), E = cf->end(); I != E; ++I) {
 305                 assert(I->subtype == NST_ALU_GROUP);
 306                 alu_group_node *g = static_cast<alu_group_node*>(*I);
 307                 prepare_alu_group(cf, g);
 308         }
 309
 310         return 0;
 311 }
 312
 313 int bc_parser::prepare_alu_group(cf_node* cf, alu_group_node *g) {
 314
 315         alu_node *n;
 316
 317         cgroup = !cgroup;
 318         memset(slots[cgroup], 0, 5*sizeof(slots[0][0]));
 319
 320         for (node_iterator I = g->begin(), E = g->end();
 321                         I != E; ++I) {
 322                 n = static_cast<alu_node*>(*I);
 323
 324                 if (!sh->assign_slot(n, slots[cgroup])) {
 325                         assert(!"alu slot assignment failed");
 326                         return -1;
 327                 }
 328
 329                 unsigned src_count = n->bc.op_ptr->src_count;
 330
 331                 if (ctx.alu_slots(n->bc.op) & AF_4SLOT)
 332                         n->flags |= NF_ALU_4SLOT;
 333
 334                 n->src.resize(src_count);
 335
 336                 unsigned flags = n->bc.op_ptr->flags;
 337
 338                 if (flags & AF_PRED) {
 339                         n->dst.resize(3);
 340                         if (n->bc.update_pred)
 341                                 n->dst[1] = sh->get_special_value(SV_ALU_PRED);
 342                         if (n->bc.update_exec_mask)
 343                                 n->dst[2] = sh->get_special_value(SV_EXEC_MASK);
 344
 345                         n->flags |= NF_DONT_HOIST;
 346
 347                 } else if (flags & AF_KILL) {
 348
 349                         n->dst.resize(2);
 350                         n->dst[1] = sh->get_special_value(SV_VALID_MASK);
 351                         sh->set_uses_kill();
 352
 353                         n->flags |= NF_DONT_HOIST | NF_DONT_MOVE |
 354                                         NF_DONT_KILL | NF_SCHEDULE_EARLY;
 355
 356                 } else {
 357                         n->dst.resize(1);
 358                 }
 359
 360                 if (flags & AF_MOVA) {
 361
 362                         n->dst[0] = sh->get_special_value(SV_AR_INDEX);
 363
 364                         n->flags |= NF_DONT_HOIST;
 365
 366                 } else if (n->bc.op_ptr->src_count == 3 || n->bc.write_mask) {
 367                         assert(!n->bc.dst_rel || n->bc.index_mode == INDEX_AR_X);
 368
 369                         value *v = sh->get_gpr_value(false, n->bc.dst_gpr, n->bc.dst_chan,
 370                                         n->bc.dst_rel);
 371
 372                         n->dst[0] = v;
 373                 }
 374
 375                 if (n->bc.pred_sel) {
 376                         sh->has_alu_predication = true;
 377                         n->pred = sh->get_special_value(SV_ALU_PRED);
 378                 }
 379
 380                 for (unsigned s = 0; s < src_count; ++s) {
 381                         bc_alu_src &src = n->bc.src[s];
 382
 383                         if (src.sel == ALU_SRC_LITERAL) {
 384                                 n->src[s] = sh->get_const_value(src.value);
 385                         } else if (src.sel == ALU_SRC_PS || src.sel == ALU_SRC_PV) {
 386                                 unsigned pgroup = !cgroup, prev_slot = src.sel == ALU_SRC_PS ?
 387                                                 SLOT_TRANS : src.chan;
 388
 389                                 // XXX shouldn't happen but llvm backend uses PS on cayman
 390                                 if (prev_slot == SLOT_TRANS && ctx.is_cayman())
 391                                         prev_slot = SLOT_X;
 392
 393                                 alu_node *prev_alu = slots[pgroup][prev_slot];
 394
 395                                 assert(prev_alu);
 396
 397                                 if (!prev_alu->dst[0]) {
 398                                         value * t = sh->create_temp_value();
 399                                         prev_alu->dst[0] = t;
 400                                 }
 401
 402                                 value *d = prev_alu->dst[0];
 403
 404                                 if (d->is_rel()) {
 405                                         d = sh->get_gpr_value(true, prev_alu->bc.dst_gpr,
 406                                                               prev_alu->bc.dst_chan,
 407                                                               prev_alu->bc.dst_rel);
 408                                 }
 409
 410                                 n->src[s] = d;
 411                         } else if (ctx.is_kcache_sel(src.sel)) {
 412                                 unsigned sel = src.sel, kc_addr;
 413                                 unsigned kc_set = ((sel >> 7) & 2) + ((sel >> 5) & 1);
 414
 415                                 bc_kcache &kc = cf->bc.kc[kc_set];
 416                                 kc_addr = (kc.addr << 4) + (sel & 0x1F);
 417                                 n->src[s] = sh->get_kcache_value(kc.bank, kc_addr, src.chan);
 418                         } else if (src.sel < MAX_GPR) {
 419                                 value *v = sh->get_gpr_value(true, src.sel, src.chan, src.rel);
 420
 421                                 n->src[s] = v;
 422
 423                         } else if (src.sel >= ALU_SRC_PARAM_OFFSET) {
 424                                 // using slot for value channel because in fact the slot
 425                                 // determines the channel that is loaded by INTERP_LOAD_P0
 426                                 // (and maybe some others).
 427                                 // otherwise GVN will consider INTERP_LOAD_P0s with the same
 428                                 // param index as equal instructions and leave only one of them
 429                                 n->src[s] = sh->get_special_ro_value(sel_chan(src.sel,
 430                                                                               n->bc.slot));
 431                         } else {
 432                                 switch (src.sel) {
 433                                 case ALU_SRC_0:
 434                                         n->src[s] = sh->get_const_value(0);
 435                                         break;
 436                                 case ALU_SRC_0_5:
 437                                         n->src[s] = sh->get_const_value(0.5f);
 438                                         break;
 439                                 case ALU_SRC_1:
 440                                         n->src[s] = sh->get_const_value(1.0f);
 441                                         break;
 442                                 case ALU_SRC_1_INT:
 443                                         n->src[s] = sh->get_const_value(1);
 444                                         break;
 445                                 case ALU_SRC_M_1_INT:
 446                                         n->src[s] = sh->get_const_value(-1);
 447                                         break;
 448                                 default:
 449                                         n->src[s] = sh->get_special_ro_value(src.sel);
 450                                         break;
 451                                 }
 452                         }
 453                 }
 454         }
 455
 456         // pack multislot instructions into alu_packed_node
 457
 458         alu_packed_node *p = NULL;
 459         for (node_iterator N, I = g->begin(), E = g->end(); I != E; I = N) {
 460                 N = I + 1;
 461                 alu_node *a = static_cast<alu_node*>(*I);
 462                 unsigned sflags = a->bc.slot_flags;
 463
 464                 if (sflags == AF_4V || (ctx.is_cayman() && sflags == AF_S)) {
 465                         if (!p)
 466                                 p = sh->create_alu_packed();
 467
 468                         a->remove();
 469                         p->push_back(a);
 470                 }
 471         }
 472
 473         if (p) {
 474                 g->push_front(p);
 475
 476                 if (p->count() == 3 && ctx.is_cayman()) {
 477                         // cayman's scalar instruction that can use 3 or 4 slots
 478
 479                         // FIXME for simplicity we'll always add 4th slot,
 480                         // but probably we might want to always remove 4th slot and make
 481                         // sure that regalloc won't choose 'w' component for dst
 482
 483                         alu_node *f = static_cast<alu_node*>(p->first);
 484                         alu_node *a = sh->create_alu();
 485                         a->src = f->src;
 486                         a->dst.resize(f->dst.size());
 487                         a->bc = f->bc;
 488                         a->bc.slot = SLOT_W;
 489                         p->push_back(a);
 490                 }
 491         }
 492
 493         return 0;
 494 }
 495
 496 int bc_parser::decode_fetch_clause(cf_node* cf) {
 497         int r;
 498         unsigned i = cf->bc.addr << 1, cnt = cf->bc.count + 1;
 499
 500         cf->subtype = NST_TEX_CLAUSE;
 501
 502         while (cnt--) {
 503                 fetch_node *n = sh->create_fetch();
 504                 cf->push_back(n);
 505                 if ((r = dec->decode_fetch(i, n->bc)))
 506                         return r;
 507                 if (n->bc.src_rel || n->bc.dst_rel)
 508                         gpr_reladdr = true;
 509
 510         }
 511         return 0;
 512 }
 513
 514 int bc_parser::prepare_fetch_clause(cf_node *cf) {
 515
 516         vvec grad_v, grad_h;
 517
 518         for (node_iterator I = cf->begin(), E = cf->end(); I != E; ++I) {
 519
 520                 fetch_node *n = static_cast<fetch_node*>(*I);
 521                 assert(n->is_valid());
 522
 523                 unsigned flags = n->bc.op_ptr->flags;
 524
 525                 unsigned vtx = flags & FF_VTX;
 526                 unsigned num_src = vtx ? ctx.vtx_src_num : 4;
 527
 528                 n->dst.resize(4);
 529
 530                 if (flags & (FF_SETGRAD | FF_USEGRAD | FF_GETGRAD)) {
 531                         sh->uses_gradients = true;
 532                 }
 533
 534                 if (flags & FF_SETGRAD) {
 535
 536                         vvec *grad = NULL;
 537
 538                         switch (n->bc.op) {
 539                                 case FETCH_OP_SET_GRADIENTS_V:
 540                                         grad = &grad_v;
 541                                         break;
 542                                 case FETCH_OP_SET_GRADIENTS_H:
 543                                         grad = &grad_h;
 544                                         break;
 545                                 default:
 546                                         assert(!"unexpected SET_GRAD instruction");
 547                                         return -1;
 548                         }
 549
 550                         if (grad->empty())
 551                                 grad->resize(4);
 552
 553                         for(unsigned s = 0; s < 4; ++s) {
 554                                 unsigned sw = n->bc.src_sel[s];
 555                                 if (sw <= SEL_W)
 556                                         (*grad)[s] = sh->get_gpr_value(true, n->bc.src_gpr,
 557                                                                        sw, false);
 558                                 else if (sw == SEL_0)
 559                                         (*grad)[s] = sh->get_const_value(0.0f);
 560                                 else if (sw == SEL_1)
 561                                         (*grad)[s] = sh->get_const_value(1.0f);
 562                         }
 563                 } else {
 564
 565                         if (flags & FF_USEGRAD) {
 566                                 n->src.resize(12);
 567                                 std::copy(grad_v.begin(), grad_v.end(), n->src.begin() + 4);
 568                                 std::copy(grad_h.begin(), grad_h.end(), n->src.begin() + 8);
 569                         } else {
 570                                 n->src.resize(4);
 571                         }
 572
 573                         for(int s = 0; s < 4; ++s) {
 574                                 if (n->bc.dst_sel[s] != SEL_MASK)
 575                                         n->dst[s] = sh->get_gpr_value(false, n->bc.dst_gpr, s, false);
 576                                 // NOTE: it doesn't matter here which components of the result we
 577                                 // are using, but original n->bc.dst_sel should be taken into
 578                                 // account when building the bytecode
 579                         }
 580                         for(unsigned s = 0; s < num_src; ++s) {
 581                                 if (n->bc.src_sel[s] <= SEL_W)
 582                                         n->src[s] = sh->get_gpr_value(true, n->bc.src_gpr,
 583                                                                       n->bc.src_sel[s], false);
 584                         }
 585
 586                 }
 587         }
 588
 589         return 0;
 590 }
 591
 592 int bc_parser::prepare_ir() {
 593
 594         for(id_cf_map::iterator I = cf_map.begin(), E = cf_map.end(); I != E; ++I) {
 595                 cf_node *c = *I;
 596
 597                 if (!c)
 598                         continue;
 599
 600                 unsigned flags = c->bc.op_ptr->flags;
 601
 602                 if (flags & CF_ALU) {
 603                         prepare_alu_clause(c);
 604                 } else if (flags & CF_FETCH) {
 605                         prepare_fetch_clause(c);
 606                 } else if (c->bc.op == CF_OP_CALL_FS) {
 607                         sh->init_call_fs(c);
 608                         c->flags |= NF_SCHEDULE_EARLY | NF_DONT_MOVE;
 609                 } else if (flags & CF_LOOP_START) {
 610                         prepare_loop(c);
 611                 } else if (c->bc.op == CF_OP_JUMP) {
 612                         prepare_if(c);
 613                 } else if (c->bc.op == CF_OP_LOOP_END) {
 614                         loop_stack.pop();
 615                 } else if (c->bc.op == CF_OP_LOOP_CONTINUE) {
 616                         assert(!loop_stack.empty());
 617                         repeat_node *rep = sh->create_repeat(loop_stack.top());
 618                         if (c->parent->first != c)
 619                                 rep->move(c->parent->first, c);
 620                         c->replace_with(rep);
 621                         sh->simplify_dep_rep(rep);
 622                 } else if (c->bc.op == CF_OP_LOOP_BREAK) {
 623                         assert(!loop_stack.empty());
 624                         depart_node *dep = sh->create_depart(loop_stack.top());
 625                         if (c->parent->first != c)
 626                                 dep->move(c->parent->first, c);
 627                         c->replace_with(dep);
 628                         sh->simplify_dep_rep(dep);
 629                 } else if (flags & CF_EXP) {
 630
 631                         // unroll burst exports
 632
 633                         assert(c->bc.op == CF_OP_EXPORT || c->bc.op == CF_OP_EXPORT_DONE);
 634
 635                         c->bc.set_op(CF_OP_EXPORT);
 636
 637                         unsigned burst_count = c->bc.burst_count;
 638                         unsigned eop = c->bc.end_of_program;
 639
 640                         c->bc.end_of_program = 0;
 641                         c->bc.burst_count = 0;
 642
 643                         do {
 644                                 c->src.resize(4);
 645
 646                                 for(int s = 0; s < 4; ++s) {
 647                                         switch (c->bc.sel[s]) {
 648                                         case SEL_0:
 649                                                 c->src[s] = sh->get_const_value(0.0f);
 650                                                 break;
 651                                         case SEL_1:
 652                                                 c->src[s] = sh->get_const_value(1.0f);
 653                                                 break;
 654                                         case SEL_MASK:
 655                                                 break;
 656                                         default:
 657                                                 if (c->bc.sel[s] <= SEL_W)
 658                                                         c->src[s] = sh->get_gpr_value(true, c->bc.rw_gpr,
 659                                                                         c->bc.sel[s], false);
 660                                                 else
 661                                                         assert(!"invalid src_sel for export");
 662                                         }
 663                                 }
 664
 665                                 if (!burst_count--)
 666                                         break;
 667
 668                                 cf_node *cf_next = sh->create_cf();
 669                                 cf_next->bc = c->bc;
 670                                 ++cf_next->bc.rw_gpr;
 671                                 ++cf_next->bc.array_base;
 672
 673                                 c->insert_after(cf_next);
 674                                 c = cf_next;
 675
 676                         } while (1);
 677
 678                         c->bc.end_of_program = eop;
 679                 } else if (flags & (CF_STRM | CF_RAT)) {
 680
 681                         unsigned burst_count = c->bc.burst_count;
 682                         unsigned eop = c->bc.end_of_program;
 683
 684                         c->bc.end_of_program = 0;
 685                         c->bc.burst_count = 0;
 686
 687                         do {
 688
 689                                 c->src.resize(4);
 690
 691                                 for(int s = 0; s < 4; ++s) {
 692                                         if (c->bc.comp_mask & (1 << s))
 693                                                 c->src[s] =
 694                                                                 sh->get_gpr_value(true, c->bc.rw_gpr, s, false);
 695                                 }
 696
 697                                 if ((flags & CF_RAT) && (c->bc.type & 1)) { // indexed write
 698                                         c->src.resize(8);
 699                                         for(int s = 0; s < 3; ++s) {
 700                                                 c->src[4 + s] =
 701                                                         sh->get_gpr_value(true, c->bc.index_gpr, s, false);
 702                                         }
 703
 704                                         // FIXME probably we can relax it a bit
 705                                         c->flags |= NF_DONT_HOIST | NF_DONT_MOVE;
 706                                 }
 707
 708                                 if (!burst_count--)
 709                                         break;
 710
 711                                 cf_node *cf_next = sh->create_cf();
 712                                 cf_next->bc = c->bc;
 713                                 ++cf_next->bc.rw_gpr;
 714
 715                                 // FIXME is it correct?
 716                                 cf_next->bc.array_base += cf_next->bc.elem_size + 1;
 717
 718                                 c->insert_after(cf_next);
 719                                 c = cf_next;
 720                         } while (1);
 721
 722                         c->bc.end_of_program = eop;
 723
 724                 }
 725         }
 726
 727         assert(loop_stack.empty());
 728         return 0;
 729 }
 730
 731 int bc_parser::prepare_loop(cf_node* c) {
 732
 733         cf_node *end = cf_map[c->bc.addr - 1];
 734         assert(end->bc.op == CF_OP_LOOP_END);
 735         assert(c->parent == end->parent);
 736
 737         region_node *reg = sh->create_region();
 738         repeat_node *rep = sh->create_repeat(reg);
 739
 740         reg->push_back(rep);
 741         c->insert_before(reg);
 742         rep->move(c, end->next);
 743
 744         loop_stack.push(reg);
 745         return 0;
 746 }
 747
 748 int bc_parser::prepare_if(cf_node* c) {
 749         cf_node *c_else = NULL, *end = cf_map[c->bc.addr];
 750
 751         BCP_DUMP(
 752                 sblog << "parsing JUMP @" << c->bc.id;
 753                 sblog << "\n";
 754         );
 755
 756         if (end->bc.op == CF_OP_ELSE) {
 757                 BCP_DUMP(
 758                         sblog << "  found ELSE : ";
 759                         dump::dump_op(end);
 760                         sblog << "\n";
 761                 );
 762
 763                 c_else = end;
 764                 end = cf_map[c_else->bc.addr];
 765         } else {
 766                 BCP_DUMP(
 767                         sblog << "  no else\n";
 768                 );
 769
 770                 c_else = end;
 771         }
 772
 773         if (c_else->parent != c->parent)
 774                 c_else = NULL;
 775
 776         if (end->parent != c->parent)
 777                 end = NULL;
 778
 779         region_node *reg = sh->create_region();
 780
 781         depart_node *dep2 = sh->create_depart(reg);
 782         depart_node *dep = sh->create_depart(reg);
 783         if_node *n_if = sh->create_if();
 784
 785         c->insert_before(reg);
 786
 787         if (c_else != end)
 788                 dep->move(c_else, end);
 789         dep2->move(c, end);
 790
 791         reg->push_back(dep);
 792         dep->push_front(n_if);
 793         n_if->push_back(dep2);
 794
 795         n_if->cond = sh->get_special_value(SV_EXEC_MASK);
 796
 797         return 0;
 798 }
 799
 800
 801 } // namespace r600_sb