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 #include "util/u_math.h"
  28
  29 namespace aco {
  30
  31 namespace {
  32
  33 enum WQMState : uint8_t {
  34    Unspecified = 0,
  35    Exact = 1 << 0,
  36    WQM = 1 << 1, /* with control flow applied */
  37    Preserve_WQM = 1 << 2,
  38    Exact_Branch = 1 << 3,
  39 };
  40
  41 enum mask_type : uint8_t {
  42    mask_type_global = 1 << 0,
  43    mask_type_exact = 1 << 1,
  44    mask_type_wqm = 1 << 2,
  45    mask_type_loop = 1 << 3, /* active lanes of a loop */
  46    mask_type_initial = 1 << 4, /* initially active lanes */
  47 };
  48
  49 struct wqm_ctx {
  50    Program* program;
  51    /* state for WQM propagation */
  52    std::set<unsigned> worklist;
  53    std::vector<uint16_t> defined_in;
  54    std::vector<bool> needs_wqm;
  55    std::vector<bool> branch_wqm; /* true if the branch condition in this block should be in wqm */
  56    bool loop;
  57    bool wqm;
  58    wqm_ctx(Program* program) : program(program),
  59                                defined_in(program->peekAllocationId(), 0xFFFF),
  60                                needs_wqm(program->peekAllocationId()),
  61                                branch_wqm(program->blocks.size()),
  62                                loop(false),
  63                                wqm(false)
  64    {
  65       for (unsigned i = 0; i < program->blocks.size(); i++)
  66          worklist.insert(i);
  67    }
  68 };
  69
  70 struct loop_info {
  71    Block* loop_header;
  72    uint16_t num_exec_masks;
  73    uint8_t needs;
  74    bool has_divergent_break;
  75    bool has_divergent_continue;
  76    bool has_discard; /* has a discard or demote */
  77    loop_info(Block* b, uint16_t num, uint8_t needs, bool breaks, bool cont, bool discard) :
  78              loop_header(b), num_exec_masks(num), needs(needs), has_divergent_break(breaks),
  79              has_divergent_continue(cont), has_discard(discard) {}
  80 };
  81
  82 struct block_info {
  83    std::vector<std::pair<Temp, uint8_t>> exec;
  84    std::vector<WQMState> instr_needs;
  85    uint8_t block_needs;
  86    uint8_t ever_again_needs;
  87    bool logical_end_wqm;
  88    /* more... */
  89 };
  90
  91 struct exec_ctx {
  92    Program *program;
  93    std::vector<block_info> info;
  94    std::vector<loop_info> loop;
  95    bool handle_wqm = false;
  96    exec_ctx(Program *program) : program(program), info(program->blocks.size()) {}
  97 };
  98
  99 bool pred_by_exec_mask(aco_ptr<Instruction>& instr) {
 100    if (instr->isSALU())
 101       return instr->reads_exec();
 102    if (instr->format == Format::SMEM || instr->isSALU())
 103       return false;
 104    if (instr->format == Format::PSEUDO_BARRIER)
 105       return false;
 106
 107    if (instr->format == Format::PSEUDO) {
 108       switch (instr->opcode) {
 109       case aco_opcode::p_create_vector:
 110          return instr->definitions[0].getTemp().type() == RegType::vgpr;
 111       case aco_opcode::p_extract_vector:
 112       case aco_opcode::p_split_vector:
 113          return instr->operands[0].getTemp().type() == RegType::vgpr;
 114       case aco_opcode::p_spill:
 115       case aco_opcode::p_reload:
 116          return false;
 117       default:
 118          break;
 119       }
 120    }
 121
 122    if (instr->opcode == aco_opcode::v_readlane_b32 ||
 123        instr->opcode == aco_opcode::v_readlane_b32_e64 ||
 124        instr->opcode == aco_opcode::v_writelane_b32 ||
 125        instr->opcode == aco_opcode::v_writelane_b32_e64)
 126       return false;
 127
 128    return true;
 129 }
 130
 131 bool needs_exact(aco_ptr<Instruction>& instr) {
 132    if (instr->format == Format::MUBUF) {
 133       MUBUF_instruction *mubuf = static_cast<MUBUF_instruction *>(instr.get());
 134       return mubuf->disable_wqm;
 135    } else if (instr->format == Format::MTBUF) {
 136       MTBUF_instruction *mtbuf = static_cast<MTBUF_instruction *>(instr.get());
 137       return mtbuf->disable_wqm;
 138    } else if (instr->format == Format::MIMG) {
 139       MIMG_instruction *mimg = static_cast<MIMG_instruction *>(instr.get());
 140       return mimg->disable_wqm;
 141    } else if (instr->format == Format::FLAT || instr->format == Format::GLOBAL) {
 142       FLAT_instruction *flat = static_cast<FLAT_instruction *>(instr.get());
 143       return flat->disable_wqm;
 144    } else {
 145       return instr->format == Format::EXP || instr->opcode == aco_opcode::p_fs_buffer_store_smem;
 146    }
 147 }
 148
 149 void set_needs_wqm(wqm_ctx &ctx, Temp tmp)
 150 {
 151    if (!ctx.needs_wqm[tmp.id()]) {
 152       ctx.needs_wqm[tmp.id()] = true;
 153       if (ctx.defined_in[tmp.id()] != 0xFFFF)
 154          ctx.worklist.insert(ctx.defined_in[tmp.id()]);
 155    }
 156 }
 157
 158 void mark_block_wqm(wqm_ctx &ctx, unsigned block_idx)
 159 {
 160    if (ctx.branch_wqm[block_idx])
 161       return;
 162
 163    ctx.branch_wqm[block_idx] = true;
 164    Block& block = ctx.program->blocks[block_idx];
 165    aco_ptr<Instruction>& branch = block.instructions.back();
 166
 167    if (branch->opcode != aco_opcode::p_branch) {
 168       assert(!branch->operands.empty() && branch->operands[0].isTemp());
 169       set_needs_wqm(ctx, branch->operands[0].getTemp());
 170    }
 171
 172    /* TODO: this sets more branch conditions to WQM than it needs to
 173     * it should be enough to stop at the "exec mask top level" */
 174    if (block.kind & block_kind_top_level)
 175       return;
 176
 177    for (unsigned pred_idx : block.logical_preds)
 178       mark_block_wqm(ctx, pred_idx);
 179 }
 180
 181 void get_block_needs(wqm_ctx &ctx, exec_ctx &exec_ctx, Block* block)
 182 {
 183    block_info& info = exec_ctx.info[block->index];
 184
 185    std::vector<WQMState> instr_needs(block->instructions.size());
 186
 187    if (block->kind & block_kind_top_level) {
 188       if (ctx.loop && ctx.wqm) {
 189          /* mark all break conditions as WQM */
 190          unsigned block_idx = block->index + 1;
 191          while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level)) {
 192             if (ctx.program->blocks[block_idx].kind & block_kind_break)
 193                mark_block_wqm(ctx, block_idx);
 194             block_idx++;
 195          }
 196       } else if (ctx.loop && !ctx.wqm) {
 197          /* Ensure a branch never results in an exec mask with only helper
 198           * invocations (which can cause a loop to repeat infinitively if it's
 199           * break branches are done in exact). */
 200          unsigned block_idx = block->index;
 201          do {
 202             if ((ctx.program->blocks[block_idx].kind & block_kind_branch))
 203                exec_ctx.info[block_idx].block_needs |= Exact_Branch;
 204             block_idx++;
 205          } while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level));
 206       }
 207
 208       ctx.loop = false;
 209       ctx.wqm = false;
 210    }
 211
 212    for (int i = block->instructions.size() - 1; i >= 0; --i) {
 213       aco_ptr<Instruction>& instr = block->instructions[i];
 214
 215       WQMState needs = needs_exact(instr) ? Exact : Unspecified;
 216       bool propagate_wqm = instr->opcode == aco_opcode::p_wqm;
 217       bool preserve_wqm = instr->opcode == aco_opcode::p_discard_if;
 218       bool pred_by_exec = pred_by_exec_mask(instr);
 219       for (const Definition& definition : instr->definitions) {
 220          if (!definition.isTemp())
 221             continue;
 222          const unsigned def = definition.tempId();
 223          ctx.defined_in[def] = block->index;
 224          if (needs == Unspecified && ctx.needs_wqm[def]) {
 225             needs = pred_by_exec ? WQM : Unspecified;
 226             propagate_wqm = true;
 227          }
 228       }
 229
 230       if (propagate_wqm) {
 231          for (const Operand& op : instr->operands) {
 232             if (op.isTemp()) {
 233                set_needs_wqm(ctx, op.getTemp());
 234             }
 235          }
 236       } else if (preserve_wqm && info.block_needs & WQM) {
 237          needs = Preserve_WQM;
 238       }
 239
 240       /* ensure the condition controlling the control flow for this phi is in WQM */
 241       if (needs == WQM && instr->opcode == aco_opcode::p_phi) {
 242          for (unsigned pred_idx : block->logical_preds) {
 243             mark_block_wqm(ctx, pred_idx);
 244             exec_ctx.info[pred_idx].logical_end_wqm = true;
 245             ctx.worklist.insert(pred_idx);
 246          }
 247       }
 248
 249       if (instr->opcode == aco_opcode::p_logical_end && info.logical_end_wqm) {
 250          assert(needs == Unspecified);
 251          needs = WQM;
 252       }
 253
 254       instr_needs[i] = needs;
 255       info.block_needs |= needs;
 256    }
 257
 258    info.instr_needs = instr_needs;
 259
 260    /* for "if (<cond>) <wqm code>" or "while (<cond>) <wqm code>",
 261     * <cond> should be computed in WQM */
 262    if (info.block_needs & WQM && !(block->kind & block_kind_top_level)) {
 263       for (unsigned pred_idx : block->logical_preds)
 264          mark_block_wqm(ctx, pred_idx);
 265       ctx.wqm = true;
 266    }
 267    if (block->kind & block_kind_loop_header)
 268       ctx.loop = true;
 269 }
 270
 271 void calculate_wqm_needs(exec_ctx& exec_ctx)
 272 {
 273    wqm_ctx ctx(exec_ctx.program);
 274
 275    while (!ctx.worklist.empty()) {
 276       unsigned block_index = *std::prev(ctx.worklist.end());
 277       ctx.worklist.erase(std::prev(ctx.worklist.end()));
 278
 279       get_block_needs(ctx, exec_ctx, &exec_ctx.program->blocks[block_index]);
 280    }
 281
 282    uint8_t ever_again_needs = 0;
 283    for (int i = exec_ctx.program->blocks.size() - 1; i >= 0; i--) {
 284       exec_ctx.info[i].ever_again_needs = ever_again_needs;
 285       Block& block = exec_ctx.program->blocks[i];
 286
 287       if (block.kind & block_kind_needs_lowering)
 288          exec_ctx.info[i].block_needs |= Exact;
 289
 290       /* if discard is used somewhere in nested CF, we need to preserve the WQM mask */
 291       if ((block.kind & block_kind_discard ||
 292            block.kind & block_kind_uses_discard_if) &&
 293           ever_again_needs & WQM)
 294          exec_ctx.info[i].block_needs |= Preserve_WQM;
 295
 296       ever_again_needs |= exec_ctx.info[i].block_needs & ~Exact_Branch;
 297       if (block.kind & block_kind_discard ||
 298           block.kind & block_kind_uses_discard_if ||
 299           block.kind & block_kind_uses_demote)
 300          ever_again_needs |= Exact;
 301
 302       /* don't propagate WQM preservation further than the next top_level block */
 303       if (block.kind & block_kind_top_level)
 304          ever_again_needs &= ~Preserve_WQM;
 305       else
 306          exec_ctx.info[i].block_needs &= ~Preserve_WQM;
 307    }
 308    exec_ctx.handle_wqm = true;
 309 }
 310
 311 void transition_to_WQM(exec_ctx& ctx, Builder bld, unsigned idx)
 312 {
 313    if (ctx.info[idx].exec.back().second & mask_type_wqm)
 314       return;
 315    if (ctx.info[idx].exec.back().second & mask_type_global) {
 316       Temp exec_mask = ctx.info[idx].exec.back().first;
 317       exec_mask = bld.sop1(Builder::s_wqm, bld.def(bld.lm, exec), bld.def(s1, scc), exec_mask);
 318       ctx.info[idx].exec.emplace_back(exec_mask, mask_type_global | mask_type_wqm);
 319       return;
 320    }
 321    /* otherwise, the WQM mask should be one below the current mask */
 322    ctx.info[idx].exec.pop_back();
 323    assert(ctx.info[idx].exec.back().second & mask_type_wqm);
 324    assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 325    ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 326                                                 ctx.info[idx].exec.back().first);
 327 }
 328
 329 void transition_to_Exact(exec_ctx& ctx, Builder bld, unsigned idx)
 330 {
 331    if (ctx.info[idx].exec.back().second & mask_type_exact)
 332       return;
 333    /* We can't remove the loop exec mask, because that can cause exec.size() to
 334     * be less than num_exec_masks. The loop exec mask also needs to be kept
 335     * around for various uses. */
 336    if ((ctx.info[idx].exec.back().second & mask_type_global) &&
 337        !(ctx.info[idx].exec.back().second & mask_type_loop)) {
 338       ctx.info[idx].exec.pop_back();
 339       assert(ctx.info[idx].exec.back().second & mask_type_exact);
 340       assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 341       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 342                                                    ctx.info[idx].exec.back().first);
 343       return;
 344    }
 345    /* otherwise, we create an exact mask and push to the stack */
 346    Temp wqm = ctx.info[idx].exec.back().first;
 347    Temp exact = bld.tmp(bld.lm);
 348    wqm = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
 349                   bld.exec(Definition(exact)), ctx.info[idx].exec[0].first, bld.exec(wqm));
 350    ctx.info[idx].exec.back().first = wqm;
 351    ctx.info[idx].exec.emplace_back(exact, mask_type_exact);
 352 }
 353
 354 unsigned add_coupling_code(exec_ctx& ctx, Block* block,
 355                            std::vector<aco_ptr<Instruction>>& instructions)
 356 {
 357    unsigned idx = block->index;
 358    Builder bld(ctx.program, &instructions);
 359    std::vector<unsigned>& preds = block->linear_preds;
 360
 361    /* start block */
 362    if (idx == 0) {
 363       aco_ptr<Instruction>& startpgm = block->instructions[0];
 364       assert(startpgm->opcode == aco_opcode::p_startpgm);
 365       Temp exec_mask = startpgm->definitions.back().getTemp();
 366       bld.insert(std::move(startpgm));
 367
 368       /* exec seems to need to be manually initialized with combined shaders */
 369       if (util_bitcount(ctx.program->stage & sw_mask) > 1) {
 370          bld.sop1(Builder::s_mov, bld.exec(Definition(exec_mask)), bld.lm == s2 ? Operand(UINT64_MAX) : Operand(UINT32_MAX));
 371          instructions[0]->definitions.pop_back();
 372       }
 373
 374       if (ctx.handle_wqm) {
 375          ctx.info[0].exec.emplace_back(exec_mask, mask_type_global | mask_type_exact | mask_type_initial);
 376          /* if this block only needs WQM, initialize already */
 377          if (ctx.info[0].block_needs == WQM)
 378             transition_to_WQM(ctx, bld, 0);
 379       } else {
 380          uint8_t mask = mask_type_global;
 381          if (ctx.program->needs_wqm) {
 382             exec_mask = bld.sop1(Builder::s_wqm, bld.def(bld.lm, exec), bld.def(s1, scc), bld.exec(exec_mask));
 383             mask |= mask_type_wqm;
 384          } else {
 385             mask |= mask_type_exact;
 386          }
 387          ctx.info[0].exec.emplace_back(exec_mask, mask);
 388       }
 389
 390       return 1;
 391    }
 392
 393    /* loop entry block */
 394    if (block->kind & block_kind_loop_header) {
 395       assert(preds[0] == idx - 1);
 396       ctx.info[idx].exec = ctx.info[idx - 1].exec;
 397       loop_info& info = ctx.loop.back();
 398       while (ctx.info[idx].exec.size() > info.num_exec_masks)
 399          ctx.info[idx].exec.pop_back();
 400
 401       /* create ssa names for outer exec masks */
 402       if (info.has_discard) {
 403          aco_ptr<Pseudo_instruction> phi;
 404          for (int i = 0; i < info.num_exec_masks - 1; i++) {
 405             phi.reset(create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1));
 406             phi->definitions[0] = bld.def(bld.lm);
 407             phi->operands[0] = Operand(ctx.info[preds[0]].exec[i].first);
 408             ctx.info[idx].exec[i].first = bld.insert(std::move(phi));
 409          }
 410       }
 411
 412       /* create ssa name for restore mask */
 413       if (info.has_divergent_break) {
 414          /* this phi might be trivial but ensures a parallelcopy on the loop header */
 415          aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 416          phi->definitions[0] = bld.def(bld.lm);
 417          phi->operands[0] = Operand(ctx.info[preds[0]].exec[info.num_exec_masks - 1].first);
 418          ctx.info[idx].exec.back().first = bld.insert(std::move(phi));
 419       }
 420
 421       /* create ssa name for loop active mask */
 422       aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 423       if (info.has_divergent_continue)
 424          phi->definitions[0] = bld.def(bld.lm);
 425       else
 426          phi->definitions[0] = bld.def(bld.lm, exec);
 427       phi->operands[0] = Operand(ctx.info[preds[0]].exec.back().first);
 428       Temp loop_active = bld.insert(std::move(phi));
 429
 430       if (info.has_divergent_break) {
 431          uint8_t mask_type = (ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact)) | mask_type_loop;
 432          ctx.info[idx].exec.emplace_back(loop_active, mask_type);
 433       } else {
 434          ctx.info[idx].exec.back().first = loop_active;
 435          ctx.info[idx].exec.back().second |= mask_type_loop;
 436       }
 437
 438       /* create a parallelcopy to move the active mask to exec */
 439       unsigned i = 0;
 440       if (info.has_divergent_continue) {
 441          while (block->instructions[i]->opcode != aco_opcode::p_logical_start) {
 442             bld.insert(std::move(block->instructions[i]));
 443             i++;
 444          }
 445          uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
 446          assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 447          ctx.info[idx].exec.emplace_back(bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 448                                                     ctx.info[idx].exec.back().first), mask_type);
 449       }
 450
 451       return i;
 452    }
 453
 454    /* loop exit block */
 455    if (block->kind & block_kind_loop_exit) {
 456       Block* header = ctx.loop.back().loop_header;
 457       loop_info& info = ctx.loop.back();
 458
 459       for (ASSERTED unsigned pred : preds)
 460          assert(ctx.info[pred].exec.size() >= info.num_exec_masks);
 461
 462       /* fill the loop header phis */
 463       std::vector<unsigned>& header_preds = header->linear_preds;
 464       int k = 0;
 465       if (info.has_discard) {
 466          while (k < info.num_exec_masks - 1) {
 467             aco_ptr<Instruction>& phi = header->instructions[k];
 468             assert(phi->opcode == aco_opcode::p_linear_phi);
 469             for (unsigned i = 1; i < phi->operands.size(); i++)
 470                phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[k].first);
 471             k++;
 472          }
 473       }
 474       aco_ptr<Instruction>& phi = header->instructions[k++];
 475       assert(phi->opcode == aco_opcode::p_linear_phi);
 476       for (unsigned i = 1; i < phi->operands.size(); i++)
 477          phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[info.num_exec_masks - 1].first);
 478
 479       if (info.has_divergent_break) {
 480          aco_ptr<Instruction>& phi = header->instructions[k];
 481          assert(phi->opcode == aco_opcode::p_linear_phi);
 482          for (unsigned i = 1; i < phi->operands.size(); i++)
 483             phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[info.num_exec_masks].first);
 484       }
 485
 486       assert(!(block->kind & block_kind_top_level) || info.num_exec_masks <= 2);
 487
 488       /* create the loop exit phis if not trivial */
 489       for (unsigned k = 0; k < info.num_exec_masks; k++) {
 490          Temp same = ctx.info[preds[0]].exec[k].first;
 491          uint8_t type = ctx.info[header_preds[0]].exec[k].second;
 492          bool trivial = true;
 493
 494          for (unsigned i = 1; i < preds.size() && trivial; i++) {
 495             if (ctx.info[preds[i]].exec[k].first != same)
 496                trivial = false;
 497          }
 498
 499          if (trivial) {
 500             ctx.info[idx].exec.emplace_back(same, type);
 501          } else {
 502             /* create phi for loop footer */
 503             aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 504             phi->definitions[0] = bld.def(bld.lm);
 505             for (unsigned i = 0; i < phi->operands.size(); i++)
 506                phi->operands[i] = Operand(ctx.info[preds[i]].exec[k].first);
 507             ctx.info[idx].exec.emplace_back(bld.insert(std::move(phi)), type);
 508          }
 509       }
 510       assert(ctx.info[idx].exec.size() == info.num_exec_masks);
 511
 512       /* create a parallelcopy to move the live mask to exec */
 513       unsigned i = 0;
 514       while (block->instructions[i]->opcode != aco_opcode::p_logical_start) {
 515          bld.insert(std::move(block->instructions[i]));
 516          i++;
 517       }
 518
 519       if (ctx.handle_wqm) {
 520          if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 2) {
 521             if ((ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == 0 ||
 522                 (ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == Exact) {
 523                ctx.info[idx].exec.back().second |= mask_type_global;
 524                transition_to_Exact(ctx, bld, idx);
 525                ctx.handle_wqm = false;
 526             }
 527          }
 528          if (ctx.info[idx].block_needs == WQM)
 529             transition_to_WQM(ctx, bld, idx);
 530          else if (ctx.info[idx].block_needs == Exact)
 531             transition_to_Exact(ctx, bld, idx);
 532       }
 533
 534       assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 535       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 536                                                    ctx.info[idx].exec.back().first);
 537
 538       ctx.loop.pop_back();
 539       return i;
 540    }
 541
 542    if (preds.size() == 1) {
 543       ctx.info[idx].exec = ctx.info[preds[0]].exec;
 544    } else {
 545       assert(preds.size() == 2);
 546       /* if one of the predecessors ends in exact mask, we pop it from stack */
 547       unsigned num_exec_masks = std::min(ctx.info[preds[0]].exec.size(),
 548                                          ctx.info[preds[1]].exec.size());
 549       if (block->kind & block_kind_top_level && !(block->kind & block_kind_merge))
 550          num_exec_masks = std::min(num_exec_masks, 2u);
 551
 552       /* create phis for diverged exec masks */
 553       for (unsigned i = 0; i < num_exec_masks; i++) {
 554          bool in_exec = i == num_exec_masks - 1 && !(block->kind & block_kind_merge);
 555          if (!in_exec && ctx.info[preds[0]].exec[i].first == ctx.info[preds[1]].exec[i].first) {
 556             assert(ctx.info[preds[0]].exec[i].second == ctx.info[preds[1]].exec[i].second);
 557             ctx.info[idx].exec.emplace_back(ctx.info[preds[0]].exec[i]);
 558             continue;
 559          }
 560
 561          Temp phi = bld.pseudo(aco_opcode::p_linear_phi, in_exec ? bld.def(bld.lm, exec) : bld.def(bld.lm),
 562                                ctx.info[preds[0]].exec[i].first,
 563                                ctx.info[preds[1]].exec[i].first);
 564          uint8_t mask_type = ctx.info[preds[0]].exec[i].second & ctx.info[preds[1]].exec[i].second;
 565          ctx.info[idx].exec.emplace_back(phi, mask_type);
 566       }
 567    }
 568
 569    unsigned i = 0;
 570    while (block->instructions[i]->opcode == aco_opcode::p_phi ||
 571           block->instructions[i]->opcode == aco_opcode::p_linear_phi) {
 572       bld.insert(std::move(block->instructions[i]));
 573       i++;
 574    }
 575
 576    if (block->kind & block_kind_merge)
 577       ctx.info[idx].exec.pop_back();
 578
 579    if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 3) {
 580       assert(ctx.info[idx].exec.back().second == mask_type_exact);
 581       assert(block->kind & block_kind_merge);
 582       ctx.info[idx].exec.pop_back();
 583    }
 584
 585    /* try to satisfy the block's needs */
 586    if (ctx.handle_wqm) {
 587       if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 2) {
 588          if ((ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == 0 ||
 589              (ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == Exact) {
 590             ctx.info[idx].exec.back().second |= mask_type_global;
 591             transition_to_Exact(ctx, bld, idx);
 592             ctx.handle_wqm = false;
 593          }
 594       }
 595       if (ctx.info[idx].block_needs == WQM)
 596          transition_to_WQM(ctx, bld, idx);
 597       else if (ctx.info[idx].block_needs == Exact)
 598          transition_to_Exact(ctx, bld, idx);
 599    }
 600
 601    if (block->kind & block_kind_merge) {
 602       Temp restore = ctx.info[idx].exec.back().first;
 603       assert(restore.size() == bld.lm.size());
 604       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), restore);
 605    }
 606
 607    return i;
 608 }
 609
 610 void lower_fs_buffer_store_smem(Builder& bld, bool need_check, aco_ptr<Instruction>& instr, Temp cur_exec)
 611 {
 612    Operand offset = instr->operands[1];
 613    if (need_check) {
 614       /* if exec is zero, then use UINT32_MAX as an offset and make this store a no-op */
 615       Temp nonempty = bld.sopc(Builder::s_cmp_lg, bld.def(s1, scc), cur_exec, Operand(0u));
 616
 617       if (offset.isLiteral())
 618          offset = bld.sop1(aco_opcode::s_mov_b32, bld.def(s1), offset);
 619
 620       offset = bld.sop2(aco_opcode::s_cselect_b32, bld.hint_m0(bld.def(s1)),
 621                         offset, Operand(UINT32_MAX), bld.scc(nonempty));
 622    } else if (offset.isConstant() && offset.constantValue() > 0xFFFFF) {
 623       offset = bld.sop1(aco_opcode::s_mov_b32, bld.hint_m0(bld.def(s1)), offset);
 624    }
 625    if (!offset.isConstant())
 626       offset.setFixed(m0);
 627
 628    switch (instr->operands[2].size()) {
 629    case 1:
 630       instr->opcode = aco_opcode::s_buffer_store_dword;
 631       break;
 632    case 2:
 633       instr->opcode = aco_opcode::s_buffer_store_dwordx2;
 634       break;
 635    case 4:
 636       instr->opcode = aco_opcode::s_buffer_store_dwordx4;
 637       break;
 638    default:
 639       unreachable("Invalid SMEM buffer store size");
 640    }
 641    instr->operands[1] = offset;
 642    /* as_uniform() needs to be done here so it's done in exact mode and helper
 643     * lanes don't contribute. */
 644    instr->operands[2] = Operand(bld.as_uniform(instr->operands[2]));
 645 }
 646
 647 void process_instructions(exec_ctx& ctx, Block* block,
 648                           std::vector<aco_ptr<Instruction>>& instructions,
 649                           unsigned idx)
 650 {
 651    WQMState state;
 652    if (ctx.info[block->index].exec.back().second & mask_type_wqm)
 653       state = WQM;
 654    else {
 655       assert(!ctx.handle_wqm || ctx.info[block->index].exec.back().second & mask_type_exact);
 656       state = Exact;
 657    }
 658
 659    /* if the block doesn't need both, WQM and Exact, we can skip processing the instructions */
 660    bool process = (ctx.handle_wqm &&
 661                    (ctx.info[block->index].block_needs & state) !=
 662                    (ctx.info[block->index].block_needs & (WQM | Exact))) ||
 663                   block->kind & block_kind_uses_discard_if ||
 664                   block->kind & block_kind_uses_demote ||
 665                   block->kind & block_kind_needs_lowering;
 666    if (!process) {
 667       std::vector<aco_ptr<Instruction>>::iterator it = std::next(block->instructions.begin(), idx);
 668       instructions.insert(instructions.end(),
 669                           std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(it),
 670                           std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
 671       return;
 672    }
 673
 674    Builder bld(ctx.program, &instructions);
 675
 676    for (; idx < block->instructions.size(); idx++) {
 677       aco_ptr<Instruction> instr = std::move(block->instructions[idx]);
 678
 679       WQMState needs = ctx.handle_wqm ? ctx.info[block->index].instr_needs[idx] : Unspecified;
 680
 681       if (instr->opcode == aco_opcode::p_discard_if) {
 682          if (ctx.info[block->index].block_needs & Preserve_WQM) {
 683             assert(block->kind & block_kind_top_level);
 684             transition_to_WQM(ctx, bld, block->index);
 685             ctx.info[block->index].exec.back().second &= ~mask_type_global;
 686          }
 687          int num = ctx.info[block->index].exec.size();
 688          assert(num);
 689          Operand cond = instr->operands[0];
 690          for (int i = num - 1; i >= 0; i--) {
 691             Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
 692                                           ctx.info[block->index].exec[i].first, cond);
 693             if (i == num - 1) {
 694                andn2->operands[0].setFixed(exec);
 695                andn2->definitions[0].setFixed(exec);
 696             }
 697             if (i == 0) {
 698                instr->opcode = aco_opcode::p_exit_early_if;
 699                instr->operands[0] = bld.scc(andn2->definitions[1].getTemp());
 700             }
 701             ctx.info[block->index].exec[i].first = andn2->definitions[0].getTemp();
 702          }
 703          assert(!ctx.handle_wqm || (ctx.info[block->index].exec[0].second & mask_type_wqm) == 0);
 704
 705       } else if (needs == WQM && state != WQM) {
 706          transition_to_WQM(ctx, bld, block->index);
 707          state = WQM;
 708       } else if (needs == Exact && state != Exact) {
 709          transition_to_Exact(ctx, bld, block->index);
 710          state = Exact;
 711       }
 712
 713       if (instr->opcode == aco_opcode::p_is_helper || instr->opcode == aco_opcode::p_load_helper) {
 714          Definition dst = instr->definitions[0];
 715          assert(dst.size() == bld.lm.size());
 716          if (state == Exact) {
 717             instr.reset(create_instruction<SOP1_instruction>(bld.w64or32(Builder::s_mov), Format::SOP1, 1, 1));
 718             instr->operands[0] = Operand(0u);
 719             instr->definitions[0] = dst;
 720          } else {
 721             std::pair<Temp, uint8_t>& exact_mask = ctx.info[block->index].exec[0];
 722             if (instr->opcode == aco_opcode::p_load_helper &&
 723                 !(ctx.info[block->index].exec[0].second & mask_type_initial)) {
 724                /* find last initial exact mask */
 725                for (int i = block->index; i >= 0; i--) {
 726                   if (ctx.program->blocks[i].kind & block_kind_top_level &&
 727                       ctx.info[i].exec[0].second & mask_type_initial) {
 728                      exact_mask = ctx.info[i].exec[0];
 729                      break;
 730                   }
 731                }
 732             }
 733
 734             assert(instr->opcode == aco_opcode::p_is_helper || exact_mask.second & mask_type_initial);
 735             assert(exact_mask.second & mask_type_exact);
 736
 737             instr.reset(create_instruction<SOP2_instruction>(bld.w64or32(Builder::s_andn2), Format::SOP2, 2, 2));
 738             instr->operands[0] = Operand(ctx.info[block->index].exec.back().first); /* current exec */
 739             instr->operands[1] = Operand(exact_mask.first);
 740             instr->definitions[0] = dst;
 741             instr->definitions[1] = bld.def(s1, scc);
 742          }
 743       } else if (instr->opcode == aco_opcode::p_demote_to_helper) {
 744          /* turn demote into discard_if with only exact masks */
 745          assert((ctx.info[block->index].exec[0].second & (mask_type_exact | mask_type_global)) == (mask_type_exact | mask_type_global));
 746          ctx.info[block->index].exec[0].second &= ~mask_type_initial;
 747
 748          int num;
 749          Temp cond, exit_cond;
 750          if (instr->operands[0].isConstant()) {
 751             assert(instr->operands[0].constantValue() == -1u);
 752             /* transition to exact and set exec to zero */
 753             Temp old_exec = ctx.info[block->index].exec.back().first;
 754             Temp new_exec = bld.tmp(bld.lm);
 755             exit_cond = bld.tmp(s1);
 756             cond = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.scc(Definition(exit_cond)),
 757                             bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
 758
 759             num = ctx.info[block->index].exec.size() - 2;
 760             if (ctx.info[block->index].exec.back().second & mask_type_exact) {
 761                ctx.info[block->index].exec.back().first = new_exec;
 762             } else {
 763                ctx.info[block->index].exec.back().first = cond;
 764                ctx.info[block->index].exec.emplace_back(new_exec, mask_type_exact);
 765             }
 766          } else {
 767             /* demote_if: transition to exact */
 768             transition_to_Exact(ctx, bld, block->index);
 769             assert(instr->operands[0].isTemp());
 770             cond = instr->operands[0].getTemp();
 771             num = ctx.info[block->index].exec.size() - 1;
 772          }
 773
 774          for (int i = num; i >= 0; i--) {
 775             if (ctx.info[block->index].exec[i].second & mask_type_exact) {
 776                Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
 777                                              ctx.info[block->index].exec[i].first, cond);
 778                if (i == (int)ctx.info[block->index].exec.size() - 1) {
 779                   andn2->operands[0].setFixed(exec);
 780                   andn2->definitions[0].setFixed(exec);
 781                }
 782
 783                ctx.info[block->index].exec[i].first = andn2->definitions[0].getTemp();
 784                exit_cond = andn2->definitions[1].getTemp();
 785             } else {
 786                assert(i != 0);
 787             }
 788          }
 789          instr->opcode = aco_opcode::p_exit_early_if;
 790          instr->operands[0] = bld.scc(exit_cond);
 791          state = Exact;
 792
 793       } else if (instr->opcode == aco_opcode::p_fs_buffer_store_smem) {
 794          bool need_check = ctx.info[block->index].exec.size() != 1 &&
 795                            !(ctx.info[block->index].exec[ctx.info[block->index].exec.size() - 2].second & Exact);
 796          lower_fs_buffer_store_smem(bld, need_check, instr, ctx.info[block->index].exec.back().first);
 797       }
 798
 799       bld.insert(std::move(instr));
 800    }
 801 }
 802
 803 void add_branch_code(exec_ctx& ctx, Block* block)
 804 {
 805    unsigned idx = block->index;
 806    Builder bld(ctx.program, block);
 807
 808    if (idx == ctx.program->blocks.size() - 1)
 809       return;
 810
 811    /* try to disable wqm handling */
 812    if (ctx.handle_wqm && block->kind & block_kind_top_level) {
 813       if (ctx.info[idx].exec.size() == 3) {
 814          assert(ctx.info[idx].exec[1].second == mask_type_wqm);
 815          ctx.info[idx].exec.pop_back();
 816       }
 817       assert(ctx.info[idx].exec.size() <= 2);
 818
 819       if (ctx.info[idx].ever_again_needs == 0 ||
 820           ctx.info[idx].ever_again_needs == Exact) {
 821          /* transition to Exact */
 822          aco_ptr<Instruction> branch = std::move(block->instructions.back());
 823          block->instructions.pop_back();
 824          ctx.info[idx].exec.back().second |= mask_type_global;
 825          transition_to_Exact(ctx, bld, idx);
 826          bld.insert(std::move(branch));
 827          ctx.handle_wqm = false;
 828
 829       } else if (ctx.info[idx].block_needs & Preserve_WQM) {
 830          /* transition to WQM and remove global flag */
 831          aco_ptr<Instruction> branch = std::move(block->instructions.back());
 832          block->instructions.pop_back();
 833          transition_to_WQM(ctx, bld, idx);
 834          ctx.info[idx].exec.back().second &= ~mask_type_global;
 835          bld.insert(std::move(branch));
 836       }
 837    }
 838
 839    if (block->kind & block_kind_loop_preheader) {
 840       /* collect information about the succeeding loop */
 841       bool has_divergent_break = false;
 842       bool has_divergent_continue = false;
 843       bool has_discard = false;
 844       uint8_t needs = 0;
 845       unsigned loop_nest_depth = ctx.program->blocks[idx + 1].loop_nest_depth;
 846
 847       for (unsigned i = idx + 1; ctx.program->blocks[i].loop_nest_depth >= loop_nest_depth; i++) {
 848          Block& loop_block = ctx.program->blocks[i];
 849          needs |= ctx.info[i].block_needs;
 850
 851          if (loop_block.kind & block_kind_uses_discard_if ||
 852              loop_block.kind & block_kind_discard ||
 853              loop_block.kind & block_kind_uses_demote)
 854             has_discard = true;
 855          if (loop_block.loop_nest_depth != loop_nest_depth)
 856             continue;
 857
 858          if (loop_block.kind & block_kind_uniform)
 859             continue;
 860          else if (loop_block.kind & block_kind_break)
 861             has_divergent_break = true;
 862          else if (loop_block.kind & block_kind_continue)
 863             has_divergent_continue = true;
 864       }
 865
 866       if (ctx.handle_wqm) {
 867          if (needs & WQM) {
 868             aco_ptr<Instruction> branch = std::move(block->instructions.back());
 869             block->instructions.pop_back();
 870             transition_to_WQM(ctx, bld, idx);
 871             bld.insert(std::move(branch));
 872          } else {
 873             aco_ptr<Instruction> branch = std::move(block->instructions.back());
 874             block->instructions.pop_back();
 875             transition_to_Exact(ctx, bld, idx);
 876             bld.insert(std::move(branch));
 877          }
 878       }
 879
 880       unsigned num_exec_masks = ctx.info[idx].exec.size();
 881       if (block->kind & block_kind_top_level)
 882          num_exec_masks = std::min(num_exec_masks, 2u);
 883
 884       ctx.loop.emplace_back(&ctx.program->blocks[block->linear_succs[0]],
 885                             num_exec_masks,
 886                             needs,
 887                             has_divergent_break,
 888                             has_divergent_continue,
 889                             has_discard);
 890    }
 891
 892    if (block->kind & block_kind_discard) {
 893
 894       assert(block->instructions.back()->format == Format::PSEUDO_BRANCH);
 895       aco_ptr<Instruction> branch = std::move(block->instructions.back());
 896       block->instructions.pop_back();
 897
 898       /* create a discard_if() instruction with the exec mask as condition */
 899       unsigned num = 0;
 900       if (ctx.loop.size()) {
 901          /* if we're in a loop, only discard from the outer exec masks */
 902          num = ctx.loop.back().num_exec_masks;
 903       } else {
 904          num = ctx.info[idx].exec.size() - 1;
 905       }
 906
 907       Temp old_exec = ctx.info[idx].exec.back().first;
 908       Temp new_exec = bld.tmp(bld.lm);
 909       Temp cond = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
 910                            bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
 911       ctx.info[idx].exec.back().first = new_exec;
 912
 913       for (int i = num - 1; i >= 0; i--) {
 914          Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
 915                                        ctx.info[block->index].exec[i].first, cond);
 916          if (i == (int)ctx.info[idx].exec.size() - 1)
 917             andn2->definitions[0].setFixed(exec);
 918          if (i == 0)
 919             bld.pseudo(aco_opcode::p_exit_early_if, bld.scc(andn2->definitions[1].getTemp()));
 920          ctx.info[block->index].exec[i].first = andn2->definitions[0].getTemp();
 921       }
 922       assert(!ctx.handle_wqm || (ctx.info[block->index].exec[0].second & mask_type_wqm) == 0);
 923
 924       if ((block->kind & (block_kind_break | block_kind_uniform)) == block_kind_break)
 925          ctx.info[idx].exec.back().first = cond;
 926       bld.insert(std::move(branch));
 927       /* no return here as it can be followed by a divergent break */
 928    }
 929
 930    if (block->kind & block_kind_continue_or_break) {
 931       assert(ctx.program->blocks[ctx.program->blocks[block->linear_succs[1]].linear_succs[0]].kind & block_kind_loop_header);
 932       assert(ctx.program->blocks[ctx.program->blocks[block->linear_succs[0]].linear_succs[0]].kind & block_kind_loop_exit);
 933       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
 934       block->instructions.pop_back();
 935
 936       while (!(ctx.info[idx].exec.back().second & mask_type_loop))
 937          ctx.info[idx].exec.pop_back();
 938
 939       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), ctx.info[idx].exec.back().first);
 940       bld.branch(aco_opcode::p_cbranch_nz, bld.exec(ctx.info[idx].exec.back().first), block->linear_succs[1], block->linear_succs[0]);
 941       return;
 942    }
 943
 944    if (block->kind & block_kind_uniform) {
 945       Pseudo_branch_instruction* branch = static_cast<Pseudo_branch_instruction*>(block->instructions.back().get());
 946       if (branch->opcode == aco_opcode::p_branch) {
 947          branch->target[0] = block->linear_succs[0];
 948       } else {
 949          branch->target[0] = block->linear_succs[1];
 950          branch->target[1] = block->linear_succs[0];
 951       }
 952       return;
 953    }
 954
 955    if (block->kind & block_kind_branch) {
 956
 957       if (ctx.handle_wqm &&
 958           ctx.info[idx].exec.size() >= 2 &&
 959           ctx.info[idx].exec.back().second == mask_type_exact &&
 960           !(ctx.info[idx].block_needs & Exact_Branch) &&
 961           ctx.info[idx].exec[ctx.info[idx].exec.size() - 2].second & mask_type_wqm) {
 962          /* return to wqm before branching */
 963          ctx.info[idx].exec.pop_back();
 964       }
 965
 966       // orig = s_and_saveexec_b64
 967       assert(block->linear_succs.size() == 2);
 968       assert(block->instructions.back()->opcode == aco_opcode::p_cbranch_z);
 969       Temp cond = block->instructions.back()->operands[0].getTemp();
 970       block->instructions.pop_back();
 971
 972       if (ctx.info[idx].block_needs & Exact_Branch)
 973          transition_to_Exact(ctx, bld, idx);
 974
 975       Temp current_exec = ctx.info[idx].exec.back().first;
 976       uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
 977
 978       Temp then_mask = bld.tmp(bld.lm);
 979       Temp old_exec = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
 980                                bld.exec(Definition(then_mask)), cond, bld.exec(current_exec));
 981
 982       ctx.info[idx].exec.back().first = old_exec;
 983
 984       /* add next current exec to the stack */
 985       ctx.info[idx].exec.emplace_back(then_mask, mask_type);
 986
 987       bld.branch(aco_opcode::p_cbranch_z, bld.exec(then_mask), block->linear_succs[1], block->linear_succs[0]);
 988       return;
 989    }
 990
 991    if (block->kind & block_kind_invert) {
 992       // exec = s_andn2_b64 (original_exec, exec)
 993       assert(block->instructions.back()->opcode == aco_opcode::p_cbranch_nz);
 994       block->instructions.pop_back();
 995       Temp then_mask = ctx.info[idx].exec.back().first;
 996       uint8_t mask_type = ctx.info[idx].exec.back().second;
 997       ctx.info[idx].exec.pop_back();
 998       Temp orig_exec = ctx.info[idx].exec.back().first;
 999       Temp else_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm, exec),
1000                                 bld.def(s1, scc), orig_exec, bld.exec(then_mask));
1001
1002       /* add next current exec to the stack */
1003       ctx.info[idx].exec.emplace_back(else_mask, mask_type);
1004
1005       bld.branch(aco_opcode::p_cbranch_z, bld.exec(else_mask), block->linear_succs[1], block->linear_succs[0]);
1006       return;
1007    }
1008
1009    if (block->kind & block_kind_break) {
1010       // loop_mask = s_andn2_b64 (loop_mask, exec)
1011       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
1012       block->instructions.pop_back();
1013
1014       Temp current_exec = ctx.info[idx].exec.back().first;
1015       Temp cond = Temp();
1016       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
1017          cond = bld.tmp(s1);
1018          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
1019          exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
1020                               exec_mask, current_exec);
1021          ctx.info[idx].exec[exec_idx].first = exec_mask;
1022          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
1023             break;
1024       }
1025
1026       /* check if the successor is the merge block, otherwise set exec to 0 */
1027       // TODO: this could be done better by directly branching to the merge block
1028       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
1029       Block& succ = ctx.program->blocks[succ_idx];
1030       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
1031          ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
1032       }
1033
1034       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
1035       return;
1036    }
1037
1038    if (block->kind & block_kind_continue) {
1039       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
1040       block->instructions.pop_back();
1041
1042       Temp current_exec = ctx.info[idx].exec.back().first;
1043       Temp cond = Temp();
1044       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
1045          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
1046             break;
1047          cond = bld.tmp(s1);
1048          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
1049          exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
1050                               exec_mask, bld.exec(current_exec));
1051          ctx.info[idx].exec[exec_idx].first = exec_mask;
1052       }
1053       assert(cond != Temp());
1054
1055       /* check if the successor is the merge block, otherwise set exec to 0 */
1056       // TODO: this could be done better by directly branching to the merge block
1057       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
1058       Block& succ = ctx.program->blocks[succ_idx];
1059       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
1060          ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
1061       }
1062
1063       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
1064       return;
1065    }
1066 }
1067
1068 void process_block(exec_ctx& ctx, Block* block)
1069 {
1070    std::vector<aco_ptr<Instruction>> instructions;
1071    instructions.reserve(block->instructions.size());
1072
1073    unsigned idx = add_coupling_code(ctx, block, instructions);
1074
1075    assert(block->index != ctx.program->blocks.size() - 1 ||
1076           ctx.info[block->index].exec.size() <= 2);
1077
1078    process_instructions(ctx, block, instructions, idx);
1079
1080    block->instructions = std::move(instructions);
1081
1082    add_branch_code(ctx, block);
1083
1084    block->live_out_exec = ctx.info[block->index].exec.back().first;
1085 }
1086
1087 } /* end namespace */
1088
1089
1090 void insert_exec_mask(Program *program)
1091 {
1092    exec_ctx ctx(program);
1093
1094    if (program->needs_wqm && program->needs_exact)
1095       calculate_wqm_needs(ctx);
1096
1097    for (Block& block : program->blocks)
1098       process_block(ctx, &block);
1099
1100 }
1101
1102 }
1103