src/amd/compiler/aco_spill.cpp

   1 /*
   2  * Copyright © 2018 Valve Corporation
   3  * Copyright © 2018 Google
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  *
  24  */
  25
  26 #include "aco_ir.h"
  27 #include "aco_builder.h"
  28 #include "sid.h"
  29
  30 #include <map>
  31 #include <stack>
  32
  33 /*
  34  * Implements the spilling algorithm on SSA-form from
  35  * "Register Spilling and Live-Range Splitting for SSA-Form Programs"
  36  * by Matthias Braun and Sebastian Hack.
  37  */
  38
  39 namespace aco {
  40
  41 namespace {
  42
  43 struct remat_info {
  44    Instruction *instr;
  45 };
  46
  47 struct spill_ctx {
  48    RegisterDemand target_pressure;
  49    Program* program;
  50    std::vector<std::vector<RegisterDemand>> register_demand;
  51    std::vector<std::map<Temp, Temp>> renames;
  52    std::vector<std::map<Temp, uint32_t>> spills_entry;
  53    std::vector<std::map<Temp, uint32_t>> spills_exit;
  54    std::vector<bool> processed;
  55    std::stack<Block*> loop_header;
  56    std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_start;
  57    std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_end;
  58    std::vector<std::pair<RegClass, std::set<uint32_t>>> interferences;
  59    std::vector<std::vector<uint32_t>> affinities;
  60    std::vector<bool> is_reloaded;
  61    std::map<Temp, remat_info> remat;
  62    std::map<Instruction *, bool> remat_used;
  63    unsigned wave_size;
  64
  65    spill_ctx(const RegisterDemand target_pressure, Program* program,
  66              std::vector<std::vector<RegisterDemand>> register_demand)
  67       : target_pressure(target_pressure), program(program),
  68         register_demand(std::move(register_demand)), renames(program->blocks.size()),
  69         spills_entry(program->blocks.size()), spills_exit(program->blocks.size()),
  70         processed(program->blocks.size(), false), wave_size(program->wave_size) {}
  71
  72    void add_affinity(uint32_t first, uint32_t second)
  73    {
  74       unsigned found_first = affinities.size();
  75       unsigned found_second = affinities.size();
  76       for (unsigned i = 0; i < affinities.size(); i++) {
  77          std::vector<uint32_t>& vec = affinities[i];
  78          for (uint32_t entry : vec) {
  79             if (entry == first)
  80                found_first = i;
  81             else if (entry == second)
  82                found_second = i;
  83          }
  84       }
  85       if (found_first == affinities.size() && found_second == affinities.size()) {
  86          affinities.emplace_back(std::vector<uint32_t>({first, second}));
  87       } else if (found_first < affinities.size() && found_second == affinities.size()) {
  88          affinities[found_first].push_back(second);
  89       } else if (found_second < affinities.size() && found_first == affinities.size()) {
  90          affinities[found_second].push_back(first);
  91       } else if (found_first != found_second) {
  92          /* merge second into first */
  93          affinities[found_first].insert(affinities[found_first].end(), affinities[found_second].begin(), affinities[found_second].end());
  94          affinities.erase(std::next(affinities.begin(), found_second));
  95       } else {
  96          assert(found_first == found_second);
  97       }
  98    }
  99
 100    uint32_t allocate_spill_id(RegClass rc)
 101    {
 102       interferences.emplace_back(rc, std::set<uint32_t>());
 103       is_reloaded.push_back(false);
 104       return next_spill_id++;
 105    }
 106
 107    uint32_t next_spill_id = 0;
 108 };
 109
 110 int32_t get_dominator(int idx_a, int idx_b, Program* program, bool is_linear)
 111 {
 112
 113    if (idx_a == -1)
 114       return idx_b;
 115    if (idx_b == -1)
 116       return idx_a;
 117    if (is_linear) {
 118       while (idx_a != idx_b) {
 119          if (idx_a > idx_b)
 120             idx_a = program->blocks[idx_a].linear_idom;
 121          else
 122             idx_b = program->blocks[idx_b].linear_idom;
 123       }
 124    } else {
 125       while (idx_a != idx_b) {
 126          if (idx_a > idx_b)
 127             idx_a = program->blocks[idx_a].logical_idom;
 128          else
 129             idx_b = program->blocks[idx_b].logical_idom;
 130       }
 131    }
 132    assert(idx_a != -1);
 133    return idx_a;
 134 }
 135
 136 void next_uses_per_block(spill_ctx& ctx, unsigned block_idx, std::set<uint32_t>& worklist)
 137 {
 138    Block* block = &ctx.program->blocks[block_idx];
 139    std::map<Temp, std::pair<uint32_t, uint32_t>> next_uses = ctx.next_use_distances_end[block_idx];
 140
 141    /* to compute the next use distance at the beginning of the block, we have to add the block's size */
 142    for (std::map<Temp, std::pair<uint32_t, uint32_t>>::iterator it = next_uses.begin(); it != next_uses.end(); ++it)
 143       it->second.second = it->second.second + block->instructions.size();
 144
 145    int idx = block->instructions.size() - 1;
 146    while (idx >= 0) {
 147       aco_ptr<Instruction>& instr = block->instructions[idx];
 148
 149       if (instr->opcode == aco_opcode::p_linear_phi ||
 150           instr->opcode == aco_opcode::p_phi)
 151          break;
 152
 153       for (const Definition& def : instr->definitions) {
 154          if (def.isTemp())
 155             next_uses.erase(def.getTemp());
 156       }
 157
 158       for (const Operand& op : instr->operands) {
 159          /* omit exec mask */
 160          if (op.isFixed() && op.physReg() == exec)
 161             continue;
 162          if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear())
 163             continue;
 164          if (op.isTemp())
 165             next_uses[op.getTemp()] = {block_idx, idx};
 166       }
 167       idx--;
 168    }
 169
 170    assert(block_idx != 0 || next_uses.empty());
 171    ctx.next_use_distances_start[block_idx] = next_uses;
 172    while (idx >= 0) {
 173       aco_ptr<Instruction>& instr = block->instructions[idx];
 174       assert(instr->opcode == aco_opcode::p_linear_phi || instr->opcode == aco_opcode::p_phi);
 175
 176       for (unsigned i = 0; i < instr->operands.size(); i++) {
 177          unsigned pred_idx = instr->opcode == aco_opcode::p_phi ?
 178                              block->logical_preds[i] :
 179                              block->linear_preds[i];
 180          if (instr->operands[i].isTemp()) {
 181             if (instr->operands[i].getTemp() == ctx.program->blocks[pred_idx].live_out_exec)
 182                continue;
 183             if (ctx.next_use_distances_end[pred_idx].find(instr->operands[i].getTemp()) == ctx.next_use_distances_end[pred_idx].end() ||
 184                 ctx.next_use_distances_end[pred_idx][instr->operands[i].getTemp()] != std::pair<uint32_t, uint32_t>{block_idx, 0})
 185                worklist.insert(pred_idx);
 186             ctx.next_use_distances_end[pred_idx][instr->operands[i].getTemp()] = {block_idx, 0};
 187          }
 188       }
 189       next_uses.erase(instr->definitions[0].getTemp());
 190       idx--;
 191    }
 192
 193    /* all remaining live vars must be live-out at the predecessors */
 194    for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : next_uses) {
 195       Temp temp = pair.first;
 196       uint32_t distance = pair.second.second;
 197       uint32_t dom = pair.second.first;
 198       std::vector<unsigned>& preds = temp.is_linear() ? block->linear_preds : block->logical_preds;
 199       for (unsigned pred_idx : preds) {
 200          if (temp == ctx.program->blocks[pred_idx].live_out_exec)
 201             continue;
 202          if (ctx.program->blocks[pred_idx].loop_nest_depth > block->loop_nest_depth)
 203             distance += 0xFFFF;
 204          if (ctx.next_use_distances_end[pred_idx].find(temp) != ctx.next_use_distances_end[pred_idx].end()) {
 205             dom = get_dominator(dom, ctx.next_use_distances_end[pred_idx][temp].first, ctx.program, temp.is_linear());
 206             distance = std::min(ctx.next_use_distances_end[pred_idx][temp].second, distance);
 207          }
 208          if (ctx.next_use_distances_end[pred_idx][temp] != std::pair<uint32_t, uint32_t>{dom, distance})
 209             worklist.insert(pred_idx);
 210          ctx.next_use_distances_end[pred_idx][temp] = {dom, distance};
 211       }
 212    }
 213
 214 }
 215
 216 void compute_global_next_uses(spill_ctx& ctx, std::vector<std::set<Temp>>& live_out)
 217 {
 218    ctx.next_use_distances_start.resize(ctx.program->blocks.size());
 219    ctx.next_use_distances_end.resize(ctx.program->blocks.size());
 220    std::set<uint32_t> worklist;
 221    for (Block& block : ctx.program->blocks)
 222       worklist.insert(block.index);
 223
 224    while (!worklist.empty()) {
 225       std::set<unsigned>::reverse_iterator b_it = worklist.rbegin();
 226       unsigned block_idx = *b_it;
 227       worklist.erase(block_idx);
 228       next_uses_per_block(ctx, block_idx, worklist);
 229    }
 230 }
 231
 232 bool should_rematerialize(aco_ptr<Instruction>& instr)
 233 {
 234    /* TODO: rematerialization is only supported for VOP1, SOP1 and PSEUDO */
 235    if (instr->format != Format::VOP1 && instr->format != Format::SOP1 && instr->format != Format::PSEUDO && instr->format != Format::SOPK)
 236       return false;
 237    /* TODO: pseudo-instruction rematerialization is only supported for p_create_vector */
 238    if (instr->format == Format::PSEUDO && instr->opcode != aco_opcode::p_create_vector)
 239       return false;
 240    if (instr->format == Format::SOPK && instr->opcode != aco_opcode::s_movk_i32)
 241       return false;
 242
 243    for (const Operand& op : instr->operands) {
 244       /* TODO: rematerialization using temporaries isn't yet supported */
 245       if (op.isTemp())
 246          return false;
 247    }
 248
 249    /* TODO: rematerialization with multiple definitions isn't yet supported */
 250    if (instr->definitions.size() > 1)
 251       return false;
 252
 253    return true;
 254 }
 255
 256 aco_ptr<Instruction> do_reload(spill_ctx& ctx, Temp tmp, Temp new_name, uint32_t spill_id)
 257 {
 258    std::map<Temp, remat_info>::iterator remat = ctx.remat.find(tmp);
 259    if (remat != ctx.remat.end()) {
 260       Instruction *instr = remat->second.instr;
 261       assert((instr->format == Format::VOP1 || instr->format == Format::SOP1 || instr->format == Format::PSEUDO || instr->format == Format::SOPK) && "unsupported");
 262       assert((instr->format != Format::PSEUDO || instr->opcode == aco_opcode::p_create_vector) && "unsupported");
 263       assert(instr->definitions.size() == 1 && "unsupported");
 264
 265       aco_ptr<Instruction> res;
 266       if (instr->format == Format::VOP1) {
 267          res.reset(create_instruction<VOP1_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
 268       } else if (instr->format == Format::SOP1) {
 269          res.reset(create_instruction<SOP1_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
 270       } else if (instr->format == Format::PSEUDO) {
 271          res.reset(create_instruction<Pseudo_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
 272       } else if (instr->format == Format::SOPK) {
 273          res.reset(create_instruction<SOPK_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
 274          static_cast<SOPK_instruction*>(res.get())->imm = static_cast<SOPK_instruction*>(instr)->imm;
 275       }
 276       for (unsigned i = 0; i < instr->operands.size(); i++) {
 277          res->operands[i] = instr->operands[i];
 278          if (instr->operands[i].isTemp()) {
 279             assert(false && "unsupported");
 280             if (ctx.remat.count(instr->operands[i].getTemp()))
 281                ctx.remat_used[ctx.remat[instr->operands[i].getTemp()].instr] = true;
 282          }
 283       }
 284       res->definitions[0] = Definition(new_name);
 285       return res;
 286    } else {
 287       aco_ptr<Pseudo_instruction> reload{create_instruction<Pseudo_instruction>(aco_opcode::p_reload, Format::PSEUDO, 1, 1)};
 288       reload->operands[0] = Operand(spill_id);
 289       reload->definitions[0] = Definition(new_name);
 290       ctx.is_reloaded[spill_id] = true;
 291       return reload;
 292    }
 293 }
 294
 295 void get_rematerialize_info(spill_ctx& ctx)
 296 {
 297    for (Block& block : ctx.program->blocks) {
 298       bool logical = false;
 299       for (aco_ptr<Instruction>& instr : block.instructions) {
 300          if (instr->opcode == aco_opcode::p_logical_start)
 301             logical = true;
 302          else if (instr->opcode == aco_opcode::p_logical_end)
 303             logical = false;
 304          if (logical && should_rematerialize(instr)) {
 305             for (const Definition& def : instr->definitions) {
 306                if (def.isTemp()) {
 307                   ctx.remat[def.getTemp()] = (remat_info){instr.get()};
 308                   ctx.remat_used[instr.get()] = false;
 309                }
 310             }
 311          }
 312       }
 313    }
 314 }
 315
 316 std::vector<std::map<Temp, uint32_t>> local_next_uses(spill_ctx& ctx, Block* block)
 317 {
 318    std::vector<std::map<Temp, uint32_t>> local_next_uses(block->instructions.size());
 319
 320    std::map<Temp, uint32_t> next_uses;
 321    for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block->index])
 322       next_uses[pair.first] = pair.second.second + block->instructions.size();
 323
 324    for (int idx = block->instructions.size() - 1; idx >= 0; idx--) {
 325       aco_ptr<Instruction>& instr = block->instructions[idx];
 326       if (!instr)
 327          break;
 328       if (instr->opcode == aco_opcode::p_phi || instr->opcode == aco_opcode::p_linear_phi)
 329          break;
 330
 331       for (const Operand& op : instr->operands) {
 332          if (op.isFixed() && op.physReg() == exec)
 333             continue;
 334          if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear())
 335             continue;
 336          if (op.isTemp())
 337             next_uses[op.getTemp()] = idx;
 338       }
 339       for (const Definition& def : instr->definitions) {
 340          if (def.isTemp())
 341             next_uses.erase(def.getTemp());
 342       }
 343       local_next_uses[idx] = next_uses;
 344    }
 345    return local_next_uses;
 346 }
 347
 348
 349 RegisterDemand init_live_in_vars(spill_ctx& ctx, Block* block, unsigned block_idx)
 350 {
 351    RegisterDemand spilled_registers;
 352
 353    /* first block, nothing was spilled before */
 354    if (block_idx == 0)
 355       return {0, 0};
 356
 357    /* loop header block */
 358    if (block->loop_nest_depth > ctx.program->blocks[block_idx - 1].loop_nest_depth) {
 359       assert(block->linear_preds[0] == block_idx - 1);
 360       assert(block->logical_preds[0] == block_idx - 1);
 361
 362       /* create new loop_info */
 363       ctx.loop_header.emplace(block);
 364
 365       /* check how many live-through variables should be spilled */
 366       RegisterDemand new_demand;
 367       unsigned i = block_idx;
 368       while (ctx.program->blocks[i].loop_nest_depth >= block->loop_nest_depth) {
 369          assert(ctx.program->blocks.size() > i);
 370          new_demand.update(ctx.program->blocks[i].register_demand);
 371          i++;
 372       }
 373       unsigned loop_end = i;
 374
 375       /* select live-through vgpr variables */
 376       while (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) {
 377          unsigned distance = 0;
 378          Temp to_spill;
 379          for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
 380             if (pair.first.type() == RegType::vgpr &&
 381                 pair.second.first >= loop_end &&
 382                 pair.second.second > distance &&
 383                 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
 384                to_spill = pair.first;
 385                distance = pair.second.second;
 386             }
 387          }
 388          if (distance == 0)
 389             break;
 390
 391          uint32_t spill_id;
 392          if (ctx.spills_exit[block_idx - 1].find(to_spill) == ctx.spills_exit[block_idx - 1].end()) {
 393             spill_id = ctx.allocate_spill_id(to_spill.regClass());
 394          } else {
 395             spill_id = ctx.spills_exit[block_idx - 1][to_spill];
 396          }
 397
 398          ctx.spills_entry[block_idx][to_spill] = spill_id;
 399          spilled_registers.vgpr += to_spill.size();
 400       }
 401
 402       /* select live-through sgpr variables */
 403       while (new_demand.sgpr - spilled_registers.sgpr > ctx.target_pressure.sgpr) {
 404          unsigned distance = 0;
 405          Temp to_spill;
 406          for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
 407             if (pair.first.type() == RegType::sgpr &&
 408                 pair.second.first >= loop_end &&
 409                 pair.second.second > distance &&
 410                 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
 411                to_spill = pair.first;
 412                distance = pair.second.second;
 413             }
 414          }
 415          if (distance == 0)
 416             break;
 417
 418          uint32_t spill_id;
 419          if (ctx.spills_exit[block_idx - 1].find(to_spill) == ctx.spills_exit[block_idx - 1].end()) {
 420             spill_id = ctx.allocate_spill_id(to_spill.regClass());
 421          } else {
 422             spill_id = ctx.spills_exit[block_idx - 1][to_spill];
 423          }
 424
 425          ctx.spills_entry[block_idx][to_spill] = spill_id;
 426          spilled_registers.sgpr += to_spill.size();
 427       }
 428
 429
 430
 431       /* shortcut */
 432       if (!RegisterDemand(new_demand - spilled_registers).exceeds(ctx.target_pressure))
 433          return spilled_registers;
 434
 435       /* if reg pressure is too high at beginning of loop, add variables with furthest use */
 436       unsigned idx = 0;
 437       while (block->instructions[idx]->opcode == aco_opcode::p_phi || block->instructions[idx]->opcode == aco_opcode::p_linear_phi)
 438          idx++;
 439
 440       assert(idx != 0 && "loop without phis: TODO");
 441       idx--;
 442       RegisterDemand reg_pressure = ctx.register_demand[block_idx][idx] - spilled_registers;
 443       while (reg_pressure.sgpr > ctx.target_pressure.sgpr) {
 444          unsigned distance = 0;
 445          Temp to_spill;
 446          for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
 447             if (pair.first.type() == RegType::sgpr &&
 448                 pair.second.second > distance &&
 449                 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
 450                to_spill = pair.first;
 451                distance = pair.second.second;
 452             }
 453          }
 454          assert(distance != 0);
 455
 456          ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
 457          spilled_registers.sgpr += to_spill.size();
 458          reg_pressure.sgpr -= to_spill.size();
 459       }
 460       while (reg_pressure.vgpr > ctx.target_pressure.vgpr) {
 461          unsigned distance = 0;
 462          Temp to_spill;
 463          for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
 464             if (pair.first.type() == RegType::vgpr &&
 465                 pair.second.second > distance &&
 466                 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
 467                to_spill = pair.first;
 468                distance = pair.second.second;
 469             }
 470          }
 471          assert(distance != 0);
 472          ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
 473          spilled_registers.vgpr += to_spill.size();
 474          reg_pressure.vgpr -= to_spill.size();
 475       }
 476
 477       return spilled_registers;
 478    }
 479
 480    /* branch block */
 481    if (block->linear_preds.size() == 1 && !(block->kind & block_kind_loop_exit)) {
 482       /* keep variables spilled if they are alive and not used in the current block */
 483       unsigned pred_idx = block->linear_preds[0];
 484       for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
 485          if (pair.first.type() == RegType::sgpr &&
 486              ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
 487              ctx.next_use_distances_start[block_idx][pair.first].second > block_idx) {
 488             ctx.spills_entry[block_idx].insert(pair);
 489             spilled_registers.sgpr += pair.first.size();
 490          }
 491       }
 492       if (block->logical_preds.size() == 1) {
 493          pred_idx = block->logical_preds[0];
 494          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
 495             if (pair.first.type() == RegType::vgpr &&
 496                 ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
 497                 ctx.next_use_distances_start[block_idx][pair.first].second > block_idx) {
 498                ctx.spills_entry[block_idx].insert(pair);
 499                spilled_registers.vgpr += pair.first.size();
 500             }
 501          }
 502       }
 503
 504       /* if register demand is still too high, we just keep all spilled live vars and process the block */
 505       if (block->register_demand.sgpr - spilled_registers.sgpr > ctx.target_pressure.sgpr) {
 506          pred_idx = block->linear_preds[0];
 507          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
 508             if (pair.first.type() == RegType::sgpr &&
 509                 ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
 510                 ctx.spills_entry[block_idx].insert(pair).second) {
 511                spilled_registers.sgpr += pair.first.size();
 512             }
 513          }
 514       }
 515       if (block->register_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr && block->logical_preds.size() == 1) {
 516          pred_idx = block->logical_preds[0];
 517          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
 518             if (pair.first.type() == RegType::vgpr &&
 519                 ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
 520                 ctx.spills_entry[block_idx].insert(pair).second) {
 521                spilled_registers.vgpr += pair.first.size();
 522             }
 523          }
 524       }
 525
 526       return spilled_registers;
 527    }
 528
 529    /* else: merge block */
 530    std::set<Temp> partial_spills;
 531
 532    /* keep variables spilled on all incoming paths */
 533    for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
 534       std::vector<unsigned>& preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds;
 535       /* If it can be rematerialized, keep the variable spilled if all predecessors do not reload it.
 536        * Otherwise, if any predecessor reloads it, ensure it's reloaded on all other predecessors.
 537        * The idea is that it's better in practice to rematerialize redundantly than to create lots of phis. */
 538       /* TODO: test this idea with more than Dawn of War III shaders (the current pipeline-db doesn't seem to exercise this path much) */
 539       bool remat = ctx.remat.count(pair.first);
 540       bool spill = !remat;
 541       uint32_t spill_id = 0;
 542       for (unsigned pred_idx : preds) {
 543          /* variable is not even live at the predecessor: probably from a phi */
 544          if (ctx.next_use_distances_end[pred_idx].find(pair.first) == ctx.next_use_distances_end[pred_idx].end()) {
 545             spill = false;
 546             break;
 547          }
 548          if (ctx.spills_exit[pred_idx].find(pair.first) == ctx.spills_exit[pred_idx].end()) {
 549             if (!remat)
 550                spill = false;
 551          } else {
 552             partial_spills.insert(pair.first);
 553             /* it might be that on one incoming path, the variable has a different spill_id, but add_couple_code() will take care of that. */
 554             spill_id = ctx.spills_exit[pred_idx][pair.first];
 555             if (remat)
 556                spill = true;
 557          }
 558       }
 559       if (spill) {
 560          ctx.spills_entry[block_idx][pair.first] = spill_id;
 561          partial_spills.erase(pair.first);
 562          spilled_registers += pair.first;
 563       }
 564    }
 565
 566    /* same for phis */
 567    unsigned idx = 0;
 568    while (block->instructions[idx]->opcode == aco_opcode::p_linear_phi ||
 569           block->instructions[idx]->opcode == aco_opcode::p_phi) {
 570       aco_ptr<Instruction>& phi = block->instructions[idx];
 571       std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
 572       bool spill = true;
 573
 574       for (unsigned i = 0; i < phi->operands.size(); i++) {
 575          if (phi->operands[i].isUndefined())
 576             continue;
 577          assert(phi->operands[i].isTemp());
 578          if (ctx.spills_exit[preds[i]].find(phi->operands[i].getTemp()) == ctx.spills_exit[preds[i]].end())
 579             spill = false;
 580          else
 581             partial_spills.insert(phi->definitions[0].getTemp());
 582       }
 583       if (spill) {
 584          ctx.spills_entry[block_idx][phi->definitions[0].getTemp()] = ctx.allocate_spill_id(phi->definitions[0].regClass());
 585          partial_spills.erase(phi->definitions[0].getTemp());
 586          spilled_registers += phi->definitions[0].getTemp();
 587       }
 588
 589       idx++;
 590    }
 591
 592    /* if reg pressure at first instruction is still too high, add partially spilled variables */
 593    RegisterDemand reg_pressure;
 594    if (idx == 0) {
 595       for (const Definition& def : block->instructions[idx]->definitions) {
 596          if (def.isTemp()) {
 597             reg_pressure -= def.getTemp();
 598          }
 599       }
 600       for (const Operand& op : block->instructions[idx]->operands) {
 601          if (op.isTemp() && op.isFirstKill()) {
 602             reg_pressure += op.getTemp();
 603          }
 604       }
 605    } else {
 606       idx--;
 607    }
 608    reg_pressure += ctx.register_demand[block_idx][idx] - spilled_registers;
 609
 610    while (reg_pressure.sgpr > ctx.target_pressure.sgpr) {
 611       assert(!partial_spills.empty());
 612
 613       std::set<Temp>::iterator it = partial_spills.begin();
 614       Temp to_spill = *it;
 615       unsigned distance = ctx.next_use_distances_start[block_idx][*it].second;
 616       while (it != partial_spills.end()) {
 617          assert(ctx.spills_entry[block_idx].find(*it) == ctx.spills_entry[block_idx].end());
 618
 619          if (it->type() == RegType::sgpr && ctx.next_use_distances_start[block_idx][*it].second > distance) {
 620             distance = ctx.next_use_distances_start[block_idx][*it].second;
 621             to_spill = *it;
 622          }
 623          ++it;
 624       }
 625       assert(distance != 0);
 626
 627       ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
 628       partial_spills.erase(to_spill);
 629       spilled_registers.sgpr += to_spill.size();
 630       reg_pressure.sgpr -= to_spill.size();
 631    }
 632
 633    while (reg_pressure.vgpr > ctx.target_pressure.vgpr) {
 634       assert(!partial_spills.empty());
 635
 636       std::set<Temp>::iterator it = partial_spills.begin();
 637       Temp to_spill = *it;
 638       unsigned distance = ctx.next_use_distances_start[block_idx][*it].second;
 639       while (it != partial_spills.end()) {
 640          assert(ctx.spills_entry[block_idx].find(*it) == ctx.spills_entry[block_idx].end());
 641
 642          if (it->type() == RegType::vgpr && ctx.next_use_distances_start[block_idx][*it].second > distance) {
 643             distance = ctx.next_use_distances_start[block_idx][*it].second;
 644             to_spill = *it;
 645          }
 646          ++it;
 647       }
 648       assert(distance != 0);
 649
 650       ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
 651       partial_spills.erase(to_spill);
 652       spilled_registers.vgpr += to_spill.size();
 653       reg_pressure.vgpr -= to_spill.size();
 654    }
 655
 656    return spilled_registers;
 657 }
 658
 659
 660 RegisterDemand get_demand_before(spill_ctx& ctx, unsigned block_idx, unsigned idx)
 661 {
 662    if (idx == 0) {
 663       RegisterDemand demand = ctx.register_demand[block_idx][idx];
 664       aco_ptr<Instruction>& instr = ctx.program->blocks[block_idx].instructions[idx];
 665       aco_ptr<Instruction> instr_before(nullptr);
 666       return get_demand_before(demand, instr, instr_before);
 667    } else {
 668       return ctx.register_demand[block_idx][idx - 1];
 669    }
 670 }
 671
 672 void add_coupling_code(spill_ctx& ctx, Block* block, unsigned block_idx)
 673 {
 674    /* no coupling code necessary */
 675    if (block->linear_preds.size() == 0)
 676       return;
 677
 678    std::vector<aco_ptr<Instruction>> instructions;
 679    /* branch block: TODO take other branch into consideration */
 680    if (block->linear_preds.size() == 1 && !(block->kind & (block_kind_loop_exit | block_kind_loop_header))) {
 681       assert(ctx.processed[block->linear_preds[0]]);
 682       assert(ctx.register_demand[block_idx].size() == block->instructions.size());
 683       std::vector<RegisterDemand> reg_demand;
 684       unsigned insert_idx = 0;
 685       unsigned pred_idx = block->linear_preds[0];
 686       RegisterDemand demand_before = get_demand_before(ctx, block_idx, 0);
 687
 688       for (std::pair<Temp, std::pair<uint32_t, uint32_t>> live : ctx.next_use_distances_start[block_idx]) {
 689          if (!live.first.is_linear())
 690             continue;
 691          /* still spilled */
 692          if (ctx.spills_entry[block_idx].find(live.first) != ctx.spills_entry[block_idx].end())
 693             continue;
 694
 695          /* in register at end of predecessor */
 696          if (ctx.spills_exit[pred_idx].find(live.first) == ctx.spills_exit[pred_idx].end()) {
 697             std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(live.first);
 698             if (it != ctx.renames[pred_idx].end())
 699                ctx.renames[block_idx].insert(*it);
 700             continue;
 701          }
 702
 703          /* variable is spilled at predecessor and live at current block: create reload instruction */
 704          Temp new_name = {ctx.program->allocateId(), live.first.regClass()};
 705          aco_ptr<Instruction> reload = do_reload(ctx, live.first, new_name, ctx.spills_exit[pred_idx][live.first]);
 706          instructions.emplace_back(std::move(reload));
 707          reg_demand.push_back(demand_before);
 708          ctx.renames[block_idx][live.first] = new_name;
 709       }
 710
 711       if (block->logical_preds.size() == 1) {
 712          do {
 713             assert(insert_idx < block->instructions.size());
 714             instructions.emplace_back(std::move(block->instructions[insert_idx]));
 715             reg_demand.push_back(ctx.register_demand[block_idx][insert_idx]);
 716             insert_idx++;
 717          } while (instructions.back()->opcode != aco_opcode::p_logical_start);
 718
 719          unsigned pred_idx = block->logical_preds[0];
 720          for (std::pair<Temp, std::pair<uint32_t, uint32_t>> live : ctx.next_use_distances_start[block_idx]) {
 721             if (live.first.is_linear())
 722                continue;
 723             /* still spilled */
 724             if (ctx.spills_entry[block_idx].find(live.first) != ctx.spills_entry[block_idx].end())
 725                continue;
 726
 727             /* in register at end of predecessor */
 728             if (ctx.spills_exit[pred_idx].find(live.first) == ctx.spills_exit[pred_idx].end()) {
 729                std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(live.first);
 730                if (it != ctx.renames[pred_idx].end())
 731                   ctx.renames[block_idx].insert(*it);
 732                continue;
 733             }
 734
 735             /* variable is spilled at predecessor and live at current block: create reload instruction */
 736             Temp new_name = {ctx.program->allocateId(), live.first.regClass()};
 737             aco_ptr<Instruction> reload = do_reload(ctx, live.first, new_name, ctx.spills_exit[pred_idx][live.first]);
 738             instructions.emplace_back(std::move(reload));
 739             reg_demand.emplace_back(reg_demand.back());
 740             ctx.renames[block_idx][live.first] = new_name;
 741          }
 742       }
 743
 744       /* combine new reload instructions with original block */
 745       if (!instructions.empty()) {
 746          reg_demand.insert(reg_demand.end(), std::next(ctx.register_demand[block->index].begin(), insert_idx),
 747                            ctx.register_demand[block->index].end());
 748          ctx.register_demand[block_idx] = std::move(reg_demand);
 749          instructions.insert(instructions.end(),
 750                              std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(std::next(block->instructions.begin(), insert_idx)),
 751                              std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
 752          block->instructions = std::move(instructions);
 753       }
 754       return;
 755    }
 756
 757    /* loop header and merge blocks: check if all (linear) predecessors have been processed */
 758    for (ASSERTED unsigned pred : block->linear_preds)
 759       assert(ctx.processed[pred]);
 760
 761    /* iterate the phi nodes for which operands to spill at the predecessor */
 762    for (aco_ptr<Instruction>& phi : block->instructions) {
 763       if (phi->opcode != aco_opcode::p_phi &&
 764           phi->opcode != aco_opcode::p_linear_phi)
 765          break;
 766
 767       /* if the phi is not spilled, add to instructions */
 768       if (ctx.spills_entry[block_idx].find(phi->definitions[0].getTemp()) == ctx.spills_entry[block_idx].end()) {
 769          instructions.emplace_back(std::move(phi));
 770          continue;
 771       }
 772
 773       std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
 774       uint32_t def_spill_id = ctx.spills_entry[block_idx][phi->definitions[0].getTemp()];
 775
 776       for (unsigned i = 0; i < phi->operands.size(); i++) {
 777          if (phi->operands[i].isUndefined())
 778             continue;
 779
 780          unsigned pred_idx = preds[i];
 781          assert(phi->operands[i].isTemp() && phi->operands[i].isKill());
 782          Temp var = phi->operands[i].getTemp();
 783
 784          /* build interferences between the phi def and all spilled variables at the predecessor blocks */
 785          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
 786             if (var == pair.first)
 787                continue;
 788             ctx.interferences[def_spill_id].second.emplace(pair.second);
 789             ctx.interferences[pair.second].second.emplace(def_spill_id);
 790          }
 791
 792          /* check if variable is already spilled at predecessor */
 793          std::map<Temp, uint32_t>::iterator spilled = ctx.spills_exit[pred_idx].find(var);
 794          if (spilled != ctx.spills_exit[pred_idx].end()) {
 795             if (spilled->second != def_spill_id)
 796                ctx.add_affinity(def_spill_id, spilled->second);
 797             continue;
 798          }
 799
 800          /* rename if necessary */
 801          std::map<Temp, Temp>::iterator rename_it = ctx.renames[pred_idx].find(var);
 802          if (rename_it != ctx.renames[pred_idx].end()) {
 803             var = rename_it->second;
 804             ctx.renames[pred_idx].erase(rename_it);
 805          }
 806
 807          uint32_t spill_id = ctx.allocate_spill_id(phi->definitions[0].regClass());
 808          ctx.add_affinity(def_spill_id, spill_id);
 809          aco_ptr<Pseudo_instruction> spill{create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
 810          spill->operands[0] = Operand(var);
 811          spill->operands[1] = Operand(spill_id);
 812          Block& pred = ctx.program->blocks[pred_idx];
 813          unsigned idx = pred.instructions.size();
 814          do {
 815             assert(idx != 0);
 816             idx--;
 817          } while (phi->opcode == aco_opcode::p_phi && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
 818          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
 819          pred.instructions.insert(it, std::move(spill));
 820          ctx.spills_exit[pred_idx][phi->operands[i].getTemp()] = spill_id;
 821       }
 822
 823       /* remove phi from instructions */
 824       phi.reset();
 825    }
 826
 827    /* iterate all (other) spilled variables for which to spill at the predecessor */
 828    // TODO: would be better to have them sorted: first vgprs and first with longest distance
 829    for (std::pair<Temp, uint32_t> pair : ctx.spills_entry[block_idx]) {
 830       std::vector<unsigned> preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds;
 831
 832       for (unsigned pred_idx : preds) {
 833          /* variable is already spilled at predecessor */
 834          std::map<Temp, uint32_t>::iterator spilled = ctx.spills_exit[pred_idx].find(pair.first);
 835          if (spilled != ctx.spills_exit[pred_idx].end()) {
 836             if (spilled->second != pair.second)
 837                ctx.add_affinity(pair.second, spilled->second);
 838             continue;
 839          }
 840
 841          /* variable is dead at predecessor, it must be from a phi: this works because of CSSA form */
 842          if (ctx.next_use_distances_end[pred_idx].find(pair.first) == ctx.next_use_distances_end[pred_idx].end())
 843             continue;
 844
 845          /* add interferences between spilled variable and predecessors exit spills */
 846          for (std::pair<Temp, uint32_t> exit_spill : ctx.spills_exit[pred_idx]) {
 847             if (exit_spill.first == pair.first)
 848                continue;
 849             ctx.interferences[exit_spill.second].second.emplace(pair.second);
 850             ctx.interferences[pair.second].second.emplace(exit_spill.second);
 851          }
 852
 853          /* variable is in register at predecessor and has to be spilled */
 854          /* rename if necessary */
 855          Temp var = pair.first;
 856          std::map<Temp, Temp>::iterator rename_it = ctx.renames[pred_idx].find(var);
 857          if (rename_it != ctx.renames[pred_idx].end()) {
 858             var = rename_it->second;
 859             ctx.renames[pred_idx].erase(rename_it);
 860          }
 861
 862          aco_ptr<Pseudo_instruction> spill{create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
 863          spill->operands[0] = Operand(var);
 864          spill->operands[1] = Operand(pair.second);
 865          Block& pred = ctx.program->blocks[pred_idx];
 866          unsigned idx = pred.instructions.size();
 867          do {
 868             assert(idx != 0);
 869             idx--;
 870          } while (pair.first.type() == RegType::vgpr && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
 871          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
 872          pred.instructions.insert(it, std::move(spill));
 873          ctx.spills_exit[pred.index][pair.first] = pair.second;
 874       }
 875    }
 876
 877    /* iterate phis for which operands to reload */
 878    for (aco_ptr<Instruction>& phi : instructions) {
 879       assert(phi->opcode == aco_opcode::p_phi || phi->opcode == aco_opcode::p_linear_phi);
 880       assert(ctx.spills_entry[block_idx].find(phi->definitions[0].getTemp()) == ctx.spills_entry[block_idx].end());
 881
 882       std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
 883       for (unsigned i = 0; i < phi->operands.size(); i++) {
 884          if (!phi->operands[i].isTemp())
 885             continue;
 886          unsigned pred_idx = preds[i];
 887
 888          /* rename operand */
 889          if (ctx.spills_exit[pred_idx].find(phi->operands[i].getTemp()) == ctx.spills_exit[pred_idx].end()) {
 890             std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(phi->operands[i].getTemp());
 891             if (it != ctx.renames[pred_idx].end())
 892                phi->operands[i].setTemp(it->second);
 893             continue;
 894          }
 895
 896          Temp tmp = phi->operands[i].getTemp();
 897
 898          /* reload phi operand at end of predecessor block */
 899          Temp new_name = {ctx.program->allocateId(), tmp.regClass()};
 900          Block& pred = ctx.program->blocks[pred_idx];
 901          unsigned idx = pred.instructions.size();
 902          do {
 903             assert(idx != 0);
 904             idx--;
 905          } while (phi->opcode == aco_opcode::p_phi && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
 906          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
 907
 908          aco_ptr<Instruction> reload = do_reload(ctx, tmp, new_name, ctx.spills_exit[pred_idx][tmp]);
 909          pred.instructions.insert(it, std::move(reload));
 910
 911          ctx.spills_exit[pred_idx].erase(tmp);
 912          ctx.renames[pred_idx][tmp] = new_name;
 913          phi->operands[i].setTemp(new_name);
 914       }
 915    }
 916
 917    /* iterate live variables for which to reload */
 918    // TODO: reload at current block if variable is spilled on all predecessors
 919    for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
 920       /* skip spilled variables */
 921       if (ctx.spills_entry[block_idx].find(pair.first) != ctx.spills_entry[block_idx].end())
 922          continue;
 923       std::vector<unsigned> preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds;
 924
 925       /* variable is dead at predecessor, it must be from a phi */
 926       bool is_dead = false;
 927       for (unsigned pred_idx : preds) {
 928          if (ctx.next_use_distances_end[pred_idx].find(pair.first) == ctx.next_use_distances_end[pred_idx].end())
 929             is_dead = true;
 930       }
 931       if (is_dead)
 932          continue;
 933       for (unsigned pred_idx : preds) {
 934          /* the variable is not spilled at the predecessor */
 935          if (ctx.spills_exit[pred_idx].find(pair.first) == ctx.spills_exit[pred_idx].end())
 936             continue;
 937
 938          /* variable is spilled at predecessor and has to be reloaded */
 939          Temp new_name = {ctx.program->allocateId(), pair.first.regClass()};
 940          Block& pred = ctx.program->blocks[pred_idx];
 941          unsigned idx = pred.instructions.size();
 942          do {
 943             assert(idx != 0);
 944             idx--;
 945          } while (pair.first.type() == RegType::vgpr && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
 946          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
 947
 948          aco_ptr<Instruction> reload = do_reload(ctx, pair.first, new_name, ctx.spills_exit[pred.index][pair.first]);
 949          pred.instructions.insert(it, std::move(reload));
 950
 951          ctx.spills_exit[pred.index].erase(pair.first);
 952          ctx.renames[pred.index][pair.first] = new_name;
 953       }
 954
 955       /* check if we have to create a new phi for this variable */
 956       Temp rename = Temp();
 957       bool is_same = true;
 958       for (unsigned pred_idx : preds) {
 959          if (ctx.renames[pred_idx].find(pair.first) == ctx.renames[pred_idx].end()) {
 960             if (rename == Temp())
 961                rename = pair.first;
 962             else
 963                is_same = rename == pair.first;
 964          } else {
 965             if (rename == Temp())
 966                rename = ctx.renames[pred_idx][pair.first];
 967             else
 968                is_same = rename == ctx.renames[pred_idx][pair.first];
 969          }
 970
 971          if (!is_same)
 972             break;
 973       }
 974
 975       if (!is_same) {
 976          /* the variable was renamed differently in the predecessors: we have to create a phi */
 977          aco_opcode opcode = pair.first.is_linear() ? aco_opcode::p_linear_phi : aco_opcode::p_phi;
 978          aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(opcode, Format::PSEUDO, preds.size(), 1)};
 979          rename = {ctx.program->allocateId(), pair.first.regClass()};
 980          for (unsigned i = 0; i < phi->operands.size(); i++) {
 981             Temp tmp;
 982             if (ctx.renames[preds[i]].find(pair.first) != ctx.renames[preds[i]].end())
 983                tmp = ctx.renames[preds[i]][pair.first];
 984             else if (preds[i] >= block_idx)
 985                tmp = rename;
 986             else
 987                tmp = pair.first;
 988             phi->operands[i] = Operand(tmp);
 989          }
 990          phi->definitions[0] = Definition(rename);
 991          instructions.emplace_back(std::move(phi));
 992       }
 993
 994       /* the variable was renamed: add new name to renames */
 995       if (!(rename == Temp() || rename == pair.first))
 996          ctx.renames[block_idx][pair.first] = rename;
 997    }
 998
 999    /* combine phis with instructions */
1000    unsigned idx = 0;
1001    while (!block->instructions[idx]) {
1002       idx++;
1003    }
1004
1005    if (!ctx.processed[block_idx]) {
1006       assert(!(block->kind & block_kind_loop_header));
1007       RegisterDemand demand_before = get_demand_before(ctx, block_idx, idx);
1008       ctx.register_demand[block->index].erase(ctx.register_demand[block->index].begin(), ctx.register_demand[block->index].begin() + idx);
1009       ctx.register_demand[block->index].insert(ctx.register_demand[block->index].begin(), instructions.size(), demand_before);
1010    }
1011
1012    std::vector<aco_ptr<Instruction>>::iterator start = std::next(block->instructions.begin(), idx);
1013    instructions.insert(instructions.end(), std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(start),
1014                std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
1015    block->instructions = std::move(instructions);
1016 }
1017
1018 void process_block(spill_ctx& ctx, unsigned block_idx, Block* block,
1019                    std::map<Temp, uint32_t> &current_spills, RegisterDemand spilled_registers)
1020 {
1021    assert(!ctx.processed[block_idx]);
1022
1023    std::vector<std::map<Temp, uint32_t>> local_next_use_distance;
1024    std::vector<aco_ptr<Instruction>> instructions;
1025    unsigned idx = 0;
1026
1027    /* phis are handled separetely */
1028    while (block->instructions[idx]->opcode == aco_opcode::p_phi ||
1029           block->instructions[idx]->opcode == aco_opcode::p_linear_phi) {
1030       aco_ptr<Instruction>& instr = block->instructions[idx];
1031       for (const Operand& op : instr->operands) {
1032          /* prevent it's definining instruction from being DCE'd if it could be rematerialized */
1033          if (op.isTemp() && ctx.remat.count(op.getTemp()))
1034             ctx.remat_used[ctx.remat[op.getTemp()].instr] = true;
1035       }
1036       instructions.emplace_back(std::move(instr));
1037       idx++;
1038    }
1039
1040    if (block->register_demand.exceeds(ctx.target_pressure))
1041       local_next_use_distance = local_next_uses(ctx, block);
1042
1043    while (idx < block->instructions.size()) {
1044       aco_ptr<Instruction>& instr = block->instructions[idx];
1045
1046       std::map<Temp, std::pair<Temp, uint32_t>> reloads;
1047       std::map<Temp, uint32_t> spills;
1048       /* rename and reload operands */
1049       for (Operand& op : instr->operands) {
1050          if (!op.isTemp())
1051             continue;
1052          if (current_spills.find(op.getTemp()) == current_spills.end()) {
1053             /* the Operand is in register: check if it was renamed */
1054             if (ctx.renames[block_idx].find(op.getTemp()) != ctx.renames[block_idx].end())
1055                op.setTemp(ctx.renames[block_idx][op.getTemp()]);
1056             /* prevent it's definining instruction from being DCE'd if it could be rematerialized */
1057             if (ctx.remat.count(op.getTemp()))
1058                ctx.remat_used[ctx.remat[op.getTemp()].instr] = true;
1059             continue;
1060          }
1061          /* the Operand is spilled: add it to reloads */
1062          Temp new_tmp = {ctx.program->allocateId(), op.regClass()};
1063          ctx.renames[block_idx][op.getTemp()] = new_tmp;
1064          reloads[new_tmp] = std::make_pair(op.getTemp(), current_spills[op.getTemp()]);
1065          current_spills.erase(op.getTemp());
1066          op.setTemp(new_tmp);
1067          spilled_registers -= new_tmp;
1068       }
1069
1070       /* check if register demand is low enough before and after the current instruction */
1071       if (block->register_demand.exceeds(ctx.target_pressure)) {
1072
1073          RegisterDemand new_demand = ctx.register_demand[block_idx][idx];
1074          new_demand.update(get_demand_before(ctx, block_idx, idx));
1075
1076          assert(!local_next_use_distance.empty());
1077
1078          /* if reg pressure is too high, spill variable with furthest next use */
1079          while (RegisterDemand(new_demand - spilled_registers).exceeds(ctx.target_pressure)) {
1080             unsigned distance = 0;
1081             Temp to_spill;
1082             bool do_rematerialize = false;
1083             if (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) {
1084                for (std::pair<Temp, uint32_t> pair : local_next_use_distance[idx]) {
1085                   bool can_rematerialize = ctx.remat.count(pair.first);
1086                   if (pair.first.type() == RegType::vgpr &&
1087                       ((pair.second > distance && can_rematerialize == do_rematerialize) ||
1088                        (can_rematerialize && !do_rematerialize && pair.second > idx)) &&
1089                       current_spills.find(pair.first) == current_spills.end() &&
1090                       ctx.spills_exit[block_idx].find(pair.first) == ctx.spills_exit[block_idx].end()) {
1091                      to_spill = pair.first;
1092                      distance = pair.second;
1093                      do_rematerialize = can_rematerialize;
1094                   }
1095                }
1096             } else {
1097                for (std::pair<Temp, uint32_t> pair : local_next_use_distance[idx]) {
1098                   bool can_rematerialize = ctx.remat.count(pair.first);
1099                   if (pair.first.type() == RegType::sgpr &&
1100                       ((pair.second > distance && can_rematerialize == do_rematerialize) ||
1101                        (can_rematerialize && !do_rematerialize && pair.second > idx)) &&
1102                       current_spills.find(pair.first) == current_spills.end() &&
1103                       ctx.spills_exit[block_idx].find(pair.first) == ctx.spills_exit[block_idx].end()) {
1104                      to_spill = pair.first;
1105                      distance = pair.second;
1106                      do_rematerialize = can_rematerialize;
1107                   }
1108                }
1109             }
1110
1111             assert(distance != 0 && distance > idx);
1112             uint32_t spill_id = ctx.allocate_spill_id(to_spill.regClass());
1113
1114             /* add interferences with currently spilled variables */
1115             for (std::pair<Temp, uint32_t> pair : current_spills) {
1116                ctx.interferences[spill_id].second.emplace(pair.second);
1117                ctx.interferences[pair.second].second.emplace(spill_id);
1118             }
1119             for (std::pair<Temp, std::pair<Temp, uint32_t>> pair : reloads) {
1120                ctx.interferences[spill_id].second.emplace(pair.second.second);
1121                ctx.interferences[pair.second.second].second.emplace(spill_id);
1122             }
1123
1124             current_spills[to_spill] = spill_id;
1125             spilled_registers += to_spill;
1126
1127             /* rename if necessary */
1128             if (ctx.renames[block_idx].find(to_spill) != ctx.renames[block_idx].end()) {
1129                to_spill = ctx.renames[block_idx][to_spill];
1130             }
1131
1132             /* add spill to new instructions */
1133             aco_ptr<Pseudo_instruction> spill{create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
1134             spill->operands[0] = Operand(to_spill);
1135             spill->operands[1] = Operand(spill_id);
1136             instructions.emplace_back(std::move(spill));
1137          }
1138       }
1139
1140       /* add reloads and instruction to new instructions */
1141       for (std::pair<Temp, std::pair<Temp, uint32_t>> pair : reloads) {
1142          aco_ptr<Instruction> reload = do_reload(ctx, pair.second.first, pair.first, pair.second.second);
1143          instructions.emplace_back(std::move(reload));
1144       }
1145       instructions.emplace_back(std::move(instr));
1146       idx++;
1147    }
1148
1149    block->instructions = std::move(instructions);
1150    ctx.spills_exit[block_idx].insert(current_spills.begin(), current_spills.end());
1151 }
1152
1153 void spill_block(spill_ctx& ctx, unsigned block_idx)
1154 {
1155    Block* block = &ctx.program->blocks[block_idx];
1156
1157    /* determine set of variables which are spilled at the beginning of the block */
1158    RegisterDemand spilled_registers = init_live_in_vars(ctx, block, block_idx);
1159
1160    /* add interferences for spilled variables */
1161    for (std::pair<Temp, uint32_t> x : ctx.spills_entry[block_idx]) {
1162       for (std::pair<Temp, uint32_t> y : ctx.spills_entry[block_idx])
1163          if (x.second != y.second)
1164             ctx.interferences[x.second].second.emplace(y.second);
1165    }
1166
1167    bool is_loop_header = block->loop_nest_depth && ctx.loop_header.top()->index == block_idx;
1168    if (!is_loop_header) {
1169       /* add spill/reload code on incoming control flow edges */
1170       add_coupling_code(ctx, block, block_idx);
1171    }
1172
1173    std::map<Temp, uint32_t> current_spills = ctx.spills_entry[block_idx];
1174
1175    /* check conditions to process this block */
1176    bool process = RegisterDemand(block->register_demand - spilled_registers).exceeds(ctx.target_pressure) ||
1177                   !ctx.renames[block_idx].empty() ||
1178                   ctx.remat_used.size();
1179
1180    std::map<Temp, uint32_t>::iterator it = current_spills.begin();
1181    while (!process && it != current_spills.end()) {
1182       if (ctx.next_use_distances_start[block_idx][it->first].first == block_idx)
1183          process = true;
1184       ++it;
1185    }
1186
1187    if (process)
1188       process_block(ctx, block_idx, block, current_spills, spilled_registers);
1189    else
1190       ctx.spills_exit[block_idx].insert(current_spills.begin(), current_spills.end());
1191
1192    ctx.processed[block_idx] = true;
1193
1194    /* check if the next block leaves the current loop */
1195    if (block->loop_nest_depth == 0 || ctx.program->blocks[block_idx + 1].loop_nest_depth >= block->loop_nest_depth)
1196       return;
1197
1198    Block* loop_header = ctx.loop_header.top();
1199
1200    /* preserve original renames at end of loop header block */
1201    std::map<Temp, Temp> renames = std::move(ctx.renames[loop_header->index]);
1202
1203    /* add coupling code to all loop header predecessors */
1204    add_coupling_code(ctx, loop_header, loop_header->index);
1205
1206    /* update remat_used for phis added in add_coupling_code() */
1207    for (aco_ptr<Instruction>& instr : loop_header->instructions) {
1208       if (!is_phi(instr))
1209          break;
1210       for (const Operand& op : instr->operands) {
1211          if (op.isTemp() && ctx.remat.count(op.getTemp()))
1212             ctx.remat_used[ctx.remat[op.getTemp()].instr] = true;
1213       }
1214    }
1215
1216    /* propagate new renames through loop: i.e. repair the SSA */
1217    renames.swap(ctx.renames[loop_header->index]);
1218    for (std::pair<Temp, Temp> rename : renames) {
1219       for (unsigned idx = loop_header->index; idx <= block_idx; idx++) {
1220          Block& current = ctx.program->blocks[idx];
1221          std::vector<aco_ptr<Instruction>>::iterator instr_it = current.instructions.begin();
1222
1223          /* first rename phis */
1224          while (instr_it != current.instructions.end()) {
1225             aco_ptr<Instruction>& phi = *instr_it;
1226             if (phi->opcode != aco_opcode::p_phi && phi->opcode != aco_opcode::p_linear_phi)
1227                break;
1228             /* no need to rename the loop header phis once again. this happened in add_coupling_code() */
1229             if (idx == loop_header->index) {
1230                instr_it++;
1231                continue;
1232             }
1233
1234             for (Operand& op : phi->operands) {
1235                if (!op.isTemp())
1236                   continue;
1237                if (op.getTemp() == rename.first)
1238                   op.setTemp(rename.second);
1239             }
1240             instr_it++;
1241          }
1242
1243          std::map<Temp, std::pair<uint32_t, uint32_t>>::iterator it = ctx.next_use_distances_start[idx].find(rename.first);
1244
1245          /* variable is not live at beginning of this block */
1246          if (it == ctx.next_use_distances_start[idx].end())
1247             continue;
1248
1249          /* if the variable is live at the block's exit, add rename */
1250          if (ctx.next_use_distances_end[idx].find(rename.first) != ctx.next_use_distances_end[idx].end())
1251             ctx.renames[idx].insert(rename);
1252
1253          /* rename all uses in this block */
1254          bool renamed = false;
1255          while (!renamed && instr_it != current.instructions.end()) {
1256             aco_ptr<Instruction>& instr = *instr_it;
1257             for (Operand& op : instr->operands) {
1258                if (!op.isTemp())
1259                   continue;
1260                if (op.getTemp() == rename.first) {
1261                   op.setTemp(rename.second);
1262                   /* we can stop with this block as soon as the variable is spilled */
1263                   if (instr->opcode == aco_opcode::p_spill)
1264                     renamed = true;
1265                }
1266             }
1267             instr_it++;
1268          }
1269       }
1270    }
1271
1272    /* remove loop header info from stack */
1273    ctx.loop_header.pop();
1274 }
1275
1276 Temp load_scratch_resource(spill_ctx& ctx, Temp& scratch_offset,
1277                            std::vector<aco_ptr<Instruction>>& instructions,
1278                            unsigned offset, bool is_top_level)
1279 {
1280    Builder bld(ctx.program);
1281    if (is_top_level) {
1282       bld.reset(&instructions);
1283    } else {
1284       /* find p_logical_end */
1285       unsigned idx = instructions.size() - 1;
1286       while (instructions[idx]->opcode != aco_opcode::p_logical_end)
1287          idx--;
1288       bld.reset(&instructions, std::next(instructions.begin(), idx));
1289    }
1290
1291    Temp private_segment_buffer = ctx.program->private_segment_buffer;
1292    if (ctx.program->stage != compute_cs)
1293       private_segment_buffer = bld.smem(aco_opcode::s_load_dwordx2, bld.def(s2), private_segment_buffer, Operand(0u));
1294
1295    if (offset)
1296       scratch_offset = bld.sop2(aco_opcode::s_add_u32, bld.def(s1), bld.def(s1, scc), scratch_offset, Operand(offset));
1297
1298    uint32_t rsrc_conf = S_008F0C_ADD_TID_ENABLE(1) |
1299                         S_008F0C_INDEX_STRIDE(ctx.program->wave_size == 64 ? 3 : 2);
1300
1301    if (ctx.program->chip_class >= GFX10) {
1302       rsrc_conf |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
1303                    S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
1304                    S_008F0C_RESOURCE_LEVEL(1);
1305    } else if (ctx.program->chip_class <= GFX7) { /* dfmt modifies stride on GFX8/GFX9 when ADD_TID_EN=1 */
1306       rsrc_conf |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
1307                    S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
1308    }
1309    /* older generations need element size = 4 bytes. element size removed in GFX9 */
1310    if (ctx.program->chip_class <= GFX8)
1311       rsrc_conf |= S_008F0C_ELEMENT_SIZE(1);
1312
1313    return bld.pseudo(aco_opcode::p_create_vector, bld.def(s4),
1314                      private_segment_buffer, Operand(-1u),
1315                      Operand(rsrc_conf));
1316 }
1317
1318 void assign_spill_slots(spill_ctx& ctx, unsigned spills_to_vgpr) {
1319    std::map<uint32_t, uint32_t> sgpr_slot;
1320    std::map<uint32_t, uint32_t> vgpr_slot;
1321    std::vector<bool> is_assigned(ctx.interferences.size());
1322
1323    /* first, handle affinities: just merge all interferences into both spill ids */
1324    for (std::vector<uint32_t>& vec : ctx.affinities) {
1325       for (unsigned i = 0; i < vec.size(); i++) {
1326          for (unsigned j = i + 1; j < vec.size(); j++) {
1327             assert(vec[i] != vec[j]);
1328             for (uint32_t id : ctx.interferences[vec[i]].second)
1329                ctx.interferences[id].second.insert(vec[j]);
1330             for (uint32_t id : ctx.interferences[vec[j]].second)
1331                ctx.interferences[id].second.insert(vec[i]);
1332             ctx.interferences[vec[i]].second.insert(ctx.interferences[vec[j]].second.begin(), ctx.interferences[vec[j]].second.end());
1333             ctx.interferences[vec[j]].second.insert(ctx.interferences[vec[i]].second.begin(), ctx.interferences[vec[i]].second.end());
1334
1335             bool reloaded = ctx.is_reloaded[vec[i]] || ctx.is_reloaded[vec[j]];
1336             ctx.is_reloaded[vec[i]] = reloaded;
1337             ctx.is_reloaded[vec[j]] = reloaded;
1338          }
1339       }
1340    }
1341    for (ASSERTED uint32_t i = 0; i < ctx.interferences.size(); i++)
1342       for (ASSERTED uint32_t id : ctx.interferences[i].second)
1343          assert(i != id);
1344
1345    /* for each spill slot, assign as many spill ids as possible */
1346    std::vector<std::set<uint32_t>> spill_slot_interferences;
1347    unsigned slot_idx = 0;
1348    bool done = false;
1349
1350    /* assign sgpr spill slots */
1351    while (!done) {
1352       done = true;
1353       for (unsigned id = 0; id < ctx.interferences.size(); id++) {
1354          if (is_assigned[id] || !ctx.is_reloaded[id])
1355             continue;
1356          if (ctx.interferences[id].first.type() != RegType::sgpr)
1357             continue;
1358
1359          /* check interferences */
1360          bool interferes = false;
1361          for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++) {
1362             if (i == spill_slot_interferences.size())
1363                spill_slot_interferences.emplace_back(std::set<uint32_t>());
1364             if (spill_slot_interferences[i].find(id) != spill_slot_interferences[i].end() || i / ctx.wave_size != slot_idx / ctx.wave_size) {
1365                interferes = true;
1366                break;
1367             }
1368          }
1369          if (interferes) {
1370             done = false;
1371             continue;
1372          }
1373
1374          /* we found a spill id which can be assigned to current spill slot */
1375          sgpr_slot[id] = slot_idx;
1376          is_assigned[id] = true;
1377          for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++)
1378             spill_slot_interferences[i].insert(ctx.interferences[id].second.begin(), ctx.interferences[id].second.end());
1379
1380          /* add all affinities: there are no additional interferences */
1381          for (std::vector<uint32_t>& vec : ctx.affinities) {
1382             bool found_affinity = false;
1383             for (uint32_t entry : vec) {
1384                if (entry == id) {
1385                   found_affinity = true;
1386                   break;
1387                }
1388             }
1389             if (!found_affinity)
1390                continue;
1391             for (uint32_t entry : vec) {
1392                sgpr_slot[entry] = slot_idx;
1393                is_assigned[entry] = true;
1394             }
1395          }
1396       }
1397       slot_idx++;
1398    }
1399
1400    unsigned sgpr_spill_slots = spill_slot_interferences.size();
1401    spill_slot_interferences.clear();
1402    slot_idx = 0;
1403    done = false;
1404
1405    /* assign vgpr spill slots */
1406    while (!done) {
1407       done = true;
1408       for (unsigned id = 0; id < ctx.interferences.size(); id++) {
1409          if (is_assigned[id] || !ctx.is_reloaded[id])
1410             continue;
1411          if (ctx.interferences[id].first.type() != RegType::vgpr)
1412             continue;
1413
1414          /* check interferences */
1415          bool interferes = false;
1416          for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++) {
1417             if (i == spill_slot_interferences.size())
1418                spill_slot_interferences.emplace_back(std::set<uint32_t>());
1419             /* check for interference and ensure that vector regs are stored next to each other */
1420             if (spill_slot_interferences[i].find(id) != spill_slot_interferences[i].end()) {
1421                interferes = true;
1422                break;
1423             }
1424          }
1425          if (interferes) {
1426             done = false;
1427             continue;
1428          }
1429
1430          /* we found a spill id which can be assigned to current spill slot */
1431          vgpr_slot[id] = slot_idx;
1432          is_assigned[id] = true;
1433          for (unsigned i = slot_idx; i < slot_idx + ctx.interferences[id].first.size(); i++)
1434             spill_slot_interferences[i].insert(ctx.interferences[id].second.begin(), ctx.interferences[id].second.end());
1435
1436          /* add all affinities: there are no additional interferences */
1437          for (std::vector<uint32_t>& vec : ctx.affinities) {
1438             bool found_affinity = false;
1439             for (uint32_t entry : vec) {
1440                if (entry == id) {
1441                   found_affinity = true;
1442                   break;
1443                }
1444             }
1445             if (!found_affinity)
1446                continue;
1447             for (uint32_t entry : vec) {
1448                vgpr_slot[entry] = slot_idx;
1449                is_assigned[entry] = true;
1450             }
1451          }
1452       }
1453       slot_idx++;
1454    }
1455
1456    unsigned vgpr_spill_slots = spill_slot_interferences.size();
1457
1458    for (unsigned id = 0; id < is_assigned.size(); id++)
1459       assert(is_assigned[id] || !ctx.is_reloaded[id]);
1460
1461    for (std::vector<uint32_t>& vec : ctx.affinities) {
1462       for (unsigned i = 0; i < vec.size(); i++) {
1463          for (unsigned j = i + 1; j < vec.size(); j++) {
1464             assert(is_assigned[vec[i]] == is_assigned[vec[j]]);
1465             if (!is_assigned[vec[i]])
1466                continue;
1467             assert(ctx.is_reloaded[vec[i]] == ctx.is_reloaded[vec[j]]);
1468             assert(ctx.interferences[vec[i]].first.type() == ctx.interferences[vec[j]].first.type());
1469             if (ctx.interferences[vec[i]].first.type() == RegType::sgpr)
1470                assert(sgpr_slot[vec[i]] == sgpr_slot[vec[j]]);
1471             else
1472                assert(vgpr_slot[vec[i]] == vgpr_slot[vec[j]]);
1473          }
1474       }
1475    }
1476
1477    /* hope, we didn't mess up */
1478    std::vector<Temp> vgpr_spill_temps((sgpr_spill_slots + ctx.wave_size - 1) / ctx.wave_size);
1479    assert(vgpr_spill_temps.size() <= spills_to_vgpr);
1480
1481    /* replace pseudo instructions with actual hardware instructions */
1482    Temp scratch_offset = ctx.program->scratch_offset, scratch_rsrc = Temp();
1483    unsigned last_top_level_block_idx = 0;
1484    std::vector<bool> reload_in_loop(vgpr_spill_temps.size());
1485    for (Block& block : ctx.program->blocks) {
1486
1487       /* after loops, we insert a user if there was a reload inside the loop */
1488       if (block.loop_nest_depth == 0) {
1489          int end_vgprs = 0;
1490          for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1491             if (reload_in_loop[i])
1492                end_vgprs++;
1493          }
1494
1495          if (end_vgprs > 0) {
1496             aco_ptr<Instruction> destr{create_instruction<Pseudo_instruction>(aco_opcode::p_end_linear_vgpr, Format::PSEUDO, end_vgprs, 0)};
1497             int k = 0;
1498             for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1499                if (reload_in_loop[i])
1500                   destr->operands[k++] = Operand(vgpr_spill_temps[i]);
1501                reload_in_loop[i] = false;
1502             }
1503             /* find insertion point */
1504             std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
1505             while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
1506                ++it;
1507             block.instructions.insert(it, std::move(destr));
1508          }
1509       }
1510
1511       if (block.kind & block_kind_top_level && !block.linear_preds.empty()) {
1512          last_top_level_block_idx = block.index;
1513
1514          /* check if any spilled variables use a created linear vgpr, otherwise destroy them */
1515          for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1516             if (vgpr_spill_temps[i] == Temp())
1517                continue;
1518
1519             bool can_destroy = true;
1520             for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[block.linear_preds[0]]) {
1521
1522                if (sgpr_slot.find(pair.second) != sgpr_slot.end() &&
1523                    sgpr_slot[pair.second] / ctx.wave_size == i) {
1524                   can_destroy = false;
1525                   break;
1526                }
1527             }
1528             if (can_destroy)
1529                vgpr_spill_temps[i] = Temp();
1530          }
1531       }
1532
1533       std::vector<aco_ptr<Instruction>>::iterator it;
1534       std::vector<aco_ptr<Instruction>> instructions;
1535       instructions.reserve(block.instructions.size());
1536       Builder bld(ctx.program, &instructions);
1537       for (it = block.instructions.begin(); it != block.instructions.end(); ++it) {
1538
1539          if ((*it)->opcode == aco_opcode::p_spill) {
1540             uint32_t spill_id = (*it)->operands[1].constantValue();
1541
1542             if (!ctx.is_reloaded[spill_id]) {
1543                /* never reloaded, so don't spill */
1544             } else if (vgpr_slot.find(spill_id) != vgpr_slot.end()) {
1545                /* spill vgpr */
1546                ctx.program->config->spilled_vgprs += (*it)->operands[0].size();
1547                uint32_t spill_slot = vgpr_slot[spill_id];
1548                bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096;
1549                unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
1550
1551                /* check if the scratch resource descriptor already exists */
1552                if (scratch_rsrc == Temp()) {
1553                   unsigned offset = add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
1554                   scratch_rsrc = load_scratch_resource(ctx, scratch_offset,
1555                                                        last_top_level_block_idx == block.index ?
1556                                                        instructions : ctx.program->blocks[last_top_level_block_idx].instructions,
1557                                                        offset,
1558                                                        last_top_level_block_idx == block.index);
1559                }
1560
1561                unsigned offset = base_offset + spill_slot * 4;
1562                aco_opcode opcode = aco_opcode::buffer_store_dword;
1563                assert((*it)->operands[0].isTemp());
1564                Temp temp = (*it)->operands[0].getTemp();
1565                assert(temp.type() == RegType::vgpr && !temp.is_linear());
1566                if (temp.size() > 1) {
1567                   Instruction* split{create_instruction<Pseudo_instruction>(aco_opcode::p_split_vector, Format::PSEUDO, 1, temp.size())};
1568                   split->operands[0] = Operand(temp);
1569                   for (unsigned i = 0; i < temp.size(); i++)
1570                      split->definitions[i] = bld.def(v1);
1571                   bld.insert(split);
1572                   for (unsigned i = 0; i < temp.size(); i++)
1573                      bld.mubuf(opcode, scratch_rsrc, Operand(), scratch_offset, split->definitions[i].getTemp(), offset + i * 4, false);
1574                } else {
1575                   bld.mubuf(opcode, scratch_rsrc, Operand(), scratch_offset, temp, offset, false);
1576                }
1577             } else if (sgpr_slot.find(spill_id) != sgpr_slot.end()) {
1578                ctx.program->config->spilled_sgprs += (*it)->operands[0].size();
1579
1580                uint32_t spill_slot = sgpr_slot[spill_id];
1581
1582                /* check if the linear vgpr already exists */
1583                if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) {
1584                   Temp linear_vgpr = {ctx.program->allocateId(), v1.as_linear()};
1585                   vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr;
1586                   aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
1587                   create->definitions[0] = Definition(linear_vgpr);
1588                   /* find the right place to insert this definition */
1589                   if (last_top_level_block_idx == block.index) {
1590                      /* insert right before the current instruction */
1591                      instructions.emplace_back(std::move(create));
1592                   } else {
1593                      assert(last_top_level_block_idx < block.index);
1594                      /* insert before the branch at last top level block */
1595                      std::vector<aco_ptr<Instruction>>& instructions = ctx.program->blocks[last_top_level_block_idx].instructions;
1596                      instructions.insert(std::next(instructions.begin(), instructions.size() - 1), std::move(create));
1597                   }
1598                }
1599
1600                /* spill sgpr: just add the vgpr temp to operands */
1601                Pseudo_instruction* spill = create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 3, 0);
1602                spill->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]);
1603                spill->operands[1] = Operand(spill_slot % ctx.wave_size);
1604                spill->operands[2] = (*it)->operands[0];
1605                instructions.emplace_back(aco_ptr<Instruction>(spill));
1606             } else {
1607                unreachable("No spill slot assigned for spill id");
1608             }
1609
1610          } else if ((*it)->opcode == aco_opcode::p_reload) {
1611             uint32_t spill_id = (*it)->operands[0].constantValue();
1612             assert(ctx.is_reloaded[spill_id]);
1613
1614             if (vgpr_slot.find(spill_id) != vgpr_slot.end()) {
1615                /* reload vgpr */
1616                uint32_t spill_slot = vgpr_slot[spill_id];
1617                bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096;
1618                unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
1619
1620                /* check if the scratch resource descriptor already exists */
1621                if (scratch_rsrc == Temp()) {
1622                   unsigned offset = add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
1623                   scratch_rsrc = load_scratch_resource(ctx, scratch_offset,
1624                                                        last_top_level_block_idx == block.index ?
1625                                                        instructions : ctx.program->blocks[last_top_level_block_idx].instructions,
1626                                                        offset,
1627                                                        last_top_level_block_idx == block.index);
1628                }
1629
1630                unsigned offset = base_offset + spill_slot * 4;
1631                aco_opcode opcode = aco_opcode::buffer_load_dword;
1632                Definition def = (*it)->definitions[0];
1633                if (def.size() > 1) {
1634                   Instruction* vec{create_instruction<Pseudo_instruction>(aco_opcode::p_create_vector, Format::PSEUDO, def.size(), 1)};
1635                   vec->definitions[0] = def;
1636                   for (unsigned i = 0; i < def.size(); i++) {
1637                      Temp tmp = bld.tmp(v1);
1638                      vec->operands[i] = Operand(tmp);
1639                      bld.mubuf(opcode, Definition(tmp), scratch_rsrc, Operand(), scratch_offset, offset + i * 4, false);
1640                   }
1641                   bld.insert(vec);
1642                } else {
1643                   bld.mubuf(opcode, def, scratch_rsrc, Operand(), scratch_offset, offset, false);
1644                }
1645             } else if (sgpr_slot.find(spill_id) != sgpr_slot.end()) {
1646                uint32_t spill_slot = sgpr_slot[spill_id];
1647                reload_in_loop[spill_slot / ctx.wave_size] = block.loop_nest_depth > 0;
1648
1649                /* check if the linear vgpr already exists */
1650                if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) {
1651                   Temp linear_vgpr = {ctx.program->allocateId(), v1.as_linear()};
1652                   vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr;
1653                   aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
1654                   create->definitions[0] = Definition(linear_vgpr);
1655                   /* find the right place to insert this definition */
1656                   if (last_top_level_block_idx == block.index) {
1657                      /* insert right before the current instruction */
1658                      instructions.emplace_back(std::move(create));
1659                   } else {
1660                      assert(last_top_level_block_idx < block.index);
1661                      /* insert before the branch at last top level block */
1662                      std::vector<aco_ptr<Instruction>>& instructions = ctx.program->blocks[last_top_level_block_idx].instructions;
1663                      instructions.insert(std::next(instructions.begin(), instructions.size() - 1), std::move(create));
1664                   }
1665                }
1666
1667                /* reload sgpr: just add the vgpr temp to operands */
1668                Pseudo_instruction* reload = create_instruction<Pseudo_instruction>(aco_opcode::p_reload, Format::PSEUDO, 2, 1);
1669                reload->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]);
1670                reload->operands[1] = Operand(spill_slot % ctx.wave_size);
1671                reload->definitions[0] = (*it)->definitions[0];
1672                instructions.emplace_back(aco_ptr<Instruction>(reload));
1673             } else {
1674                unreachable("No spill slot assigned for spill id");
1675             }
1676          } else if (!ctx.remat_used.count(it->get()) || ctx.remat_used[it->get()]) {
1677             instructions.emplace_back(std::move(*it));
1678          }
1679
1680       }
1681       block.instructions = std::move(instructions);
1682    }
1683
1684    /* update required scratch memory */
1685    ctx.program->config->scratch_bytes_per_wave += align(vgpr_spill_slots * 4 * ctx.program->wave_size, 1024);
1686
1687    /* SSA elimination inserts copies for logical phis right before p_logical_end
1688     * So if a linear vgpr is used between that p_logical_end and the branch,
1689     * we need to ensure logical phis don't choose a definition which aliases
1690     * the linear vgpr.
1691     * TODO: Moving the spills and reloads to before p_logical_end might produce
1692     *       slightly better code. */
1693    for (Block& block : ctx.program->blocks) {
1694       /* loops exits are already handled */
1695       if (block.logical_preds.size() <= 1)
1696          continue;
1697
1698       bool has_logical_phis = false;
1699       for (aco_ptr<Instruction>& instr : block.instructions) {
1700          if (instr->opcode == aco_opcode::p_phi) {
1701             has_logical_phis = true;
1702             break;
1703          } else if (instr->opcode != aco_opcode::p_linear_phi) {
1704             break;
1705          }
1706       }
1707       if (!has_logical_phis)
1708          continue;
1709
1710       std::set<Temp> vgprs;
1711       for (unsigned pred_idx : block.logical_preds) {
1712          Block& pred = ctx.program->blocks[pred_idx];
1713          for (int i = pred.instructions.size() - 1; i >= 0; i--) {
1714             aco_ptr<Instruction>& pred_instr = pred.instructions[i];
1715             if (pred_instr->opcode == aco_opcode::p_logical_end) {
1716                break;
1717             } else if (pred_instr->opcode == aco_opcode::p_spill ||
1718                        pred_instr->opcode == aco_opcode::p_reload) {
1719                vgprs.insert(pred_instr->operands[0].getTemp());
1720             }
1721          }
1722       }
1723       if (!vgprs.size())
1724          continue;
1725
1726       aco_ptr<Instruction> destr{create_instruction<Pseudo_instruction>(aco_opcode::p_end_linear_vgpr, Format::PSEUDO, vgprs.size(), 0)};
1727       int k = 0;
1728       for (Temp tmp : vgprs) {
1729          destr->operands[k++] = Operand(tmp);
1730       }
1731       /* find insertion point */
1732       std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
1733       while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
1734          ++it;
1735       block.instructions.insert(it, std::move(destr));
1736    }
1737 }
1738
1739 } /* end namespace */
1740
1741
1742 void spill(Program* program, live& live_vars, const struct radv_nir_compiler_options *options)
1743 {
1744    program->config->spilled_vgprs = 0;
1745    program->config->spilled_sgprs = 0;
1746
1747    /* no spilling when register pressure is low enough */
1748    if (program->num_waves > 0)
1749       return;
1750
1751    /* lower to CSSA before spilling to ensure correctness w.r.t. phis */
1752    lower_to_cssa(program, live_vars, options);
1753
1754    /* calculate target register demand */
1755    RegisterDemand register_target = program->max_reg_demand;
1756    if (register_target.sgpr > program->sgpr_limit)
1757       register_target.vgpr += (register_target.sgpr - program->sgpr_limit + program->wave_size - 1 + 32) / program->wave_size;
1758    register_target.sgpr = program->sgpr_limit;
1759
1760    if (register_target.vgpr > program->vgpr_limit)
1761       register_target.sgpr = program->sgpr_limit - 5;
1762    int spills_to_vgpr = (program->max_reg_demand.sgpr - register_target.sgpr + program->wave_size - 1 + 32) / program->wave_size;
1763    register_target.vgpr = program->vgpr_limit - spills_to_vgpr;
1764
1765    /* initialize ctx */
1766    spill_ctx ctx(register_target, program, live_vars.register_demand);
1767    compute_global_next_uses(ctx, live_vars.live_out);
1768    get_rematerialize_info(ctx);
1769
1770    /* create spills and reloads */
1771    for (unsigned i = 0; i < program->blocks.size(); i++)
1772       spill_block(ctx, i);
1773
1774    /* assign spill slots and DCE rematerialized code */
1775    assign_spill_slots(ctx, spills_to_vgpr);
1776
1777    /* update live variable information */
1778    live_vars = live_var_analysis(program, options);
1779
1780    assert(program->num_waves > 0);
1781 }
1782
1783 }
1784