src/amd/compiler/aco_insert_exec_mask.cpp

   1 /*
   2  * Copyright © 2019 Valve Corporation
   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 (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 NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include "aco_ir.h"
  26 #include "aco_builder.h"
  27
  28 namespace aco {
  29
  30 namespace {
  31
  32 enum WQMState : uint8_t {
  33    Unspecified = 0,
  34    Exact = 1 << 0,
  35    WQM = 1 << 1, /* with control flow applied */
  36    Preserve_WQM = 1 << 2,
  37    Exact_Branch = 1 << 3,
  38 };
  39
  40 enum mask_type : uint8_t {
  41    mask_type_global = 1 << 0,
  42    mask_type_exact = 1 << 1,
  43    mask_type_wqm = 1 << 2,
  44    mask_type_loop = 1 << 3, /* active lanes of a loop */
  45    mask_type_initial = 1 << 4, /* initially active lanes */
  46 };
  47
  48 struct wqm_ctx {
  49    Program* program;
  50    /* state for WQM propagation */
  51    std::set<unsigned> worklist;
  52    std::vector<uint16_t> defined_in;
  53    std::vector<bool> needs_wqm;
  54    std::vector<bool> branch_wqm; /* true if the branch condition in this block should be in wqm */
  55    bool loop;
  56    bool wqm;
  57    wqm_ctx(Program* program) : program(program),
  58                                defined_in(program->peekAllocationId(), 0xFFFF),
  59                                needs_wqm(program->peekAllocationId()),
  60                                branch_wqm(program->blocks.size()),
  61                                loop(false),
  62                                wqm(false)
  63    {
  64       for (unsigned i = 0; i < program->blocks.size(); i++)
  65          worklist.insert(i);
  66    }
  67 };
  68
  69 struct loop_info {
  70    Block* loop_header;
  71    uint16_t num_exec_masks;
  72    uint8_t needs;
  73    bool has_divergent_break;
  74    bool has_divergent_continue;
  75    bool has_discard;
  76    loop_info(Block* b, uint16_t num, uint8_t needs, bool breaks, bool cont, bool discard) :
  77              loop_header(b), num_exec_masks(num), needs(needs), has_divergent_break(breaks),
  78              has_divergent_continue(cont), has_discard(discard) {}
  79 };
  80
  81 struct block_info {
  82    std::vector<std::pair<Temp, uint8_t>> exec;
  83    std::vector<WQMState> instr_needs;
  84    uint8_t block_needs;
  85    uint8_t ever_again_needs;
  86    /* more... */
  87 };
  88
  89 struct exec_ctx {
  90    Program *program;
  91    std::vector<block_info> info;
  92    std::vector<loop_info> loop;
  93    bool handle_wqm = false;
  94    exec_ctx(Program *program) : program(program), info(program->blocks.size()) {}
  95 };
  96
  97 bool pred_by_exec_mask(aco_ptr<Instruction>& instr) {
  98    if (instr->format == Format::SMEM || instr->isSALU())
  99       return false;
 100    if (instr->format == Format::PSEUDO_BARRIER)
 101       return false;
 102
 103    if (instr->format == Format::PSEUDO) {
 104       switch (instr->opcode) {
 105       case aco_opcode::p_create_vector:
 106          return instr->definitions[0].getTemp().type() == RegType::vgpr;
 107       case aco_opcode::p_extract_vector:
 108       case aco_opcode::p_split_vector:
 109          return instr->operands[0].getTemp().type() == RegType::vgpr;
 110       case aco_opcode::p_spill:
 111       case aco_opcode::p_reload:
 112          return false;
 113       default:
 114          break;
 115       }
 116    }
 117
 118    if (instr->opcode == aco_opcode::v_readlane_b32 ||
 119        instr->opcode == aco_opcode::v_writelane_b32)
 120       return false;
 121
 122    return true;
 123 }
 124
 125 bool needs_exact(aco_ptr<Instruction>& instr) {
 126    if (instr->format == Format::MUBUF) {
 127       MUBUF_instruction *mubuf = static_cast<MUBUF_instruction *>(instr.get());
 128       return mubuf->disable_wqm;
 129    } else if (instr->format == Format::MTBUF) {
 130       MTBUF_instruction *mtbuf = static_cast<MTBUF_instruction *>(instr.get());
 131       return mtbuf->disable_wqm;
 132    } else if (instr->format == Format::MIMG) {
 133       MIMG_instruction *mimg = static_cast<MIMG_instruction *>(instr.get());
 134       return mimg->disable_wqm;
 135    } else {
 136       return instr->format == Format::EXP || instr->opcode == aco_opcode::p_fs_buffer_store_smem;
 137    }
 138 }
 139
 140 void set_needs_wqm(wqm_ctx &ctx, Temp tmp)
 141 {
 142    if (!ctx.needs_wqm[tmp.id()]) {
 143       ctx.needs_wqm[tmp.id()] = true;
 144       if (ctx.defined_in[tmp.id()] != 0xFFFF)
 145          ctx.worklist.insert(ctx.defined_in[tmp.id()]);
 146    }
 147 }
 148
 149 void mark_block_wqm(wqm_ctx &ctx, unsigned block_idx)
 150 {
 151    if (ctx.branch_wqm[block_idx])
 152       return;
 153
 154    ctx.branch_wqm[block_idx] = true;
 155    Block& block = ctx.program->blocks[block_idx];
 156    aco_ptr<Instruction>& branch = block.instructions.back();
 157
 158    if (branch->opcode != aco_opcode::p_branch) {
 159       assert(!branch->operands.empty() && branch->operands[0].isTemp());
 160       set_needs_wqm(ctx, branch->operands[0].getTemp());
 161    }
 162
 163    /* TODO: this sets more branch conditions to WQM than it needs to
 164     * it should be enough to stop at the "exec mask top level" */
 165    if (block.kind & block_kind_top_level)
 166       return;
 167
 168    for (unsigned pred_idx : block.logical_preds)
 169       mark_block_wqm(ctx, pred_idx);
 170 }
 171
 172 void get_block_needs(wqm_ctx &ctx, exec_ctx &exec_ctx, Block* block)
 173 {
 174    block_info& info = exec_ctx.info[block->index];
 175
 176    std::vector<WQMState> instr_needs(block->instructions.size());
 177
 178    if (block->kind & block_kind_top_level) {
 179       if (ctx.loop && ctx.wqm) {
 180          /* mark all break conditions as WQM */
 181          unsigned block_idx = block->index + 1;
 182          while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level)) {
 183             if (ctx.program->blocks[block_idx].kind & block_kind_break)
 184                mark_block_wqm(ctx, block_idx);
 185             block_idx++;
 186          }
 187       } else if (ctx.loop && !ctx.wqm) {
 188          /* Ensure a branch never results in an exec mask with only helper
 189           * invocations (which can cause a loop to repeat infinitively if it's
 190           * break branches are done in exact). */
 191          unsigned block_idx = block->index;
 192          do {
 193             if ((ctx.program->blocks[block_idx].kind & block_kind_branch))
 194                exec_ctx.info[block_idx].block_needs |= Exact_Branch;
 195             block_idx++;
 196          } while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level));
 197       }
 198
 199       ctx.loop = false;
 200       ctx.wqm = false;
 201    }
 202
 203    for (int i = block->instructions.size() - 1; i >= 0; --i)
 204    {
 205       aco_ptr<Instruction>& instr = block->instructions[i];
 206
 207       WQMState needs = needs_exact(instr) ? Exact : Unspecified;
 208       bool propagate_wqm = instr->opcode == aco_opcode::p_wqm;
 209       bool preserve_wqm = instr->opcode == aco_opcode::p_discard_if;
 210       bool pred_by_exec = pred_by_exec_mask(instr);
 211       for (const Definition& definition : instr->definitions) {
 212          if (!definition.isTemp())
 213             continue;
 214          const unsigned def = definition.tempId();
 215          ctx.defined_in[def] = block->index;
 216          if (needs == Unspecified && ctx.needs_wqm[def]) {
 217             needs = pred_by_exec ? WQM : Unspecified;
 218             propagate_wqm = true;
 219          }
 220       }
 221
 222       if (propagate_wqm) {
 223          for (const Operand& op : instr->operands) {
 224             if (op.isTemp()) {
 225                set_needs_wqm(ctx, op.getTemp());
 226             }
 227          }
 228       } else if (preserve_wqm && info.block_needs & WQM) {
 229          needs = Preserve_WQM;
 230       }
 231
 232       /* ensure the condition controlling the control flow for this phi is in WQM */
 233       if (needs == WQM && instr->opcode == aco_opcode::p_phi) {
 234          for (unsigned pred_idx : block->logical_preds)
 235             mark_block_wqm(ctx, pred_idx);
 236       }
 237
 238       instr_needs[i] = needs;
 239       info.block_needs |= needs;
 240    }
 241
 242    info.instr_needs = instr_needs;
 243
 244    /* for "if (<cond>) <wqm code>" or "while (<cond>) <wqm code>",
 245     * <cond> should be computed in WQM */
 246    if (info.block_needs & WQM && !(block->kind & block_kind_top_level)) {
 247       for (unsigned pred_idx : block->logical_preds)
 248          mark_block_wqm(ctx, pred_idx);
 249       ctx.wqm = true;
 250    }
 251    if (block->kind & block_kind_loop_header)
 252       ctx.loop = true;
 253 }
 254
 255 void calculate_wqm_needs(exec_ctx& exec_ctx)
 256 {
 257    wqm_ctx ctx(exec_ctx.program);
 258
 259    while (!ctx.worklist.empty()) {
 260       unsigned block_index = *std::prev(ctx.worklist.end());
 261       ctx.worklist.erase(std::prev(ctx.worklist.end()));
 262
 263       get_block_needs(ctx, exec_ctx, &exec_ctx.program->blocks[block_index]);
 264    }
 265
 266    uint8_t ever_again_needs = 0;
 267    for (int i = exec_ctx.program->blocks.size() - 1; i >= 0; i--) {
 268       exec_ctx.info[i].ever_again_needs = ever_again_needs;
 269       Block& block = exec_ctx.program->blocks[i];
 270
 271       if (block.kind & block_kind_needs_lowering)
 272          exec_ctx.info[i].block_needs |= Exact;
 273
 274       /* if discard is used somewhere in nested CF, we need to preserve the WQM mask */
 275       if ((block.kind & block_kind_discard ||
 276            block.kind & block_kind_uses_discard_if) &&
 277           ever_again_needs & WQM)
 278          exec_ctx.info[i].block_needs |= Preserve_WQM;
 279
 280       ever_again_needs |= exec_ctx.info[i].block_needs & ~Exact_Branch;
 281       if (block.kind & block_kind_discard ||
 282           block.kind & block_kind_uses_discard_if)
 283          ever_again_needs |= Exact;
 284
 285       /* don't propagate WQM preservation further than the next top_level block */
 286       if (block.kind & block_kind_top_level)
 287          ever_again_needs &= ~Preserve_WQM;
 288       else
 289          exec_ctx.info[i].block_needs &= ~Preserve_WQM;
 290    }
 291    exec_ctx.handle_wqm = true;
 292 }
 293
 294 void transition_to_WQM(exec_ctx& ctx, Builder bld, unsigned idx)
 295 {
 296    if (ctx.info[idx].exec.back().second & mask_type_wqm)
 297       return;
 298    if (ctx.info[idx].exec.back().second & mask_type_global) {
 299       Temp exec_mask = ctx.info[idx].exec.back().first;
 300       exec_mask = bld.sop1(aco_opcode::s_wqm_b64, bld.def(s2, exec), bld.def(s1, scc), exec_mask);
 301       ctx.info[idx].exec.emplace_back(exec_mask, mask_type_global | mask_type_wqm);
 302       return;
 303    }
 304    /* otherwise, the WQM mask should be one below the current mask */
 305    ctx.info[idx].exec.pop_back();
 306    assert(ctx.info[idx].exec.back().second & mask_type_wqm);
 307    ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
 308                                                 ctx.info[idx].exec.back().first);
 309 }
 310
 311 void transition_to_Exact(exec_ctx& ctx, Builder bld, unsigned idx)
 312 {
 313    if (ctx.info[idx].exec.back().second & mask_type_exact)
 314       return;
 315    if (ctx.info[idx].exec.back().second & mask_type_global) {
 316       ctx.info[idx].exec.pop_back();
 317       assert(ctx.info[idx].exec.back().second & mask_type_exact);
 318       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
 319                                                    ctx.info[idx].exec.back().first);
 320       return;
 321    }
 322    /* otherwise, we create an exact mask and push to the stack */
 323    Temp wqm = ctx.info[idx].exec.back().first;
 324    Temp exact = bld.tmp(s2);
 325    wqm = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
 326                   bld.exec(Definition(exact)), ctx.info[idx].exec[0].first, bld.exec(wqm));
 327    ctx.info[idx].exec.back().first = wqm;
 328    ctx.info[idx].exec.emplace_back(exact, mask_type_exact);
 329 }
 330
 331 unsigned add_coupling_code(exec_ctx& ctx, Block* block,
 332                            std::vector<aco_ptr<Instruction>>& instructions)
 333 {
 334    unsigned idx = block->index;
 335    Builder bld(ctx.program, &instructions);
 336    std::vector<unsigned>& preds = block->linear_preds;
 337
 338    /* start block */
 339    if (idx == 0) {
 340       aco_ptr<Instruction>& startpgm = block->instructions[0];
 341       assert(startpgm->opcode == aco_opcode::p_startpgm);
 342       Temp exec_mask = startpgm->definitions.back().getTemp();
 343       bld.insert(std::move(startpgm));
 344
 345       if (ctx.handle_wqm) {
 346          ctx.info[0].exec.emplace_back(exec_mask, mask_type_global | mask_type_exact | mask_type_initial);
 347          /* if this block only needs WQM, initialize already */
 348          if (ctx.info[0].block_needs == WQM)
 349             transition_to_WQM(ctx, bld, 0);
 350       } else {
 351          uint8_t mask = mask_type_global;
 352          if (ctx.program->needs_wqm) {
 353             exec_mask = bld.sop1(aco_opcode::s_wqm_b64, bld.def(s2, exec), bld.def(s1, scc), bld.exec(exec_mask));
 354             mask |= mask_type_wqm;
 355          } else {
 356             mask |= mask_type_exact;
 357          }
 358          ctx.info[0].exec.emplace_back(exec_mask, mask);
 359       }
 360
 361       return 1;
 362    }
 363
 364    /* loop entry block */
 365    if (block->kind & block_kind_loop_header) {
 366       assert(preds[0] == idx - 1);
 367       ctx.info[idx].exec = ctx.info[idx - 1].exec;
 368       loop_info& info = ctx.loop.back();
 369       while (ctx.info[idx].exec.size() > info.num_exec_masks)
 370          ctx.info[idx].exec.pop_back();
 371
 372       /* create ssa names for outer exec masks */
 373       if (info.has_discard) {
 374          aco_ptr<Pseudo_instruction> phi;
 375          for (int i = 0; i < info.num_exec_masks - 1; i++) {
 376             phi.reset(create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1));
 377             phi->definitions[0] = bld.def(s2);
 378             phi->operands[0] = Operand(ctx.info[preds[0]].exec[i].first);
 379             ctx.info[idx].exec[i].first = bld.insert(std::move(phi));
 380          }
 381       }
 382
 383       /* create ssa name for restore mask */
 384       if (info.has_divergent_break) {
 385          /* this phi might be trivial but ensures a parallelcopy on the loop header */
 386          aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 387          phi->definitions[0] = bld.def(s2);
 388          phi->operands[0] = Operand(ctx.info[preds[0]].exec[info.num_exec_masks - 1].first);
 389          ctx.info[idx].exec.back().first = bld.insert(std::move(phi));
 390       }
 391
 392       /* create ssa name for loop active mask */
 393       aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 394       if (info.has_divergent_continue)
 395          phi->definitions[0] = bld.def(s2);
 396       else
 397          phi->definitions[0] = bld.def(s2, exec);
 398       phi->operands[0] = Operand(ctx.info[preds[0]].exec.back().first);
 399       Temp loop_active = bld.insert(std::move(phi));
 400
 401       if (info.has_divergent_break) {
 402          uint8_t mask_type = (ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact)) | mask_type_loop;
 403          ctx.info[idx].exec.emplace_back(loop_active, mask_type);
 404       } else {
 405          ctx.info[idx].exec.back().first = loop_active;
 406          ctx.info[idx].exec.back().second |= mask_type_loop;
 407       }
 408
 409       /* create a parallelcopy to move the active mask to exec */
 410       unsigned i = 0;
 411       if (info.has_divergent_continue) {
 412          while (block->instructions[i]->opcode != aco_opcode::p_logical_start) {
 413             bld.insert(std::move(block->instructions[i]));
 414             i++;
 415          }
 416          uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
 417          ctx.info[idx].exec.emplace_back(bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
 418                                                     ctx.info[idx].exec.back().first), mask_type);
 419       }
 420
 421       return i;
 422    }
 423
 424    /* loop exit block */
 425    if (block->kind & block_kind_loop_exit) {
 426       Block* header = ctx.loop.back().loop_header;
 427       loop_info& info = ctx.loop.back();
 428
 429       for (ASSERTED unsigned pred : preds)
 430          assert(ctx.info[pred].exec.size() >= info.num_exec_masks);
 431
 432       /* fill the loop header phis */
 433       std::vector<unsigned>& header_preds = header->linear_preds;
 434       int k = 0;
 435       if (info.has_discard) {
 436          while (k < info.num_exec_masks - 1) {
 437             aco_ptr<Instruction>& phi = header->instructions[k];
 438             assert(phi->opcode == aco_opcode::p_linear_phi);
 439             for (unsigned i = 1; i < phi->operands.size(); i++)
 440                phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[k].first);
 441             k++;
 442          }
 443       }
 444       aco_ptr<Instruction>& phi = header->instructions[k++];
 445       assert(phi->opcode == aco_opcode::p_linear_phi);
 446       for (unsigned i = 1; i < phi->operands.size(); i++)
 447          phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[info.num_exec_masks - 1].first);
 448
 449       if (info.has_divergent_break) {
 450          aco_ptr<Instruction>& phi = header->instructions[k];
 451          assert(phi->opcode == aco_opcode::p_linear_phi);
 452          for (unsigned i = 1; i < phi->operands.size(); i++)
 453             phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[info.num_exec_masks].first);
 454       }
 455
 456       assert(!(block->kind & block_kind_top_level) || info.num_exec_masks <= 2);
 457
 458       /* create the loop exit phis if not trivial */
 459       for (unsigned k = 0; k < info.num_exec_masks; k++) {
 460          Temp same = ctx.info[preds[0]].exec[k].first;
 461          uint8_t type = ctx.info[header_preds[0]].exec[k].second;
 462          bool trivial = true;
 463
 464          for (unsigned i = 1; i < preds.size() && trivial; i++) {
 465             if (ctx.info[preds[i]].exec[k].first != same)
 466                trivial = false;
 467          }
 468
 469          if (trivial) {
 470             ctx.info[idx].exec.emplace_back(same, type);
 471          } else {
 472             /* create phi for loop footer */
 473             aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 474             phi->definitions[0] = bld.def(s2);
 475             for (unsigned i = 0; i < phi->operands.size(); i++)
 476                phi->operands[i] = Operand(ctx.info[preds[i]].exec[k].first);
 477             ctx.info[idx].exec.emplace_back(bld.insert(std::move(phi)), type);
 478          }
 479       }
 480       assert(ctx.info[idx].exec.size() == info.num_exec_masks);
 481
 482       /* create a parallelcopy to move the live mask to exec */
 483       unsigned i = 0;
 484       while (block->instructions[i]->opcode != aco_opcode::p_logical_start) {
 485          bld.insert(std::move(block->instructions[i]));
 486          i++;
 487       }
 488
 489       if (ctx.handle_wqm) {
 490          if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 2) {
 491             if ((ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == 0 ||
 492                 (ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == Exact) {
 493                ctx.info[idx].exec.back().second |= mask_type_global;
 494                transition_to_Exact(ctx, bld, idx);
 495                ctx.handle_wqm = false;
 496             }
 497          }
 498          if (ctx.info[idx].block_needs == WQM)
 499             transition_to_WQM(ctx, bld, idx);
 500          else if (ctx.info[idx].block_needs == Exact)
 501             transition_to_Exact(ctx, bld, idx);
 502       }
 503
 504       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec),
 505                                                    ctx.info[idx].exec.back().first);
 506
 507       ctx.loop.pop_back();
 508       return i;
 509    }
 510
 511    if (preds.size() == 1) {
 512       ctx.info[idx].exec = ctx.info[preds[0]].exec;
 513    } else {
 514       assert(preds.size() == 2);
 515       /* if one of the predecessors ends in exact mask, we pop it from stack */
 516       unsigned num_exec_masks = std::min(ctx.info[preds[0]].exec.size(),
 517                                          ctx.info[preds[1]].exec.size());
 518       if (block->kind & block_kind_top_level && !(block->kind & block_kind_merge))
 519          num_exec_masks = std::min(num_exec_masks, 2u);
 520
 521       /* create phis for diverged exec masks */
 522       for (unsigned i = 0; i < num_exec_masks; i++) {
 523          bool in_exec = i == num_exec_masks - 1 && !(block->kind & block_kind_merge);
 524          if (!in_exec && ctx.info[preds[0]].exec[i].first == ctx.info[preds[1]].exec[i].first) {
 525             assert(ctx.info[preds[0]].exec[i].second == ctx.info[preds[1]].exec[i].second);
 526             ctx.info[idx].exec.emplace_back(ctx.info[preds[0]].exec[i]);
 527             continue;
 528          }
 529
 530          Temp phi = bld.pseudo(aco_opcode::p_linear_phi, in_exec ? bld.def(s2, exec) : bld.def(s2),
 531                                ctx.info[preds[0]].exec[i].first,
 532                                ctx.info[preds[1]].exec[i].first);
 533          uint8_t mask_type = ctx.info[preds[0]].exec[i].second & ctx.info[preds[1]].exec[i].second;
 534          ctx.info[idx].exec.emplace_back(phi, mask_type);
 535       }
 536    }
 537
 538    unsigned i = 0;
 539    while (block->instructions[i]->opcode == aco_opcode::p_phi ||
 540           block->instructions[i]->opcode == aco_opcode::p_linear_phi) {
 541       bld.insert(std::move(block->instructions[i]));
 542       i++;
 543    }
 544
 545    if (block->kind & block_kind_merge)
 546       ctx.info[idx].exec.pop_back();
 547
 548    if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 3) {
 549       assert(ctx.info[idx].exec.back().second == mask_type_exact);
 550       assert(block->kind & block_kind_merge);
 551       ctx.info[idx].exec.pop_back();
 552    }
 553
 554    /* try to satisfy the block's needs */
 555    if (ctx.handle_wqm) {
 556       if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 2) {
 557          if ((ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == 0 ||
 558              (ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == Exact) {
 559             ctx.info[idx].exec.back().second |= mask_type_global;
 560             transition_to_Exact(ctx, bld, idx);
 561             ctx.handle_wqm = false;
 562          }
 563       }
 564       if (ctx.info[idx].block_needs == WQM)
 565          transition_to_WQM(ctx, bld, idx);
 566       else if (ctx.info[idx].block_needs == Exact)
 567          transition_to_Exact(ctx, bld, idx);
 568    }
 569
 570    if (block->kind & block_kind_merge) {
 571       Temp restore = ctx.info[idx].exec.back().first;
 572       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(s2, exec), restore);
 573    }
 574
 575    return i;
 576 }
 577
 578 void lower_fs_buffer_store_smem(Builder& bld, bool need_check, aco_ptr<Instruction>& instr, Temp cur_exec)
 579 {
 580    Operand offset = instr->operands[1];
 581    if (need_check) {
 582       /* if exec is zero, then use UINT32_MAX as an offset and make this store a no-op */
 583       Temp nonempty = bld.sopc(aco_opcode::s_cmp_lg_u64, bld.def(s1, scc), cur_exec, Operand(0u));
 584
 585       if (offset.isLiteral())
 586          offset = bld.sop1(aco_opcode::s_mov_b32, bld.def(s1), offset);
 587
 588       offset = bld.sop2(aco_opcode::s_cselect_b32, bld.hint_m0(bld.def(s1)),
 589                         offset, Operand(UINT32_MAX), bld.scc(nonempty));
 590    } else if (offset.isConstant() && offset.constantValue() > 0xFFFFF) {
 591       offset = bld.sop1(aco_opcode::s_mov_b32, bld.hint_m0(bld.def(s1)), offset);
 592    }
 593    if (!offset.isConstant())
 594       offset.setFixed(m0);
 595
 596    switch (instr->operands[2].size()) {
 597    case 1:
 598       instr->opcode = aco_opcode::s_buffer_store_dword;
 599       break;
 600    case 2:
 601       instr->opcode = aco_opcode::s_buffer_store_dwordx2;
 602       break;
 603    case 4:
 604       instr->opcode = aco_opcode::s_buffer_store_dwordx4;
 605       break;
 606    default:
 607       unreachable("Invalid SMEM buffer store size");
 608    }
 609    instr->operands[1] = offset;
 610    /* as_uniform() needs to be done here so it's done in exact mode and helper
 611     * lanes don't contribute. */
 612    instr->operands[2] = Operand(bld.as_uniform(instr->operands[2]));
 613 }
 614
 615 void process_instructions(exec_ctx& ctx, Block* block,
 616                           std::vector<aco_ptr<Instruction>>& instructions,
 617                           unsigned idx)
 618 {
 619    WQMState state;
 620    if (ctx.info[block->index].exec.back().second & mask_type_wqm)
 621       state = WQM;
 622    else {
 623       assert(!ctx.handle_wqm || ctx.info[block->index].exec.back().second & mask_type_exact);
 624       state = Exact;
 625    }
 626
 627    /* if the block doesn't need both, WQM and Exact, we can skip processing the instructions */
 628    bool process = (ctx.handle_wqm &&
 629                    (ctx.info[block->index].block_needs & state) !=
 630                    (ctx.info[block->index].block_needs & (WQM | Exact))) ||
 631                   block->kind & block_kind_uses_discard_if ||
 632                   block->kind & block_kind_needs_lowering;
 633    if (!process) {
 634       std::vector<aco_ptr<Instruction>>::iterator it = std::next(block->instructions.begin(), idx);
 635       instructions.insert(instructions.end(),
 636                           std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(it),
 637                           std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
 638       return;
 639    }
 640
 641    Builder bld(ctx.program, &instructions);
 642
 643    for (; idx < block->instructions.size(); idx++) {
 644       aco_ptr<Instruction> instr = std::move(block->instructions[idx]);
 645
 646       WQMState needs = ctx.handle_wqm ? ctx.info[block->index].instr_needs[idx] : Unspecified;
 647
 648       if (instr->opcode == aco_opcode::p_discard_if) {
 649          if (ctx.info[block->index].block_needs & Preserve_WQM) {
 650             assert(block->kind & block_kind_top_level);
 651             transition_to_WQM(ctx, bld, block->index);
 652             ctx.info[block->index].exec.back().second &= ~mask_type_global;
 653          }
 654          unsigned num = ctx.info[block->index].exec.size();
 655          assert(num);
 656          Operand cond = instr->operands[0];
 657          instr.reset(create_instruction<Pseudo_instruction>(aco_opcode::p_discard_if, Format::PSEUDO, num + 1, num + 1));
 658          for (unsigned i = 0; i < num; i++) {
 659             instr->operands[i] = Operand(ctx.info[block->index].exec[i].first);
 660             if (i == num - 1)
 661                instr->operands[i].setFixed(exec);
 662             Temp new_mask = bld.tmp(s2);
 663             instr->definitions[i] = Definition(new_mask);
 664             ctx.info[block->index].exec[i].first = new_mask;
 665          }
 666          assert((ctx.info[block->index].exec[0].second & mask_type_wqm) == 0);
 667          instr->definitions[num - 1].setFixed(exec);
 668          instr->operands[num] = cond;
 669          instr->definitions[num] = bld.def(s1, scc);
 670
 671       } else if (needs == WQM && state != WQM) {
 672          transition_to_WQM(ctx, bld, block->index);
 673          state = WQM;
 674       } else if (needs == Exact && state != Exact) {
 675          transition_to_Exact(ctx, bld, block->index);
 676          state = Exact;
 677       }
 678
 679       if (instr->opcode == aco_opcode::p_is_helper || instr->opcode == aco_opcode::p_load_helper) {
 680          Definition dst = instr->definitions[0];
 681          if (state == Exact) {
 682             instr.reset(create_instruction<SOP1_instruction>(aco_opcode::s_mov_b64, Format::SOP1, 1, 1));
 683             instr->operands[0] = Operand(0u);
 684             instr->definitions[0] = dst;
 685          } else {
 686             std::pair<Temp, uint8_t>& exact_mask = ctx.info[block->index].exec[0];
 687             if (instr->opcode == aco_opcode::p_load_helper &&
 688                 !(ctx.info[block->index].exec[0].second & mask_type_initial)) {
 689                /* find last initial exact mask */
 690                for (int i = block->index; i >= 0; i--) {
 691                   if (ctx.program->blocks[i].kind & block_kind_top_level &&
 692                       ctx.info[i].exec[0].second & mask_type_initial) {
 693                      exact_mask = ctx.info[i].exec[0];
 694                      break;
 695                   }
 696                }
 697             }
 698
 699             assert(instr->opcode == aco_opcode::p_is_helper || exact_mask.second & mask_type_initial);
 700             assert(exact_mask.second & mask_type_exact);
 701
 702             instr.reset(create_instruction<SOP2_instruction>(aco_opcode::s_andn2_b64, Format::SOP2, 2, 2));
 703             instr->operands[0] = Operand(ctx.info[block->index].exec.back().first); /* current exec */
 704             instr->operands[1] = Operand(exact_mask.first);
 705             instr->definitions[0] = dst;
 706             instr->definitions[1] = bld.def(s1, scc);
 707          }
 708       } else if (instr->opcode == aco_opcode::p_demote_to_helper) {
 709          /* turn demote into discard_if with only exact masks */
 710          assert((ctx.info[block->index].exec[0].second & (mask_type_exact | mask_type_global)) == (mask_type_exact | mask_type_global));
 711          ctx.info[block->index].exec[0].second &= ~mask_type_initial;
 712
 713          int num = 0;
 714          Temp cond;
 715          if (instr->operands.empty()) {
 716             /* transition to exact and set exec to zero */
 717             Temp old_exec = ctx.info[block->index].exec.back().first;
 718             Temp new_exec = bld.tmp(s2);
 719             cond = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
 720                             bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
 721             if (ctx.info[block->index].exec.back().second & mask_type_exact) {
 722                ctx.info[block->index].exec.back().first = new_exec;
 723             } else {
 724                ctx.info[block->index].exec.back().first = cond;
 725                ctx.info[block->index].exec.emplace_back(new_exec, mask_type_exact);
 726             }
 727          } else {
 728             /* demote_if: transition to exact */
 729             transition_to_Exact(ctx, bld, block->index);
 730             assert(instr->operands[0].isTemp());
 731             cond = instr->operands[0].getTemp();
 732             num = 1;
 733          }
 734
 735          for (unsigned i = 0; i < ctx.info[block->index].exec.size() - 1; i++)
 736             num += ctx.info[block->index].exec[i].second & mask_type_exact ? 1 : 0;
 737          instr.reset(create_instruction<Instruction>(aco_opcode::p_discard_if, Format::PSEUDO, num + 1, num + 1));
 738          int k = 0;
 739          for (unsigned i = 0; k < num; i++) {
 740             if (ctx.info[block->index].exec[i].second & mask_type_exact) {
 741                instr->operands[k] = Operand(ctx.info[block->index].exec[i].first);
 742                Temp new_mask = bld.tmp(s2);
 743                instr->definitions[k] = Definition(new_mask);
 744                if (i == ctx.info[block->index].exec.size() - 1)
 745                   instr->definitions[k].setFixed(exec);
 746                k++;
 747                ctx.info[block->index].exec[i].first = new_mask;
 748             }
 749          }
 750          assert(k == num);
 751          instr->definitions[num] = bld.def(s1, scc);
 752          instr->operands[num] = Operand(cond);
 753          state = Exact;
 754
 755       } else if (instr->opcode == aco_opcode::p_fs_buffer_store_smem) {
 756          bool need_check = ctx.info[block->index].exec.size() != 1 &&
 757                            !(ctx.info[block->index].exec[ctx.info[block->index].exec.size() - 2].second & Exact);
 758          lower_fs_buffer_store_smem(bld, need_check, instr, ctx.info[block->index].exec.back().first);
 759       }
 760
 761       bld.insert(std::move(instr));
 762    }
 763 }
 764
 765 void add_branch_code(exec_ctx& ctx, Block* block)
 766 {
 767    unsigned idx = block->index;
 768    Builder bld(ctx.program, block);
 769
 770    if (idx == ctx.program->blocks.size() - 1)
 771       return;
 772
 773    /* try to disable wqm handling */
 774    if (ctx.handle_wqm && block->kind & block_kind_top_level) {
 775       if (ctx.info[idx].exec.size() == 3) {
 776          assert(ctx.info[idx].exec[1].second == mask_type_wqm);
 777          ctx.info[idx].exec.pop_back();
 778       }
 779       assert(ctx.info[idx].exec.size() <= 2);
 780
 781       if (ctx.info[idx].ever_again_needs == 0 ||
 782           ctx.info[idx].ever_again_needs == Exact) {
 783          /* transition to Exact */
 784          aco_ptr<Instruction> branch = std::move(block->instructions.back());
 785          block->instructions.pop_back();
 786          ctx.info[idx].exec.back().second |= mask_type_global;
 787          transition_to_Exact(ctx, bld, idx);
 788          bld.insert(std::move(branch));
 789          ctx.handle_wqm = false;
 790
 791       } else if (ctx.info[idx].block_needs & Preserve_WQM) {
 792          /* transition to WQM and remove global flag */
 793          aco_ptr<Instruction> branch = std::move(block->instructions.back());
 794          block->instructions.pop_back();
 795          transition_to_WQM(ctx, bld, idx);
 796          ctx.info[idx].exec.back().second &= ~mask_type_global;
 797          bld.insert(std::move(branch));
 798       }
 799    }
 800
 801    if (block->kind & block_kind_loop_preheader) {
 802       /* collect information about the succeeding loop */
 803       bool has_divergent_break = false;
 804       bool has_divergent_continue = false;
 805       bool has_discard = false;
 806       uint8_t needs = 0;
 807       unsigned loop_nest_depth = ctx.program->blocks[idx + 1].loop_nest_depth;
 808
 809       for (unsigned i = idx + 1; ctx.program->blocks[i].loop_nest_depth >= loop_nest_depth; i++) {
 810          Block& loop_block = ctx.program->blocks[i];
 811          needs |= ctx.info[i].block_needs;
 812
 813          if (loop_block.kind & block_kind_uses_discard_if ||
 814              loop_block.kind & block_kind_discard)
 815             has_discard = true;
 816          if (loop_block.loop_nest_depth != loop_nest_depth)
 817             continue;
 818
 819          if (loop_block.kind & block_kind_uniform)
 820             continue;
 821          else if (loop_block.kind & block_kind_break)
 822             has_divergent_break = true;
 823          else if (loop_block.kind & block_kind_continue)
 824             has_divergent_continue = true;
 825       }
 826
 827       if (ctx.handle_wqm) {
 828          if (needs & WQM) {
 829             aco_ptr<Instruction> branch = std::move(block->instructions.back());
 830             block->instructions.pop_back();
 831             transition_to_WQM(ctx, bld, idx);
 832             bld.insert(std::move(branch));
 833          } else {
 834             aco_ptr<Instruction> branch = std::move(block->instructions.back());
 835             block->instructions.pop_back();
 836             transition_to_Exact(ctx, bld, idx);
 837             bld.insert(std::move(branch));
 838          }
 839       }
 840
 841       unsigned num_exec_masks = ctx.info[idx].exec.size();
 842       if (block->kind & block_kind_top_level)
 843          num_exec_masks = std::min(num_exec_masks, 2u);
 844
 845       ctx.loop.emplace_back(&ctx.program->blocks[block->linear_succs[0]],
 846                             num_exec_masks,
 847                             needs,
 848                             has_divergent_break,
 849                             has_divergent_continue,
 850                             has_discard);
 851    }
 852
 853    if (block->kind & block_kind_discard) {
 854
 855       assert(block->instructions.back()->format == Format::PSEUDO_BRANCH);
 856       aco_ptr<Instruction> branch = std::move(block->instructions.back());
 857       block->instructions.pop_back();
 858
 859       /* create a discard_if() instruction with the exec mask as condition */
 860       unsigned num = 0;
 861       if (ctx.loop.size()) {
 862          /* if we're in a loop, only discard from the outer exec masks */
 863          num = ctx.loop.back().num_exec_masks;
 864       } else {
 865          num = ctx.info[idx].exec.size() - 1;
 866       }
 867
 868       Temp old_exec = ctx.info[idx].exec.back().first;
 869       Temp new_exec = bld.tmp(s2);
 870       Temp cond = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
 871                            bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
 872       ctx.info[idx].exec.back().first = new_exec;
 873
 874       aco_ptr<Pseudo_instruction> discard{create_instruction<Pseudo_instruction>(aco_opcode::p_discard_if, Format::PSEUDO, num + 1, num + 1)};
 875       for (unsigned i = 0; i < num; i++) {
 876          discard->operands[i] = Operand(ctx.info[block->index].exec[i].first);
 877          Temp new_mask = bld.tmp(s2);
 878          discard->definitions[i] = Definition(new_mask);
 879          ctx.info[block->index].exec[i].first = new_mask;
 880       }
 881       assert(!ctx.handle_wqm || (ctx.info[block->index].exec[0].second & mask_type_wqm) == 0);
 882       discard->operands[num] = Operand(cond);
 883       discard->definitions[num] = bld.def(s1, scc);
 884
 885       bld.insert(std::move(discard));
 886       if ((block->kind & (block_kind_break | block_kind_uniform)) == block_kind_break)
 887          ctx.info[idx].exec.back().first = cond;
 888       bld.insert(std::move(branch));
 889       /* no return here as it can be followed by a divergent break */
 890    }
 891
 892    if (block->kind & block_kind_continue_or_break) {
 893       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
 894       block->instructions.pop_back();
 895
 896       /* because of how linear_succs is created, this needs to be swapped */
 897       std::swap(block->linear_succs[0], block->linear_succs[1]);
 898
 899       assert(ctx.program->blocks[block->linear_succs[1]].kind & block_kind_loop_header);
 900       assert(ctx.program->blocks[ctx.program->blocks[block->linear_succs[0]].linear_succs[0]].kind & block_kind_loop_exit);
 901
 902       if (ctx.info[idx].exec.back().second & mask_type_loop) {
 903          bld.branch(aco_opcode::p_cbranch_nz, bld.exec(ctx.info[idx].exec.back().first), block->linear_succs[1], block->linear_succs[0]);
 904       } else {
 905          Temp cond = Temp();
 906          for (int exec_idx = ctx.info[idx].exec.size() - 1; exec_idx >= 0; exec_idx--) {
 907             if (ctx.info[idx].exec[exec_idx].second & mask_type_loop) {
 908                cond = bld.sopc(aco_opcode::s_cmp_lg_u64, bld.def(s1, scc), ctx.info[idx].exec[exec_idx].first, Operand(0u));
 909                break;
 910             }
 911          }
 912          assert(cond != Temp());
 913
 914          bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
 915       }
 916       return;
 917    }
 918
 919    if (block->kind & block_kind_uniform) {
 920       Pseudo_branch_instruction* branch = static_cast<Pseudo_branch_instruction*>(block->instructions.back().get());
 921       if (branch->opcode == aco_opcode::p_branch) {
 922          branch->target[0] = block->linear_succs[0];
 923       } else {
 924          branch->target[0] = block->linear_succs[1];
 925          branch->target[1] = block->linear_succs[0];
 926       }
 927       return;
 928    }
 929
 930    if (block->kind & block_kind_branch) {
 931
 932       if (ctx.handle_wqm &&
 933           ctx.info[idx].exec.size() >= 2 &&
 934           ctx.info[idx].exec.back().second == mask_type_exact &&
 935           !(ctx.info[idx].block_needs & Exact_Branch) &&
 936           ctx.info[idx].exec[ctx.info[idx].exec.size() - 2].second & mask_type_wqm) {
 937          /* return to wqm before branching */
 938          ctx.info[idx].exec.pop_back();
 939       }
 940
 941       // orig = s_and_saveexec_b64
 942       assert(block->linear_succs.size() == 2);
 943       assert(block->instructions.back()->opcode == aco_opcode::p_cbranch_z);
 944       Temp cond = block->instructions.back()->operands[0].getTemp();
 945       block->instructions.pop_back();
 946
 947       if (ctx.info[idx].block_needs & Exact_Branch)
 948          transition_to_Exact(ctx, bld, idx);
 949
 950       Temp current_exec = ctx.info[idx].exec.back().first;
 951       uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
 952
 953       Temp then_mask = bld.tmp(s2);
 954       Temp old_exec = bld.sop1(aco_opcode::s_and_saveexec_b64, bld.def(s2), bld.def(s1, scc),
 955                                bld.exec(Definition(then_mask)), cond, bld.exec(current_exec));
 956
 957       ctx.info[idx].exec.back().first = old_exec;
 958
 959       /* add next current exec to the stack */
 960       ctx.info[idx].exec.emplace_back(then_mask, mask_type);
 961
 962       bld.branch(aco_opcode::p_cbranch_z, bld.exec(then_mask), block->linear_succs[1], block->linear_succs[0]);
 963       return;
 964    }
 965
 966    if (block->kind & block_kind_invert) {
 967       // exec = s_andn2_b64 (original_exec, exec)
 968       assert(block->instructions.back()->opcode == aco_opcode::p_cbranch_nz);
 969       block->instructions.pop_back();
 970       Temp then_mask = ctx.info[idx].exec.back().first;
 971       uint8_t mask_type = ctx.info[idx].exec.back().second;
 972       ctx.info[idx].exec.pop_back();
 973       Temp orig_exec = ctx.info[idx].exec.back().first;
 974       Temp else_mask = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2, exec),
 975                                 bld.def(s1, scc), orig_exec, bld.exec(then_mask));
 976
 977       /* add next current exec to the stack */
 978       ctx.info[idx].exec.emplace_back(else_mask, mask_type);
 979
 980       bld.branch(aco_opcode::p_cbranch_z, bld.exec(else_mask), block->linear_succs[1], block->linear_succs[0]);
 981       return;
 982    }
 983
 984    if (block->kind & block_kind_break) {
 985       // loop_mask = s_andn2_b64 (loop_mask, exec)
 986       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
 987       block->instructions.pop_back();
 988
 989       Temp current_exec = ctx.info[idx].exec.back().first;
 990       Temp cond = Temp();
 991       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
 992          cond = bld.tmp(s1);
 993          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
 994          exec_mask = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.scc(Definition(cond)),
 995                               exec_mask, current_exec);
 996          ctx.info[idx].exec[exec_idx].first = exec_mask;
 997          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
 998             break;
 999       }
1000
1001       /* check if the successor is the merge block, otherwise set exec to 0 */
1002       // TODO: this could be done better by directly branching to the merge block
1003       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
1004       Block& succ = ctx.program->blocks[succ_idx];
1005       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
1006          ctx.info[idx].exec.back().first = bld.sop1(aco_opcode::s_mov_b64, bld.def(s2, exec), Operand(0u));
1007       }
1008
1009       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
1010       return;
1011    }
1012
1013    if (block->kind & block_kind_continue) {
1014       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
1015       block->instructions.pop_back();
1016
1017       Temp current_exec = ctx.info[idx].exec.back().first;
1018       Temp cond = Temp();
1019       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
1020          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
1021             break;
1022          cond = bld.tmp(s1);
1023          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
1024          exec_mask = bld.sop2(aco_opcode::s_andn2_b64, bld.def(s2), bld.scc(Definition(cond)),
1025                               exec_mask, bld.exec(current_exec));
1026          ctx.info[idx].exec[exec_idx].first = exec_mask;
1027       }
1028       assert(cond != Temp());
1029
1030       /* check if the successor is the merge block, otherwise set exec to 0 */
1031       // TODO: this could be done better by directly branching to the merge block
1032       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
1033       Block& succ = ctx.program->blocks[succ_idx];
1034       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
1035          ctx.info[idx].exec.back().first = bld.sop1(aco_opcode::s_mov_b64, bld.def(s2, exec), Operand(0u));
1036       }
1037
1038       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
1039       return;
1040    }
1041 }
1042
1043 void process_block(exec_ctx& ctx, Block* block)
1044 {
1045    std::vector<aco_ptr<Instruction>> instructions;
1046    instructions.reserve(block->instructions.size());
1047
1048    unsigned idx = add_coupling_code(ctx, block, instructions);
1049
1050    assert(block->index != ctx.program->blocks.size() - 1 ||
1051           ctx.info[block->index].exec.size() <= 2);
1052
1053    process_instructions(ctx, block, instructions, idx);
1054
1055    block->instructions = std::move(instructions);
1056
1057    add_branch_code(ctx, block);
1058
1059    block->live_out_exec = ctx.info[block->index].exec.back().first;
1060 }
1061
1062 } /* end namespace */
1063
1064
1065 void insert_exec_mask(Program *program)
1066 {
1067    exec_ctx ctx(program);
1068
1069    if (program->needs_wqm && program->needs_exact)
1070       calculate_wqm_needs(ctx);
1071
1072    for (Block& block : program->blocks)
1073       process_block(ctx, &block);
1074
1075 }
1076
1077 }
1078