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       case aco_opcode::p_extract_vector:
 111       case aco_opcode::p_split_vector:
 112          for (Definition def : instr->definitions) {
 113             if (def.getTemp().type() == RegType::vgpr)
 114                return true;
 115          }
 116          return false;
 117       case aco_opcode::p_spill:
 118       case aco_opcode::p_reload:
 119          return false;
 120       default:
 121          break;
 122       }
 123    }
 124
 125    if (instr->opcode == aco_opcode::v_readlane_b32 ||
 126        instr->opcode == aco_opcode::v_readlane_b32_e64 ||
 127        instr->opcode == aco_opcode::v_writelane_b32 ||
 128        instr->opcode == aco_opcode::v_writelane_b32_e64)
 129       return false;
 130
 131    return true;
 132 }
 133
 134 bool needs_exact(aco_ptr<Instruction>& instr) {
 135    if (instr->format == Format::MUBUF) {
 136       MUBUF_instruction *mubuf = static_cast<MUBUF_instruction *>(instr.get());
 137       return mubuf->disable_wqm;
 138    } else if (instr->format == Format::MTBUF) {
 139       MTBUF_instruction *mtbuf = static_cast<MTBUF_instruction *>(instr.get());
 140       return mtbuf->disable_wqm;
 141    } else if (instr->format == Format::MIMG) {
 142       MIMG_instruction *mimg = static_cast<MIMG_instruction *>(instr.get());
 143       return mimg->disable_wqm;
 144    } else if (instr->format == Format::FLAT || instr->format == Format::GLOBAL) {
 145       FLAT_instruction *flat = static_cast<FLAT_instruction *>(instr.get());
 146       return flat->disable_wqm;
 147    } else {
 148       return instr->format == Format::EXP || instr->opcode == aco_opcode::p_fs_buffer_store_smem;
 149    }
 150 }
 151
 152 void set_needs_wqm(wqm_ctx &ctx, Temp tmp)
 153 {
 154    if (!ctx.needs_wqm[tmp.id()]) {
 155       ctx.needs_wqm[tmp.id()] = true;
 156       if (ctx.defined_in[tmp.id()] != 0xFFFF)
 157          ctx.worklist.insert(ctx.defined_in[tmp.id()]);
 158    }
 159 }
 160
 161 void mark_block_wqm(wqm_ctx &ctx, unsigned block_idx)
 162 {
 163    if (ctx.branch_wqm[block_idx])
 164       return;
 165
 166    ctx.branch_wqm[block_idx] = true;
 167    Block& block = ctx.program->blocks[block_idx];
 168    aco_ptr<Instruction>& branch = block.instructions.back();
 169
 170    if (branch->opcode != aco_opcode::p_branch) {
 171       assert(!branch->operands.empty() && branch->operands[0].isTemp());
 172       set_needs_wqm(ctx, branch->operands[0].getTemp());
 173    }
 174
 175    /* TODO: this sets more branch conditions to WQM than it needs to
 176     * it should be enough to stop at the "exec mask top level" */
 177    if (block.kind & block_kind_top_level)
 178       return;
 179
 180    for (unsigned pred_idx : block.logical_preds)
 181       mark_block_wqm(ctx, pred_idx);
 182 }
 183
 184 void get_block_needs(wqm_ctx &ctx, exec_ctx &exec_ctx, Block* block)
 185 {
 186    block_info& info = exec_ctx.info[block->index];
 187
 188    std::vector<WQMState> instr_needs(block->instructions.size());
 189
 190    if (block->kind & block_kind_top_level) {
 191       if (ctx.loop && ctx.wqm) {
 192          /* mark all break conditions as WQM */
 193          unsigned block_idx = block->index + 1;
 194          while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level)) {
 195             if (ctx.program->blocks[block_idx].kind & block_kind_break)
 196                mark_block_wqm(ctx, block_idx);
 197             block_idx++;
 198          }
 199       } else if (ctx.loop && !ctx.wqm) {
 200          /* Ensure a branch never results in an exec mask with only helper
 201           * invocations (which can cause a loop to repeat infinitively if it's
 202           * break branches are done in exact). */
 203          unsigned block_idx = block->index;
 204          do {
 205             if ((ctx.program->blocks[block_idx].kind & block_kind_branch))
 206                exec_ctx.info[block_idx].block_needs |= Exact_Branch;
 207             block_idx++;
 208          } while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level));
 209       }
 210
 211       ctx.loop = false;
 212       ctx.wqm = false;
 213    }
 214
 215    for (int i = block->instructions.size() - 1; i >= 0; --i) {
 216       aco_ptr<Instruction>& instr = block->instructions[i];
 217
 218       WQMState needs = needs_exact(instr) ? Exact : Unspecified;
 219       bool propagate_wqm = instr->opcode == aco_opcode::p_wqm;
 220       bool preserve_wqm = instr->opcode == aco_opcode::p_discard_if;
 221       bool pred_by_exec = pred_by_exec_mask(instr);
 222       for (const Definition& definition : instr->definitions) {
 223          if (!definition.isTemp())
 224             continue;
 225          const unsigned def = definition.tempId();
 226          ctx.defined_in[def] = block->index;
 227          if (needs == Unspecified && ctx.needs_wqm[def]) {
 228             needs = pred_by_exec ? WQM : Unspecified;
 229             propagate_wqm = true;
 230          }
 231       }
 232
 233       if (propagate_wqm) {
 234          for (const Operand& op : instr->operands) {
 235             if (op.isTemp()) {
 236                set_needs_wqm(ctx, op.getTemp());
 237             }
 238          }
 239       } else if (preserve_wqm && info.block_needs & WQM) {
 240          needs = Preserve_WQM;
 241       }
 242
 243       /* ensure the condition controlling the control flow for this phi is in WQM */
 244       if (needs == WQM && instr->opcode == aco_opcode::p_phi) {
 245          for (unsigned pred_idx : block->logical_preds) {
 246             mark_block_wqm(ctx, pred_idx);
 247             exec_ctx.info[pred_idx].logical_end_wqm = true;
 248             ctx.worklist.insert(pred_idx);
 249          }
 250       }
 251
 252       if ((instr->opcode == aco_opcode::p_logical_end && info.logical_end_wqm) ||
 253           instr->opcode == aco_opcode::p_wqm) {
 254          assert(needs != Exact);
 255          needs = WQM;
 256       }
 257
 258       instr_needs[i] = needs;
 259       info.block_needs |= needs;
 260    }
 261
 262    info.instr_needs = instr_needs;
 263
 264    /* for "if (<cond>) <wqm code>" or "while (<cond>) <wqm code>",
 265     * <cond> should be computed in WQM */
 266    if (info.block_needs & WQM && !(block->kind & block_kind_top_level)) {
 267       for (unsigned pred_idx : block->logical_preds)
 268          mark_block_wqm(ctx, pred_idx);
 269       ctx.wqm = true;
 270    }
 271    if (block->kind & block_kind_loop_header)
 272       ctx.loop = true;
 273 }
 274
 275 void calculate_wqm_needs(exec_ctx& exec_ctx)
 276 {
 277    wqm_ctx ctx(exec_ctx.program);
 278
 279    while (!ctx.worklist.empty()) {
 280       unsigned block_index = *std::prev(ctx.worklist.end());
 281       ctx.worklist.erase(std::prev(ctx.worklist.end()));
 282
 283       get_block_needs(ctx, exec_ctx, &exec_ctx.program->blocks[block_index]);
 284    }
 285
 286    uint8_t ever_again_needs = 0;
 287    for (int i = exec_ctx.program->blocks.size() - 1; i >= 0; i--) {
 288       exec_ctx.info[i].ever_again_needs = ever_again_needs;
 289       Block& block = exec_ctx.program->blocks[i];
 290
 291       if (block.kind & block_kind_needs_lowering)
 292          exec_ctx.info[i].block_needs |= Exact;
 293
 294       /* if discard is used somewhere in nested CF, we need to preserve the WQM mask */
 295       if ((block.kind & block_kind_discard ||
 296            block.kind & block_kind_uses_discard_if) &&
 297           ever_again_needs & WQM)
 298          exec_ctx.info[i].block_needs |= Preserve_WQM;
 299
 300       ever_again_needs |= exec_ctx.info[i].block_needs & ~Exact_Branch;
 301       if (block.kind & block_kind_discard ||
 302           block.kind & block_kind_uses_discard_if ||
 303           block.kind & block_kind_uses_demote)
 304          ever_again_needs |= Exact;
 305
 306       /* don't propagate WQM preservation further than the next top_level block */
 307       if (block.kind & block_kind_top_level)
 308          ever_again_needs &= ~Preserve_WQM;
 309       else
 310          exec_ctx.info[i].block_needs &= ~Preserve_WQM;
 311    }
 312    exec_ctx.handle_wqm = true;
 313 }
 314
 315 void transition_to_WQM(exec_ctx& ctx, Builder bld, unsigned idx)
 316 {
 317    if (ctx.info[idx].exec.back().second & mask_type_wqm)
 318       return;
 319    if (ctx.info[idx].exec.back().second & mask_type_global) {
 320       Temp exec_mask = ctx.info[idx].exec.back().first;
 321       /* TODO: we might generate better code if we pass the uncopied "exec_mask"
 322        * directly to the s_wqm (we still need to keep this parallelcopy for
 323        * potential later uses of exec_mask though). We currently can't do this
 324        * because of a RA bug. */
 325       exec_mask = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm), bld.exec(exec_mask));
 326       ctx.info[idx].exec.back().first = exec_mask;
 327
 328       exec_mask = bld.sop1(Builder::s_wqm, bld.def(bld.lm, exec), bld.def(s1, scc), exec_mask);
 329       ctx.info[idx].exec.emplace_back(exec_mask, mask_type_global | mask_type_wqm);
 330       return;
 331    }
 332    /* otherwise, the WQM mask should be one below the current mask */
 333    ctx.info[idx].exec.pop_back();
 334    assert(ctx.info[idx].exec.back().second & mask_type_wqm);
 335    assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 336    ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 337                                                 ctx.info[idx].exec.back().first);
 338 }
 339
 340 void transition_to_Exact(exec_ctx& ctx, Builder bld, unsigned idx)
 341 {
 342    if (ctx.info[idx].exec.back().second & mask_type_exact)
 343       return;
 344    /* We can't remove the loop exec mask, because that can cause exec.size() to
 345     * be less than num_exec_masks. The loop exec mask also needs to be kept
 346     * around for various uses. */
 347    if ((ctx.info[idx].exec.back().second & mask_type_global) &&
 348        !(ctx.info[idx].exec.back().second & mask_type_loop)) {
 349       ctx.info[idx].exec.pop_back();
 350       assert(ctx.info[idx].exec.back().second & mask_type_exact);
 351       assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 352       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 353                                                    ctx.info[idx].exec.back().first);
 354       return;
 355    }
 356    /* otherwise, we create an exact mask and push to the stack */
 357    Temp wqm = ctx.info[idx].exec.back().first;
 358    Temp exact = bld.tmp(bld.lm);
 359    wqm = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
 360                   bld.exec(Definition(exact)), ctx.info[idx].exec[0].first, bld.exec(wqm));
 361    ctx.info[idx].exec.back().first = wqm;
 362    ctx.info[idx].exec.emplace_back(exact, mask_type_exact);
 363 }
 364
 365 unsigned add_coupling_code(exec_ctx& ctx, Block* block,
 366                            std::vector<aco_ptr<Instruction>>& instructions)
 367 {
 368    unsigned idx = block->index;
 369    Builder bld(ctx.program, &instructions);
 370    std::vector<unsigned>& preds = block->linear_preds;
 371
 372    /* start block */
 373    if (idx == 0) {
 374       aco_ptr<Instruction>& startpgm = block->instructions[0];
 375       assert(startpgm->opcode == aco_opcode::p_startpgm);
 376       Temp exec_mask = startpgm->definitions.back().getTemp();
 377       bld.insert(std::move(startpgm));
 378
 379       /* exec seems to need to be manually initialized with combined shaders */
 380       if (util_bitcount(ctx.program->stage & sw_mask) > 1 || (ctx.program->stage & hw_ngg_gs)) {
 381          bld.sop1(Builder::s_mov, bld.exec(Definition(exec_mask)), bld.lm == s2 ? Operand(UINT64_MAX) : Operand(UINT32_MAX));
 382          instructions[0]->definitions.pop_back();
 383       }
 384
 385       if (ctx.handle_wqm) {
 386          ctx.info[0].exec.emplace_back(exec_mask, mask_type_global | mask_type_exact | mask_type_initial);
 387          /* if this block only needs WQM, initialize already */
 388          if (ctx.info[0].block_needs == WQM)
 389             transition_to_WQM(ctx, bld, 0);
 390       } else {
 391          uint8_t mask = mask_type_global;
 392          if (ctx.program->needs_wqm) {
 393             exec_mask = bld.sop1(Builder::s_wqm, bld.def(bld.lm, exec), bld.def(s1, scc), bld.exec(exec_mask));
 394             mask |= mask_type_wqm;
 395          } else {
 396             mask |= mask_type_exact;
 397          }
 398          ctx.info[0].exec.emplace_back(exec_mask, mask);
 399       }
 400
 401       return 1;
 402    }
 403
 404    /* loop entry block */
 405    if (block->kind & block_kind_loop_header) {
 406       assert(preds[0] == idx - 1);
 407       ctx.info[idx].exec = ctx.info[idx - 1].exec;
 408       loop_info& info = ctx.loop.back();
 409       while (ctx.info[idx].exec.size() > info.num_exec_masks)
 410          ctx.info[idx].exec.pop_back();
 411
 412       /* create ssa names for outer exec masks */
 413       if (info.has_discard) {
 414          aco_ptr<Pseudo_instruction> phi;
 415          for (int i = 0; i < info.num_exec_masks - 1; i++) {
 416             phi.reset(create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1));
 417             phi->definitions[0] = bld.def(bld.lm);
 418             phi->operands[0] = Operand(ctx.info[preds[0]].exec[i].first);
 419             ctx.info[idx].exec[i].first = bld.insert(std::move(phi));
 420          }
 421       }
 422
 423       /* create ssa name for restore mask */
 424       if (info.has_divergent_break) {
 425          /* this phi might be trivial but ensures a parallelcopy on the loop header */
 426          aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 427          phi->definitions[0] = bld.def(bld.lm);
 428          phi->operands[0] = Operand(ctx.info[preds[0]].exec[info.num_exec_masks - 1].first);
 429          ctx.info[idx].exec.back().first = bld.insert(std::move(phi));
 430       }
 431
 432       /* create ssa name for loop active mask */
 433       aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 434       if (info.has_divergent_continue)
 435          phi->definitions[0] = bld.def(bld.lm);
 436       else
 437          phi->definitions[0] = bld.def(bld.lm, exec);
 438       phi->operands[0] = Operand(ctx.info[preds[0]].exec.back().first);
 439       Temp loop_active = bld.insert(std::move(phi));
 440
 441       if (info.has_divergent_break) {
 442          uint8_t mask_type = (ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact)) | mask_type_loop;
 443          ctx.info[idx].exec.emplace_back(loop_active, mask_type);
 444       } else {
 445          ctx.info[idx].exec.back().first = loop_active;
 446          ctx.info[idx].exec.back().second |= mask_type_loop;
 447       }
 448
 449       /* create a parallelcopy to move the active mask to exec */
 450       unsigned i = 0;
 451       if (info.has_divergent_continue) {
 452          while (block->instructions[i]->opcode != aco_opcode::p_logical_start) {
 453             bld.insert(std::move(block->instructions[i]));
 454             i++;
 455          }
 456          uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
 457          assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 458          ctx.info[idx].exec.emplace_back(bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 459                                                     ctx.info[idx].exec.back().first), mask_type);
 460       }
 461
 462       return i;
 463    }
 464
 465    /* loop exit block */
 466    if (block->kind & block_kind_loop_exit) {
 467       Block* header = ctx.loop.back().loop_header;
 468       loop_info& info = ctx.loop.back();
 469
 470       for (ASSERTED unsigned pred : preds)
 471          assert(ctx.info[pred].exec.size() >= info.num_exec_masks);
 472
 473       /* fill the loop header phis */
 474       std::vector<unsigned>& header_preds = header->linear_preds;
 475       int k = 0;
 476       if (info.has_discard) {
 477          while (k < info.num_exec_masks - 1) {
 478             aco_ptr<Instruction>& phi = header->instructions[k];
 479             assert(phi->opcode == aco_opcode::p_linear_phi);
 480             for (unsigned i = 1; i < phi->operands.size(); i++)
 481                phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[k].first);
 482             k++;
 483          }
 484       }
 485       aco_ptr<Instruction>& phi = header->instructions[k++];
 486       assert(phi->opcode == aco_opcode::p_linear_phi);
 487       for (unsigned i = 1; i < phi->operands.size(); i++)
 488          phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[info.num_exec_masks - 1].first);
 489
 490       if (info.has_divergent_break) {
 491          aco_ptr<Instruction>& phi = header->instructions[k];
 492          assert(phi->opcode == aco_opcode::p_linear_phi);
 493          for (unsigned i = 1; i < phi->operands.size(); i++)
 494             phi->operands[i] = Operand(ctx.info[header_preds[i]].exec[info.num_exec_masks].first);
 495       }
 496
 497       assert(!(block->kind & block_kind_top_level) || info.num_exec_masks <= 2);
 498
 499       /* create the loop exit phis if not trivial */
 500       bool need_parallelcopy = false;
 501       for (unsigned k = 0; k < info.num_exec_masks; k++) {
 502          Temp same = ctx.info[preds[0]].exec[k].first;
 503          uint8_t type = ctx.info[header_preds[0]].exec[k].second;
 504          bool trivial = true;
 505
 506          for (unsigned i = 1; i < preds.size() && trivial; i++) {
 507             if (ctx.info[preds[i]].exec[k].first != same)
 508                trivial = false;
 509          }
 510
 511          if (k == info.num_exec_masks - 1u) {
 512             bool all_liveout_exec = true;
 513             bool all_not_liveout_exec = true;
 514             for (unsigned pred : preds) {
 515                all_liveout_exec = all_liveout_exec && same == ctx.program->blocks[pred].live_out_exec;
 516                all_not_liveout_exec = all_not_liveout_exec && same != ctx.program->blocks[pred].live_out_exec;
 517             }
 518             if (!all_liveout_exec && !all_not_liveout_exec)
 519                trivial = false;
 520             else if (all_not_liveout_exec)
 521                need_parallelcopy = true;
 522
 523             need_parallelcopy |= !trivial;
 524          }
 525
 526          if (trivial) {
 527             ctx.info[idx].exec.emplace_back(same, type);
 528          } else {
 529             /* create phi for loop footer */
 530             aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
 531             phi->definitions[0] = bld.def(bld.lm);
 532             if (k == info.num_exec_masks - 1u) {
 533                phi->definitions[0].setFixed(exec);
 534                need_parallelcopy = false;
 535             }
 536             for (unsigned i = 0; i < phi->operands.size(); i++)
 537                phi->operands[i] = Operand(ctx.info[preds[i]].exec[k].first);
 538             ctx.info[idx].exec.emplace_back(bld.insert(std::move(phi)), type);
 539          }
 540       }
 541       assert(ctx.info[idx].exec.size() == info.num_exec_masks);
 542
 543       /* create a parallelcopy to move the live mask to exec */
 544       unsigned i = 0;
 545       while (block->instructions[i]->opcode != aco_opcode::p_logical_start) {
 546          bld.insert(std::move(block->instructions[i]));
 547          i++;
 548       }
 549
 550       if (ctx.handle_wqm) {
 551          if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 2) {
 552             if ((ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == 0 ||
 553                 (ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == Exact) {
 554                ctx.info[idx].exec.back().second |= mask_type_global;
 555                transition_to_Exact(ctx, bld, idx);
 556                ctx.handle_wqm = false;
 557             }
 558          }
 559          if (ctx.info[idx].block_needs == WQM)
 560             transition_to_WQM(ctx, bld, idx);
 561          else if (ctx.info[idx].block_needs == Exact)
 562             transition_to_Exact(ctx, bld, idx);
 563       }
 564
 565       assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
 566       if (need_parallelcopy) {
 567          /* only create this parallelcopy is needed, since the operand isn't
 568           * fixed to exec which causes the spiller to miscalculate register demand */
 569          /* TODO: Fix register_demand calculation for spilling on loop exits.
 570           * The problem is only mitigated because the register demand could be
 571           * higher if the exec phi doesn't get assigned to exec. */
 572          ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
 573                                                       ctx.info[idx].exec.back().first);
 574       }
 575
 576       ctx.loop.pop_back();
 577       return i;
 578    }
 579
 580    if (preds.size() == 1) {
 581       ctx.info[idx].exec = ctx.info[preds[0]].exec;
 582    } else {
 583       assert(preds.size() == 2);
 584       /* if one of the predecessors ends in exact mask, we pop it from stack */
 585       unsigned num_exec_masks = std::min(ctx.info[preds[0]].exec.size(),
 586                                          ctx.info[preds[1]].exec.size());
 587       if (block->kind & block_kind_top_level && !(block->kind & block_kind_merge))
 588          num_exec_masks = std::min(num_exec_masks, 2u);
 589
 590       /* create phis for diverged exec masks */
 591       for (unsigned i = 0; i < num_exec_masks; i++) {
 592          bool in_exec = i == num_exec_masks - 1 && !(block->kind & block_kind_merge);
 593          if (!in_exec && ctx.info[preds[0]].exec[i].first == ctx.info[preds[1]].exec[i].first) {
 594             assert(ctx.info[preds[0]].exec[i].second == ctx.info[preds[1]].exec[i].second);
 595             ctx.info[idx].exec.emplace_back(ctx.info[preds[0]].exec[i]);
 596             continue;
 597          }
 598
 599          Temp phi = bld.pseudo(aco_opcode::p_linear_phi, in_exec ? bld.def(bld.lm, exec) : bld.def(bld.lm),
 600                                ctx.info[preds[0]].exec[i].first,
 601                                ctx.info[preds[1]].exec[i].first);
 602          uint8_t mask_type = ctx.info[preds[0]].exec[i].second & ctx.info[preds[1]].exec[i].second;
 603          ctx.info[idx].exec.emplace_back(phi, mask_type);
 604       }
 605    }
 606
 607    unsigned i = 0;
 608    while (block->instructions[i]->opcode == aco_opcode::p_phi ||
 609           block->instructions[i]->opcode == aco_opcode::p_linear_phi) {
 610       bld.insert(std::move(block->instructions[i]));
 611       i++;
 612    }
 613
 614    if (block->kind & block_kind_merge)
 615       ctx.info[idx].exec.pop_back();
 616
 617    if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 3) {
 618       assert(ctx.info[idx].exec.back().second == mask_type_exact);
 619       assert(block->kind & block_kind_merge);
 620       ctx.info[idx].exec.pop_back();
 621    }
 622
 623    /* try to satisfy the block's needs */
 624    if (ctx.handle_wqm) {
 625       if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 2) {
 626          if ((ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == 0 ||
 627              (ctx.info[idx].block_needs | ctx.info[idx].ever_again_needs) == Exact) {
 628             ctx.info[idx].exec.back().second |= mask_type_global;
 629             transition_to_Exact(ctx, bld, idx);
 630             ctx.handle_wqm = false;
 631          }
 632       }
 633       if (ctx.info[idx].block_needs == WQM)
 634          transition_to_WQM(ctx, bld, idx);
 635       else if (ctx.info[idx].block_needs == Exact)
 636          transition_to_Exact(ctx, bld, idx);
 637    }
 638
 639    if (block->kind & block_kind_merge) {
 640       Temp restore = ctx.info[idx].exec.back().first;
 641       assert(restore.size() == bld.lm.size());
 642       ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), restore);
 643    }
 644
 645    return i;
 646 }
 647
 648 void lower_fs_buffer_store_smem(Builder& bld, bool need_check, aco_ptr<Instruction>& instr, Temp cur_exec)
 649 {
 650    Operand offset = instr->operands[1];
 651    if (need_check) {
 652       /* if exec is zero, then use UINT32_MAX as an offset and make this store a no-op */
 653       Temp nonempty = bld.sopc(Builder::s_cmp_lg, bld.def(s1, scc), cur_exec, Operand(0u));
 654
 655       if (offset.isLiteral())
 656          offset = bld.sop1(aco_opcode::s_mov_b32, bld.def(s1), offset);
 657
 658       offset = bld.sop2(aco_opcode::s_cselect_b32, bld.hint_m0(bld.def(s1)),
 659                         offset, Operand(UINT32_MAX), bld.scc(nonempty));
 660    } else if (offset.isConstant() && offset.constantValue() > 0xFFFFF) {
 661       offset = bld.sop1(aco_opcode::s_mov_b32, bld.hint_m0(bld.def(s1)), offset);
 662    }
 663    if (!offset.isConstant())
 664       offset.setFixed(m0);
 665
 666    switch (instr->operands[2].size()) {
 667    case 1:
 668       instr->opcode = aco_opcode::s_buffer_store_dword;
 669       break;
 670    case 2:
 671       instr->opcode = aco_opcode::s_buffer_store_dwordx2;
 672       break;
 673    case 4:
 674       instr->opcode = aco_opcode::s_buffer_store_dwordx4;
 675       break;
 676    default:
 677       unreachable("Invalid SMEM buffer store size");
 678    }
 679    instr->operands[1] = offset;
 680    /* as_uniform() needs to be done here so it's done in exact mode and helper
 681     * lanes don't contribute. */
 682    instr->operands[2] = Operand(bld.as_uniform(instr->operands[2]));
 683 }
 684
 685 void process_instructions(exec_ctx& ctx, Block* block,
 686                           std::vector<aco_ptr<Instruction>>& instructions,
 687                           unsigned idx)
 688 {
 689    WQMState state;
 690    if (ctx.info[block->index].exec.back().second & mask_type_wqm)
 691       state = WQM;
 692    else {
 693       assert(!ctx.handle_wqm || ctx.info[block->index].exec.back().second & mask_type_exact);
 694       state = Exact;
 695    }
 696
 697    /* if the block doesn't need both, WQM and Exact, we can skip processing the instructions */
 698    bool process = (ctx.handle_wqm &&
 699                    (ctx.info[block->index].block_needs & state) !=
 700                    (ctx.info[block->index].block_needs & (WQM | Exact))) ||
 701                   block->kind & block_kind_uses_discard_if ||
 702                   block->kind & block_kind_uses_demote ||
 703                   block->kind & block_kind_needs_lowering;
 704    if (!process) {
 705       std::vector<aco_ptr<Instruction>>::iterator it = std::next(block->instructions.begin(), idx);
 706       instructions.insert(instructions.end(),
 707                           std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(it),
 708                           std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
 709       return;
 710    }
 711
 712    Builder bld(ctx.program, &instructions);
 713
 714    for (; idx < block->instructions.size(); idx++) {
 715       aco_ptr<Instruction> instr = std::move(block->instructions[idx]);
 716
 717       WQMState needs = ctx.handle_wqm ? ctx.info[block->index].instr_needs[idx] : Unspecified;
 718
 719       if (instr->opcode == aco_opcode::p_discard_if) {
 720          if (ctx.info[block->index].block_needs & Preserve_WQM) {
 721             assert(block->kind & block_kind_top_level);
 722             transition_to_WQM(ctx, bld, block->index);
 723             ctx.info[block->index].exec.back().second &= ~mask_type_global;
 724          }
 725          int num = ctx.info[block->index].exec.size();
 726          assert(num);
 727          Operand cond = instr->operands[0];
 728          for (int i = num - 1; i >= 0; i--) {
 729             Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
 730                                           ctx.info[block->index].exec[i].first, cond);
 731             if (i == num - 1) {
 732                andn2->operands[0].setFixed(exec);
 733                andn2->definitions[0].setFixed(exec);
 734             }
 735             if (i == 0) {
 736                instr->opcode = aco_opcode::p_exit_early_if;
 737                instr->operands[0] = bld.scc(andn2->definitions[1].getTemp());
 738             }
 739             ctx.info[block->index].exec[i].first = andn2->definitions[0].getTemp();
 740          }
 741          assert(!ctx.handle_wqm || (ctx.info[block->index].exec[0].second & mask_type_wqm) == 0);
 742
 743       } else if (needs == WQM && state != WQM) {
 744          transition_to_WQM(ctx, bld, block->index);
 745          state = WQM;
 746       } else if (needs == Exact && state != Exact) {
 747          transition_to_Exact(ctx, bld, block->index);
 748          state = Exact;
 749       }
 750
 751       if (instr->opcode == aco_opcode::p_is_helper || instr->opcode == aco_opcode::p_load_helper) {
 752          Definition dst = instr->definitions[0];
 753          assert(dst.size() == bld.lm.size());
 754          if (state == Exact) {
 755             instr.reset(create_instruction<SOP1_instruction>(bld.w64or32(Builder::s_mov), Format::SOP1, 1, 1));
 756             instr->operands[0] = Operand(0u);
 757             instr->definitions[0] = dst;
 758          } else {
 759             std::pair<Temp, uint8_t>& exact_mask = ctx.info[block->index].exec[0];
 760             if (instr->opcode == aco_opcode::p_load_helper &&
 761                 !(ctx.info[block->index].exec[0].second & mask_type_initial)) {
 762                /* find last initial exact mask */
 763                for (int i = block->index; i >= 0; i--) {
 764                   if (ctx.program->blocks[i].kind & block_kind_top_level &&
 765                       ctx.info[i].exec[0].second & mask_type_initial) {
 766                      exact_mask = ctx.info[i].exec[0];
 767                      break;
 768                   }
 769                }
 770             }
 771
 772             assert(instr->opcode == aco_opcode::p_is_helper || exact_mask.second & mask_type_initial);
 773             assert(exact_mask.second & mask_type_exact);
 774
 775             instr.reset(create_instruction<SOP2_instruction>(bld.w64or32(Builder::s_andn2), Format::SOP2, 2, 2));
 776             instr->operands[0] = Operand(ctx.info[block->index].exec.back().first); /* current exec */
 777             instr->operands[1] = Operand(exact_mask.first);
 778             instr->definitions[0] = dst;
 779             instr->definitions[1] = bld.def(s1, scc);
 780          }
 781       } else if (instr->opcode == aco_opcode::p_demote_to_helper) {
 782          /* turn demote into discard_if with only exact masks */
 783          assert((ctx.info[block->index].exec[0].second & (mask_type_exact | mask_type_global)) == (mask_type_exact | mask_type_global));
 784          ctx.info[block->index].exec[0].second &= ~mask_type_initial;
 785
 786          int num;
 787          Temp cond, exit_cond;
 788          if (instr->operands[0].isConstant()) {
 789             assert(instr->operands[0].constantValue() == -1u);
 790             /* transition to exact and set exec to zero */
 791             Temp old_exec = ctx.info[block->index].exec.back().first;
 792             Temp new_exec = bld.tmp(bld.lm);
 793             exit_cond = bld.tmp(s1);
 794             cond = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.scc(Definition(exit_cond)),
 795                             bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
 796
 797             num = ctx.info[block->index].exec.size() - 2;
 798             if (ctx.info[block->index].exec.back().second & mask_type_exact) {
 799                ctx.info[block->index].exec.back().first = new_exec;
 800             } else {
 801                ctx.info[block->index].exec.back().first = cond;
 802                ctx.info[block->index].exec.emplace_back(new_exec, mask_type_exact);
 803             }
 804          } else {
 805             /* demote_if: transition to exact */
 806             transition_to_Exact(ctx, bld, block->index);
 807             assert(instr->operands[0].isTemp());
 808             cond = instr->operands[0].getTemp();
 809             num = ctx.info[block->index].exec.size() - 1;
 810          }
 811
 812          for (int i = num; i >= 0; i--) {
 813             if (ctx.info[block->index].exec[i].second & mask_type_exact) {
 814                Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
 815                                              ctx.info[block->index].exec[i].first, cond);
 816                if (i == (int)ctx.info[block->index].exec.size() - 1) {
 817                   andn2->operands[0].setFixed(exec);
 818                   andn2->definitions[0].setFixed(exec);
 819                }
 820
 821                ctx.info[block->index].exec[i].first = andn2->definitions[0].getTemp();
 822                exit_cond = andn2->definitions[1].getTemp();
 823             } else {
 824                assert(i != 0);
 825             }
 826          }
 827          instr->opcode = aco_opcode::p_exit_early_if;
 828          instr->operands[0] = bld.scc(exit_cond);
 829          state = Exact;
 830
 831       } else if (instr->opcode == aco_opcode::p_fs_buffer_store_smem) {
 832          bool need_check = ctx.info[block->index].exec.size() != 1 &&
 833                            !(ctx.info[block->index].exec[ctx.info[block->index].exec.size() - 2].second & Exact);
 834          lower_fs_buffer_store_smem(bld, need_check, instr, ctx.info[block->index].exec.back().first);
 835       }
 836
 837       bld.insert(std::move(instr));
 838    }
 839 }
 840
 841 void add_branch_code(exec_ctx& ctx, Block* block)
 842 {
 843    unsigned idx = block->index;
 844    Builder bld(ctx.program, block);
 845
 846    if (idx == ctx.program->blocks.size() - 1)
 847       return;
 848
 849    /* try to disable wqm handling */
 850    if (ctx.handle_wqm && block->kind & block_kind_top_level) {
 851       if (ctx.info[idx].exec.size() == 3) {
 852          assert(ctx.info[idx].exec[1].second == mask_type_wqm);
 853          ctx.info[idx].exec.pop_back();
 854       }
 855       assert(ctx.info[idx].exec.size() <= 2);
 856
 857       if (ctx.info[idx].ever_again_needs == 0 ||
 858           ctx.info[idx].ever_again_needs == Exact) {
 859          /* transition to Exact */
 860          aco_ptr<Instruction> branch = std::move(block->instructions.back());
 861          block->instructions.pop_back();
 862          ctx.info[idx].exec.back().second |= mask_type_global;
 863          transition_to_Exact(ctx, bld, idx);
 864          bld.insert(std::move(branch));
 865          ctx.handle_wqm = false;
 866
 867       } else if (ctx.info[idx].block_needs & Preserve_WQM) {
 868          /* transition to WQM and remove global flag */
 869          aco_ptr<Instruction> branch = std::move(block->instructions.back());
 870          block->instructions.pop_back();
 871          transition_to_WQM(ctx, bld, idx);
 872          ctx.info[idx].exec.back().second &= ~mask_type_global;
 873          bld.insert(std::move(branch));
 874       }
 875    }
 876
 877    if (block->kind & block_kind_loop_preheader) {
 878       /* collect information about the succeeding loop */
 879       bool has_divergent_break = false;
 880       bool has_divergent_continue = false;
 881       bool has_discard = false;
 882       uint8_t needs = 0;
 883       unsigned loop_nest_depth = ctx.program->blocks[idx + 1].loop_nest_depth;
 884
 885       for (unsigned i = idx + 1; ctx.program->blocks[i].loop_nest_depth >= loop_nest_depth; i++) {
 886          Block& loop_block = ctx.program->blocks[i];
 887          needs |= ctx.info[i].block_needs;
 888
 889          if (loop_block.kind & block_kind_uses_discard_if ||
 890              loop_block.kind & block_kind_discard ||
 891              loop_block.kind & block_kind_uses_demote)
 892             has_discard = true;
 893          if (loop_block.loop_nest_depth != loop_nest_depth)
 894             continue;
 895
 896          if (loop_block.kind & block_kind_uniform)
 897             continue;
 898          else if (loop_block.kind & block_kind_break)
 899             has_divergent_break = true;
 900          else if (loop_block.kind & block_kind_continue)
 901             has_divergent_continue = true;
 902       }
 903
 904       if (ctx.handle_wqm) {
 905          if (needs & WQM) {
 906             aco_ptr<Instruction> branch = std::move(block->instructions.back());
 907             block->instructions.pop_back();
 908             transition_to_WQM(ctx, bld, idx);
 909             bld.insert(std::move(branch));
 910          } else {
 911             aco_ptr<Instruction> branch = std::move(block->instructions.back());
 912             block->instructions.pop_back();
 913             transition_to_Exact(ctx, bld, idx);
 914             bld.insert(std::move(branch));
 915          }
 916       }
 917
 918       unsigned num_exec_masks = ctx.info[idx].exec.size();
 919       if (block->kind & block_kind_top_level)
 920          num_exec_masks = std::min(num_exec_masks, 2u);
 921
 922       ctx.loop.emplace_back(&ctx.program->blocks[block->linear_succs[0]],
 923                             num_exec_masks,
 924                             needs,
 925                             has_divergent_break,
 926                             has_divergent_continue,
 927                             has_discard);
 928    }
 929
 930    if (block->kind & block_kind_discard) {
 931
 932       assert(block->instructions.back()->format == Format::PSEUDO_BRANCH);
 933       aco_ptr<Instruction> branch = std::move(block->instructions.back());
 934       block->instructions.pop_back();
 935
 936       /* create a discard_if() instruction with the exec mask as condition */
 937       unsigned num = 0;
 938       if (ctx.loop.size()) {
 939          /* if we're in a loop, only discard from the outer exec masks */
 940          num = ctx.loop.back().num_exec_masks;
 941       } else {
 942          num = ctx.info[idx].exec.size() - 1;
 943       }
 944
 945       Temp old_exec = ctx.info[idx].exec.back().first;
 946       Temp new_exec = bld.tmp(bld.lm);
 947       Temp cond = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
 948                            bld.exec(Definition(new_exec)), Operand(0u), bld.exec(old_exec));
 949       ctx.info[idx].exec.back().first = new_exec;
 950
 951       for (int i = num - 1; i >= 0; i--) {
 952          Instruction *andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
 953                                        ctx.info[block->index].exec[i].first, cond);
 954          if (i == (int)ctx.info[idx].exec.size() - 1)
 955             andn2->definitions[0].setFixed(exec);
 956          if (i == 0)
 957             bld.pseudo(aco_opcode::p_exit_early_if, bld.scc(andn2->definitions[1].getTemp()));
 958          ctx.info[block->index].exec[i].first = andn2->definitions[0].getTemp();
 959       }
 960       assert(!ctx.handle_wqm || (ctx.info[block->index].exec[0].second & mask_type_wqm) == 0);
 961
 962       if ((block->kind & (block_kind_break | block_kind_uniform)) == block_kind_break)
 963          ctx.info[idx].exec.back().first = cond;
 964       bld.insert(std::move(branch));
 965       /* no return here as it can be followed by a divergent break */
 966    }
 967
 968    if (block->kind & block_kind_continue_or_break) {
 969       assert(ctx.program->blocks[ctx.program->blocks[block->linear_succs[1]].linear_succs[0]].kind & block_kind_loop_header);
 970       assert(ctx.program->blocks[ctx.program->blocks[block->linear_succs[0]].linear_succs[0]].kind & block_kind_loop_exit);
 971       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
 972       block->instructions.pop_back();
 973
 974       bool need_parallelcopy = false;
 975       while (!(ctx.info[idx].exec.back().second & mask_type_loop)) {
 976          ctx.info[idx].exec.pop_back();
 977          need_parallelcopy = true;
 978       }
 979
 980       if (need_parallelcopy)
 981          ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), ctx.info[idx].exec.back().first);
 982       bld.branch(aco_opcode::p_cbranch_nz, bld.exec(ctx.info[idx].exec.back().first), block->linear_succs[1], block->linear_succs[0]);
 983       return;
 984    }
 985
 986    if (block->kind & block_kind_uniform) {
 987       Pseudo_branch_instruction* branch = static_cast<Pseudo_branch_instruction*>(block->instructions.back().get());
 988       if (branch->opcode == aco_opcode::p_branch) {
 989          branch->target[0] = block->linear_succs[0];
 990       } else {
 991          branch->target[0] = block->linear_succs[1];
 992          branch->target[1] = block->linear_succs[0];
 993       }
 994       return;
 995    }
 996
 997    if (block->kind & block_kind_branch) {
 998
 999       if (ctx.handle_wqm &&
1000           ctx.info[idx].exec.size() >= 2 &&
1001           ctx.info[idx].exec.back().second == mask_type_exact &&
1002           !(ctx.info[idx].block_needs & Exact_Branch) &&
1003           ctx.info[idx].exec[ctx.info[idx].exec.size() - 2].second & mask_type_wqm) {
1004          /* return to wqm before branching */
1005          ctx.info[idx].exec.pop_back();
1006       }
1007
1008       // orig = s_and_saveexec_b64
1009       assert(block->linear_succs.size() == 2);
1010       assert(block->instructions.back()->opcode == aco_opcode::p_cbranch_z);
1011       Temp cond = block->instructions.back()->operands[0].getTemp();
1012       block->instructions.pop_back();
1013
1014       if (ctx.info[idx].block_needs & Exact_Branch)
1015          transition_to_Exact(ctx, bld, idx);
1016
1017       Temp current_exec = ctx.info[idx].exec.back().first;
1018       uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
1019
1020       Temp then_mask = bld.tmp(bld.lm);
1021       Temp old_exec = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
1022                                bld.exec(Definition(then_mask)), cond, bld.exec(current_exec));
1023
1024       ctx.info[idx].exec.back().first = old_exec;
1025
1026       /* add next current exec to the stack */
1027       ctx.info[idx].exec.emplace_back(then_mask, mask_type);
1028
1029       bld.branch(aco_opcode::p_cbranch_z, bld.exec(then_mask), block->linear_succs[1], block->linear_succs[0]);
1030       return;
1031    }
1032
1033    if (block->kind & block_kind_invert) {
1034       // exec = s_andn2_b64 (original_exec, exec)
1035       assert(block->instructions.back()->opcode == aco_opcode::p_cbranch_nz);
1036       block->instructions.pop_back();
1037       Temp then_mask = ctx.info[idx].exec.back().first;
1038       uint8_t mask_type = ctx.info[idx].exec.back().second;
1039       ctx.info[idx].exec.pop_back();
1040       Temp orig_exec = ctx.info[idx].exec.back().first;
1041       Temp else_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm, exec),
1042                                 bld.def(s1, scc), orig_exec, bld.exec(then_mask));
1043
1044       /* add next current exec to the stack */
1045       ctx.info[idx].exec.emplace_back(else_mask, mask_type);
1046
1047       bld.branch(aco_opcode::p_cbranch_z, bld.exec(else_mask), block->linear_succs[1], block->linear_succs[0]);
1048       return;
1049    }
1050
1051    if (block->kind & block_kind_break) {
1052       // loop_mask = s_andn2_b64 (loop_mask, exec)
1053       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
1054       block->instructions.pop_back();
1055
1056       Temp current_exec = ctx.info[idx].exec.back().first;
1057       Temp cond = Temp();
1058       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
1059          cond = bld.tmp(s1);
1060          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
1061          exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
1062                               exec_mask, bld.exec(current_exec));
1063          ctx.info[idx].exec[exec_idx].first = exec_mask;
1064          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
1065             break;
1066       }
1067
1068       /* check if the successor is the merge block, otherwise set exec to 0 */
1069       // TODO: this could be done better by directly branching to the merge block
1070       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
1071       Block& succ = ctx.program->blocks[succ_idx];
1072       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
1073          ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
1074       }
1075
1076       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
1077       return;
1078    }
1079
1080    if (block->kind & block_kind_continue) {
1081       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
1082       block->instructions.pop_back();
1083
1084       Temp current_exec = ctx.info[idx].exec.back().first;
1085       Temp cond = Temp();
1086       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
1087          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
1088             break;
1089          cond = bld.tmp(s1);
1090          Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
1091          exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
1092                               exec_mask, bld.exec(current_exec));
1093          ctx.info[idx].exec[exec_idx].first = exec_mask;
1094       }
1095       assert(cond != Temp());
1096
1097       /* check if the successor is the merge block, otherwise set exec to 0 */
1098       // TODO: this could be done better by directly branching to the merge block
1099       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
1100       Block& succ = ctx.program->blocks[succ_idx];
1101       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
1102          ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
1103       }
1104
1105       bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
1106       return;
1107    }
1108 }
1109
1110 void process_block(exec_ctx& ctx, Block* block)
1111 {
1112    std::vector<aco_ptr<Instruction>> instructions;
1113    instructions.reserve(block->instructions.size());
1114
1115    unsigned idx = add_coupling_code(ctx, block, instructions);
1116
1117    assert(block->index != ctx.program->blocks.size() - 1 ||
1118           ctx.info[block->index].exec.size() <= 2);
1119
1120    process_instructions(ctx, block, instructions, idx);
1121
1122    block->instructions = std::move(instructions);
1123
1124    add_branch_code(ctx, block);
1125
1126    block->live_out_exec = ctx.info[block->index].exec.back().first;
1127 }
1128
1129 } /* end namespace */
1130
1131
1132 void insert_exec_mask(Program *program)
1133 {
1134    exec_ctx ctx(program);
1135
1136    if (program->needs_wqm && program->needs_exact)
1137       calculate_wqm_needs(ctx);
1138
1139    for (Block& block : program->blocks)
1140       process_block(ctx, &block);
1141
1142 }
1143
1144 }
1145