/* * Copyright © 2018 Valve Corporation * Copyright © 2018 Google * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS * IN THE SOFTWARE. * */ #include "aco_builder.h" #include "aco_ir.h" #include "common/sid.h" #include #include #include #include #include #include #include #include namespace std { template <> struct hash { size_t operator()(aco::Temp temp) const noexcept { uint32_t v; std::memcpy(&v, &temp, sizeof(temp)); return std::hash{}(v); } }; } // namespace std /* * Implements the spilling algorithm on SSA-form from * "Register Spilling and Live-Range Splitting for SSA-Form Programs" * by Matthias Braun and Sebastian Hack. */ namespace aco { namespace { struct remat_info { Instruction* instr; }; struct spill_ctx { RegisterDemand target_pressure; Program* program; std::vector> register_demand; std::vector> renames; std::vector> spills_entry; std::vector> spills_exit; std::vector processed; std::stack> loop_header; std::vector>> next_use_distances_start; std::vector>> next_use_distances_end; std::vector>> local_next_use_distance; /* Working buffer */ std::vector>> interferences; std::vector> affinities; std::vector is_reloaded; std::unordered_map remat; std::set unused_remats; unsigned wave_size; spill_ctx(const RegisterDemand target_pressure_, Program* program_, std::vector> register_demand_) : target_pressure(target_pressure_), program(program_), register_demand(std::move(register_demand_)), renames(program->blocks.size()), spills_entry(program->blocks.size()), spills_exit(program->blocks.size()), processed(program->blocks.size(), false), wave_size(program->wave_size) {} void add_affinity(uint32_t first, uint32_t second) { unsigned found_first = affinities.size(); unsigned found_second = affinities.size(); for (unsigned i = 0; i < affinities.size(); i++) { std::vector& vec = affinities[i]; for (uint32_t entry : vec) { if (entry == first) found_first = i; else if (entry == second) found_second = i; } } if (found_first == affinities.size() && found_second == affinities.size()) { affinities.emplace_back(std::vector({first, second})); } else if (found_first < affinities.size() && found_second == affinities.size()) { affinities[found_first].push_back(second); } else if (found_second < affinities.size() && found_first == affinities.size()) { affinities[found_second].push_back(first); } else if (found_first != found_second) { /* merge second into first */ affinities[found_first].insert(affinities[found_first].end(), affinities[found_second].begin(), affinities[found_second].end()); affinities.erase(std::next(affinities.begin(), found_second)); } else { assert(found_first == found_second); } } void add_interference(uint32_t first, uint32_t second) { if (interferences[first].first.type() != interferences[second].first.type()) return; bool inserted = interferences[first].second.insert(second).second; if (inserted) interferences[second].second.insert(first); } uint32_t allocate_spill_id(RegClass rc) { interferences.emplace_back(rc, std::unordered_set()); is_reloaded.push_back(false); return next_spill_id++; } uint32_t next_spill_id = 0; }; int32_t get_dominator(int idx_a, int idx_b, Program* program, bool is_linear) { if (idx_a == -1) return idx_b; if (idx_b == -1) return idx_a; if (is_linear) { while (idx_a != idx_b) { if (idx_a > idx_b) idx_a = program->blocks[idx_a].linear_idom; else idx_b = program->blocks[idx_b].linear_idom; } } else { while (idx_a != idx_b) { if (idx_a > idx_b) idx_a = program->blocks[idx_a].logical_idom; else idx_b = program->blocks[idx_b].logical_idom; } } assert(idx_a != -1); return idx_a; } void next_uses_per_block(spill_ctx& ctx, unsigned block_idx, uint32_t& worklist) { Block* block = &ctx.program->blocks[block_idx]; ctx.next_use_distances_start[block_idx] = ctx.next_use_distances_end[block_idx]; auto& next_use_distances_start = ctx.next_use_distances_start[block_idx]; /* to compute the next use distance at the beginning of the block, we have to add the block's * size */ for (std::unordered_map>::iterator it = next_use_distances_start.begin(); it != next_use_distances_start.end(); ++it) it->second.second = it->second.second + block->instructions.size(); int idx = block->instructions.size() - 1; while (idx >= 0) { aco_ptr& instr = block->instructions[idx]; if (instr->opcode == aco_opcode::p_linear_phi || instr->opcode == aco_opcode::p_phi) break; for (const Definition& def : instr->definitions) { if (def.isTemp()) next_use_distances_start.erase(def.getTemp()); } for (const Operand& op : instr->operands) { /* omit exec mask */ if (op.isFixed() && op.physReg() == exec) continue; if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear()) continue; if (op.isTemp()) next_use_distances_start[op.getTemp()] = {block_idx, idx}; } idx--; } assert(block_idx != 0 || next_use_distances_start.empty()); std::unordered_set phi_defs; while (idx >= 0) { aco_ptr& instr = block->instructions[idx]; assert(instr->opcode == aco_opcode::p_linear_phi || instr->opcode == aco_opcode::p_phi); std::pair distance{block_idx, 0}; auto it = instr->definitions[0].isTemp() ? next_use_distances_start.find(instr->definitions[0].getTemp()) : next_use_distances_start.end(); if (it != next_use_distances_start.end() && phi_defs.insert(instr->definitions[0].getTemp()).second) { distance = it->second; } for (unsigned i = 0; i < instr->operands.size(); i++) { unsigned pred_idx = instr->opcode == aco_opcode::p_phi ? block->logical_preds[i] : block->linear_preds[i]; if (instr->operands[i].isTemp()) { auto insert_result = ctx.next_use_distances_end[pred_idx].insert( std::make_pair(instr->operands[i].getTemp(), distance)); const bool inserted = insert_result.second; std::pair& entry_distance = insert_result.first->second; if (inserted || entry_distance != distance) worklist = std::max(worklist, pred_idx + 1); entry_distance = distance; } } idx--; } /* all remaining live vars must be live-out at the predecessors */ for (std::pair>& pair : next_use_distances_start) { Temp temp = pair.first; if (phi_defs.count(temp)) { continue; } uint32_t distance = pair.second.second; uint32_t dom = pair.second.first; std::vector& preds = temp.is_linear() ? block->linear_preds : block->logical_preds; for (unsigned pred_idx : preds) { if (ctx.program->blocks[pred_idx].loop_nest_depth > block->loop_nest_depth) distance += 0xFFFF; auto insert_result = ctx.next_use_distances_end[pred_idx].insert( std::make_pair(temp, std::pair{})); const bool inserted = insert_result.second; std::pair& entry_distance = insert_result.first->second; if (!inserted) { dom = get_dominator(dom, entry_distance.first, ctx.program, temp.is_linear()); distance = std::min(entry_distance.second, distance); } if (entry_distance != std::pair{dom, distance}) { worklist = std::max(worklist, pred_idx + 1); entry_distance = {dom, distance}; } } } } void compute_global_next_uses(spill_ctx& ctx) { ctx.next_use_distances_start.resize(ctx.program->blocks.size()); ctx.next_use_distances_end.resize(ctx.program->blocks.size()); uint32_t worklist = ctx.program->blocks.size(); while (worklist) { unsigned block_idx = --worklist; next_uses_per_block(ctx, block_idx, worklist); } } bool should_rematerialize(aco_ptr& instr) { /* TODO: rematerialization is only supported for VOP1, SOP1 and PSEUDO */ if (instr->format != Format::VOP1 && instr->format != Format::SOP1 && instr->format != Format::PSEUDO && instr->format != Format::SOPK) return false; /* TODO: pseudo-instruction rematerialization is only supported for * p_create_vector/p_parallelcopy */ if (instr->isPseudo() && instr->opcode != aco_opcode::p_create_vector && instr->opcode != aco_opcode::p_parallelcopy) return false; if (instr->isSOPK() && instr->opcode != aco_opcode::s_movk_i32) return false; for (const Operand& op : instr->operands) { /* TODO: rematerialization using temporaries isn't yet supported */ if (!op.isConstant()) return false; } /* TODO: rematerialization with multiple definitions isn't yet supported */ if (instr->definitions.size() > 1) return false; return true; } aco_ptr do_reload(spill_ctx& ctx, Temp tmp, Temp new_name, uint32_t spill_id) { std::unordered_map::iterator remat = ctx.remat.find(tmp); if (remat != ctx.remat.end()) { Instruction* instr = remat->second.instr; assert((instr->isVOP1() || instr->isSOP1() || instr->isPseudo() || instr->isSOPK()) && "unsupported"); assert((instr->format != Format::PSEUDO || instr->opcode == aco_opcode::p_create_vector || instr->opcode == aco_opcode::p_parallelcopy) && "unsupported"); assert(instr->definitions.size() == 1 && "unsupported"); aco_ptr res; if (instr->isVOP1()) { res.reset(create_instruction( instr->opcode, instr->format, instr->operands.size(), instr->definitions.size())); } else if (instr->isSOP1()) { res.reset(create_instruction( instr->opcode, instr->format, instr->operands.size(), instr->definitions.size())); } else if (instr->isPseudo()) { res.reset(create_instruction( instr->opcode, instr->format, instr->operands.size(), instr->definitions.size())); } else if (instr->isSOPK()) { res.reset(create_instruction( instr->opcode, instr->format, instr->operands.size(), instr->definitions.size())); res->sopk().imm = instr->sopk().imm; } for (unsigned i = 0; i < instr->operands.size(); i++) { res->operands[i] = instr->operands[i]; if (instr->operands[i].isTemp()) { assert(false && "unsupported"); if (ctx.remat.count(instr->operands[i].getTemp())) ctx.unused_remats.erase(ctx.remat[instr->operands[i].getTemp()].instr); } } res->definitions[0] = Definition(new_name); return res; } else { aco_ptr reload{ create_instruction(aco_opcode::p_reload, Format::PSEUDO, 1, 1)}; reload->operands[0] = Operand::c32(spill_id); reload->definitions[0] = Definition(new_name); ctx.is_reloaded[spill_id] = true; return reload; } } void get_rematerialize_info(spill_ctx& ctx) { for (Block& block : ctx.program->blocks) { bool logical = false; for (aco_ptr& instr : block.instructions) { if (instr->opcode == aco_opcode::p_logical_start) logical = true; else if (instr->opcode == aco_opcode::p_logical_end) logical = false; if (logical && should_rematerialize(instr)) { for (const Definition& def : instr->definitions) { if (def.isTemp()) { ctx.remat[def.getTemp()] = remat_info{instr.get()}; ctx.unused_remats.insert(instr.get()); } } } } } } void update_local_next_uses(spill_ctx& ctx, Block* block, std::vector>>& local_next_uses) { if (local_next_uses.size() < block->instructions.size()) { /* Allocate more next-use-maps. Note that by never reducing the vector size, we enable * future calls to this function to re-use already allocated map memory. */ local_next_uses.resize(block->instructions.size()); } local_next_uses[block->instructions.size() - 1].clear(); for (std::pair>& pair : ctx.next_use_distances_end[block->index]) { local_next_uses[block->instructions.size() - 1].push_back(std::make_pair( (Temp)pair.first, pair.second.second + block->instructions.size())); } for (int idx = block->instructions.size() - 1; idx >= 0; idx--) { aco_ptr& instr = block->instructions[idx]; if (!instr) break; if (instr->opcode == aco_opcode::p_phi || instr->opcode == aco_opcode::p_linear_phi) break; if (idx != (int)block->instructions.size() - 1) { local_next_uses[idx] = local_next_uses[idx + 1]; } for (const Operand& op : instr->operands) { if (op.isFixed() && op.physReg() == exec) continue; if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear()) continue; if (op.isTemp()) { auto it = std::find_if(local_next_uses[idx].begin(), local_next_uses[idx].end(), [op](auto& pair) { return pair.first == op.getTemp(); }); if (it == local_next_uses[idx].end()) { local_next_uses[idx].push_back(std::make_pair(op.getTemp(), idx)); } else { it->second = idx; } } } for (const Definition& def : instr->definitions) { if (def.isTemp()) { auto it = std::find_if(local_next_uses[idx].begin(), local_next_uses[idx].end(), [def](auto& pair) { return pair.first == def.getTemp(); }); if (it != local_next_uses[idx].end()) { local_next_uses[idx].erase(it); } } } } } RegisterDemand get_demand_before(spill_ctx& ctx, unsigned block_idx, unsigned idx) { if (idx == 0) { RegisterDemand demand = ctx.register_demand[block_idx][idx]; aco_ptr& instr = ctx.program->blocks[block_idx].instructions[idx]; aco_ptr instr_before(nullptr); return get_demand_before(demand, instr, instr_before); } else { return ctx.register_demand[block_idx][idx - 1]; } } RegisterDemand get_live_in_demand(spill_ctx& ctx, unsigned block_idx) { unsigned idx = 0; RegisterDemand reg_pressure = RegisterDemand(); Block& block = ctx.program->blocks[block_idx]; for (aco_ptr& phi : block.instructions) { if (!is_phi(phi)) break; idx++; /* Killed phi definitions increase pressure in the predecessor but not * the block they're in. Since the loops below are both to control * pressure of the start of this block and the ends of it's * predecessors, we need to count killed unspilled phi definitions here. */ if (phi->definitions[0].isTemp() && phi->definitions[0].isKill() && !ctx.spills_entry[block_idx].count(phi->definitions[0].getTemp())) reg_pressure += phi->definitions[0].getTemp(); } reg_pressure += get_demand_before(ctx, block_idx, idx); /* Consider register pressure from linear predecessors. This can affect * reg_pressure if the branch instructions define sgprs. */ for (unsigned pred : block.linear_preds) reg_pressure.sgpr = std::max(reg_pressure.sgpr, ctx.register_demand[pred].back().sgpr); return reg_pressure; } RegisterDemand init_live_in_vars(spill_ctx& ctx, Block* block, unsigned block_idx) { RegisterDemand spilled_registers; /* first block, nothing was spilled before */ if (block_idx == 0) return {0, 0}; /* next use distances at the beginning of the current block */ const auto& next_use_distances = ctx.next_use_distances_start[block_idx]; /* loop header block */ if (block->loop_nest_depth > ctx.program->blocks[block_idx - 1].loop_nest_depth) { assert(block->linear_preds[0] == block_idx - 1); assert(block->logical_preds[0] == block_idx - 1); /* create new loop_info */ ctx.loop_header.emplace(block); /* check how many live-through variables should be spilled */ RegisterDemand reg_pressure = get_live_in_demand(ctx, block_idx); RegisterDemand loop_demand = reg_pressure; unsigned i = block_idx; while (ctx.program->blocks[i].loop_nest_depth >= block->loop_nest_depth) { assert(ctx.program->blocks.size() > i); loop_demand.update(ctx.program->blocks[i++].register_demand); } unsigned loop_end = i; for (auto spilled : ctx.spills_exit[block_idx - 1]) { auto it = next_use_distances.find(spilled.first); /* variable is not live at loop entry: probably a phi operand */ if (it == next_use_distances.end()) continue; /* keep constants and live-through variables spilled */ if (it->second.first >= loop_end || ctx.remat.count(spilled.first)) { ctx.spills_entry[block_idx][spilled.first] = spilled.second; spilled_registers += spilled.first; loop_demand -= spilled.first; } } /* select live-through variables and constants */ RegType type = RegType::vgpr; while (loop_demand.exceeds(ctx.target_pressure)) { /* if VGPR demand is low enough, select SGPRs */ if (type == RegType::vgpr && loop_demand.vgpr <= ctx.target_pressure.vgpr) type = RegType::sgpr; /* if SGPR demand is low enough, break */ if (type == RegType::sgpr && loop_demand.sgpr <= ctx.target_pressure.sgpr) break; unsigned distance = 0; Temp to_spill; for (const std::pair>& pair : next_use_distances) { if (pair.first.type() == type && (pair.second.first >= loop_end || (ctx.remat.count(pair.first) && type == RegType::sgpr)) && pair.second.second > distance && !ctx.spills_entry[block_idx].count(pair.first)) { to_spill = pair.first; distance = pair.second.second; } } /* select SGPRs or break */ if (distance == 0) { if (type == RegType::sgpr) break; type = RegType::sgpr; continue; } uint32_t spill_id; if (!ctx.spills_exit[block_idx - 1].count(to_spill)) { spill_id = ctx.allocate_spill_id(to_spill.regClass()); } else { spill_id = ctx.spills_exit[block_idx - 1][to_spill]; } ctx.spills_entry[block_idx][to_spill] = spill_id; spilled_registers += to_spill; loop_demand -= to_spill; } /* shortcut */ if (!loop_demand.exceeds(ctx.target_pressure)) return spilled_registers; /* if reg pressure is too high at beginning of loop, add variables with furthest use */ reg_pressure -= spilled_registers; while (reg_pressure.exceeds(ctx.target_pressure)) { unsigned distance = 0; Temp to_spill; type = reg_pressure.vgpr > ctx.target_pressure.vgpr ? RegType::vgpr : RegType::sgpr; for (const std::pair>& pair : next_use_distances) { if (pair.first.type() == type && pair.second.second > distance && !ctx.spills_entry[block_idx].count(pair.first)) { to_spill = pair.first; distance = pair.second.second; } } assert(distance != 0); ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass()); spilled_registers += to_spill; reg_pressure -= to_spill; } return spilled_registers; } /* branch block */ if (block->linear_preds.size() == 1 && !(block->kind & block_kind_loop_exit)) { /* keep variables spilled if they are alive and not used in the current block */ unsigned pred_idx = block->linear_preds[0]; for (std::pair pair : ctx.spills_exit[pred_idx]) { if (pair.first.type() != RegType::sgpr) { continue; } auto next_use_distance_it = next_use_distances.find(pair.first); if (next_use_distance_it != next_use_distances.end() && next_use_distance_it->second.first != block_idx) { ctx.spills_entry[block_idx].insert(pair); spilled_registers.sgpr += pair.first.size(); } } if (block->logical_preds.size() == 1) { pred_idx = block->logical_preds[0]; for (std::pair pair : ctx.spills_exit[pred_idx]) { if (pair.first.type() != RegType::vgpr) { continue; } auto next_use_distance_it = next_use_distances.find(pair.first); if (next_use_distance_it != next_use_distances.end() && next_use_distance_it->second.first != block_idx) { ctx.spills_entry[block_idx].insert(pair); spilled_registers.vgpr += pair.first.size(); } } } /* if register demand is still too high, we just keep all spilled live vars * and process the block */ if (block->register_demand.sgpr - spilled_registers.sgpr > ctx.target_pressure.sgpr) { pred_idx = block->linear_preds[0]; for (std::pair pair : ctx.spills_exit[pred_idx]) { if (pair.first.type() == RegType::sgpr && next_use_distances.count(pair.first) && ctx.spills_entry[block_idx].insert(pair).second) { spilled_registers.sgpr += pair.first.size(); } } } if (block->register_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr && block->logical_preds.size() == 1) { pred_idx = block->logical_preds[0]; for (std::pair pair : ctx.spills_exit[pred_idx]) { if (pair.first.type() == RegType::vgpr && next_use_distances.count(pair.first) && ctx.spills_entry[block_idx].insert(pair).second) { spilled_registers.vgpr += pair.first.size(); } } } return spilled_registers; } /* else: merge block */ std::set partial_spills; /* keep variables spilled on all incoming paths */ for (const std::pair>& pair : next_use_distances) { std::vector& preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds; /* If it can be rematerialized, keep the variable spilled if all predecessors do not reload * it. Otherwise, if any predecessor reloads it, ensure it's reloaded on all other * predecessors. The idea is that it's better in practice to rematerialize redundantly than to * create lots of phis. */ /* TODO: test this idea with more than Dawn of War III shaders (the current pipeline-db * doesn't seem to exercise this path much) */ bool remat = ctx.remat.count(pair.first); bool spill = !remat; uint32_t spill_id = 0; for (unsigned pred_idx : preds) { /* variable is not even live at the predecessor: probably from a phi */ if (!ctx.next_use_distances_end[pred_idx].count(pair.first)) { spill = false; break; } if (!ctx.spills_exit[pred_idx].count(pair.first)) { if (!remat) spill = false; } else { partial_spills.insert(pair.first); /* it might be that on one incoming path, the variable has a different spill_id, but * add_couple_code() will take care of that. */ spill_id = ctx.spills_exit[pred_idx][pair.first]; if (remat) spill = true; } } if (spill) { ctx.spills_entry[block_idx][pair.first] = spill_id; partial_spills.erase(pair.first); spilled_registers += pair.first; } } /* same for phis */ for (aco_ptr& phi : block->instructions) { if (!is_phi(phi)) break; if (!phi->definitions[0].isTemp()) continue; std::vector& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds; bool spill = true; for (unsigned i = 0; i < phi->operands.size(); i++) { /* non-temp operands can increase the register pressure */ if (!phi->operands[i].isTemp()) { partial_spills.insert(phi->definitions[0].getTemp()); continue; } if (!ctx.spills_exit[preds[i]].count(phi->operands[i].getTemp())) spill = false; else partial_spills.insert(phi->definitions[0].getTemp()); } if (spill) { ctx.spills_entry[block_idx][phi->definitions[0].getTemp()] = ctx.allocate_spill_id(phi->definitions[0].regClass()); partial_spills.erase(phi->definitions[0].getTemp()); spilled_registers += phi->definitions[0].getTemp(); } } /* if reg pressure at first instruction is still too high, add partially spilled variables */ RegisterDemand reg_pressure = get_live_in_demand(ctx, block_idx); reg_pressure -= spilled_registers; while (reg_pressure.exceeds(ctx.target_pressure)) { assert(!partial_spills.empty()); std::set::iterator it = partial_spills.begin(); Temp to_spill = Temp(); unsigned distance = 0; RegType type = reg_pressure.vgpr > ctx.target_pressure.vgpr ? RegType::vgpr : RegType::sgpr; while (it != partial_spills.end()) { assert(!ctx.spills_entry[block_idx].count(*it)); if (it->type() == type && next_use_distances.at(*it).second > distance) { distance = next_use_distances.at(*it).second; to_spill = *it; } ++it; } assert(distance != 0); ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass()); partial_spills.erase(to_spill); spilled_registers += to_spill; reg_pressure -= to_spill; } return spilled_registers; } void add_coupling_code(spill_ctx& ctx, Block* block, unsigned block_idx) { /* no coupling code necessary */ if (block->linear_preds.size() == 0) return; std::vector> instructions; /* branch block: TODO take other branch into consideration */ if (block->linear_preds.size() == 1 && !(block->kind & (block_kind_loop_exit | block_kind_loop_header))) { assert(ctx.processed[block->linear_preds[0]]); assert(ctx.register_demand[block_idx].size() == block->instructions.size()); std::vector reg_demand; unsigned insert_idx = 0; RegisterDemand demand_before = get_demand_before(ctx, block_idx, 0); for (std::pair>& live : ctx.next_use_distances_start[block_idx]) { const unsigned pred_idx = block->linear_preds[0]; if (!live.first.is_linear()) continue; /* still spilled */ if (ctx.spills_entry[block_idx].count(live.first)) continue; /* in register at end of predecessor */ auto spills_exit_it = ctx.spills_exit[pred_idx].find(live.first); if (spills_exit_it == ctx.spills_exit[pred_idx].end()) { std::map::iterator it = ctx.renames[pred_idx].find(live.first); if (it != ctx.renames[pred_idx].end()) ctx.renames[block_idx].insert(*it); continue; } /* variable is spilled at predecessor and live at current block: create reload instruction */ Temp new_name = ctx.program->allocateTmp(live.first.regClass()); aco_ptr reload = do_reload(ctx, live.first, new_name, spills_exit_it->second); instructions.emplace_back(std::move(reload)); reg_demand.push_back(demand_before); ctx.renames[block_idx][live.first] = new_name; } if (block->logical_preds.size() == 1) { do { assert(insert_idx < block->instructions.size()); instructions.emplace_back(std::move(block->instructions[insert_idx])); reg_demand.push_back(ctx.register_demand[block_idx][insert_idx]); insert_idx++; } while (instructions.back()->opcode != aco_opcode::p_logical_start); unsigned pred_idx = block->logical_preds[0]; for (std::pair>& live : ctx.next_use_distances_start[block_idx]) { if (live.first.is_linear()) continue; /* still spilled */ if (ctx.spills_entry[block_idx].count(live.first)) continue; /* in register at end of predecessor */ auto spills_exit_it = ctx.spills_exit[pred_idx].find(live.first); if (spills_exit_it == ctx.spills_exit[pred_idx].end()) { std::map::iterator it = ctx.renames[pred_idx].find(live.first); if (it != ctx.renames[pred_idx].end()) ctx.renames[block_idx].insert(*it); continue; } /* variable is spilled at predecessor and live at current block: * create reload instruction */ Temp new_name = ctx.program->allocateTmp(live.first.regClass()); aco_ptr reload = do_reload(ctx, live.first, new_name, spills_exit_it->second); instructions.emplace_back(std::move(reload)); reg_demand.emplace_back(reg_demand.back()); ctx.renames[block_idx][live.first] = new_name; } } /* combine new reload instructions with original block */ if (!instructions.empty()) { reg_demand.insert(reg_demand.end(), std::next(ctx.register_demand[block->index].begin(), insert_idx), ctx.register_demand[block->index].end()); ctx.register_demand[block_idx] = std::move(reg_demand); instructions.insert(instructions.end(), std::move_iterator>::iterator>( std::next(block->instructions.begin(), insert_idx)), std::move_iterator>::iterator>( block->instructions.end())); block->instructions = std::move(instructions); } return; } /* loop header and merge blocks: check if all (linear) predecessors have been processed */ for (ASSERTED unsigned pred : block->linear_preds) assert(ctx.processed[pred]); /* iterate the phi nodes for which operands to spill at the predecessor */ for (aco_ptr& phi : block->instructions) { if (!is_phi(phi)) break; /* if the phi is not spilled, add to instructions */ if (!phi->definitions[0].isTemp() || !ctx.spills_entry[block_idx].count(phi->definitions[0].getTemp())) { instructions.emplace_back(std::move(phi)); continue; } std::vector& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds; uint32_t def_spill_id = ctx.spills_entry[block_idx][phi->definitions[0].getTemp()]; for (unsigned i = 0; i < phi->operands.size(); i++) { if (phi->operands[i].isUndefined()) continue; unsigned pred_idx = preds[i]; Operand spill_op = phi->operands[i]; if (spill_op.isTemp()) { assert(phi->operands[i].isKill()); Temp var = phi->operands[i].getTemp(); std::map::iterator rename_it = ctx.renames[pred_idx].find(var); /* prevent the definining instruction from being DCE'd if it could be rematerialized */ if (rename_it == ctx.renames[preds[i]].end() && ctx.remat.count(var)) ctx.unused_remats.erase(ctx.remat[var].instr); /* check if variable is already spilled at predecessor */ auto spilled = ctx.spills_exit[pred_idx].find(var); if (spilled != ctx.spills_exit[pred_idx].end()) { if (spilled->second != def_spill_id) ctx.add_affinity(def_spill_id, spilled->second); continue; } /* rename if necessary */ if (rename_it != ctx.renames[pred_idx].end()) { spill_op.setTemp(rename_it->second); ctx.renames[pred_idx].erase(rename_it); } } uint32_t spill_id = ctx.allocate_spill_id(phi->definitions[0].regClass()); /* add interferences and affinity */ for (std::pair pair : ctx.spills_exit[pred_idx]) ctx.add_interference(spill_id, pair.second); ctx.add_affinity(def_spill_id, spill_id); aco_ptr spill{ create_instruction(aco_opcode::p_spill, Format::PSEUDO, 2, 0)}; spill->operands[0] = spill_op; spill->operands[1] = Operand::c32(spill_id); Block& pred = ctx.program->blocks[pred_idx]; unsigned idx = pred.instructions.size(); do { assert(idx != 0); idx--; } while (phi->opcode == aco_opcode::p_phi && pred.instructions[idx]->opcode != aco_opcode::p_logical_end); std::vector>::iterator it = std::next(pred.instructions.begin(), idx); pred.instructions.insert(it, std::move(spill)); if (spill_op.isTemp()) ctx.spills_exit[pred_idx][spill_op.getTemp()] = spill_id; } /* remove phi from instructions */ phi.reset(); } /* iterate all (other) spilled variables for which to spill at the predecessor */ // TODO: would be better to have them sorted: first vgprs and first with longest distance for (std::pair pair : ctx.spills_entry[block_idx]) { std::vector preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds; for (unsigned pred_idx : preds) { /* variable is already spilled at predecessor */ auto spilled = ctx.spills_exit[pred_idx].find(pair.first); if (spilled != ctx.spills_exit[pred_idx].end()) { if (spilled->second != pair.second) ctx.add_affinity(pair.second, spilled->second); continue; } /* variable is dead at predecessor, it must be from a phi: this works because of CSSA form */ if (!ctx.next_use_distances_end[pred_idx].count(pair.first)) continue; /* add interferences between spilled variable and predecessors exit spills */ for (std::pair exit_spill : ctx.spills_exit[pred_idx]) { if (exit_spill.first == pair.first) continue; ctx.add_interference(exit_spill.second, pair.second); } /* variable is in register at predecessor and has to be spilled */ /* rename if necessary */ Temp var = pair.first; std::map::iterator rename_it = ctx.renames[pred_idx].find(var); if (rename_it != ctx.renames[pred_idx].end()) { var = rename_it->second; ctx.renames[pred_idx].erase(rename_it); } aco_ptr spill{ create_instruction(aco_opcode::p_spill, Format::PSEUDO, 2, 0)}; spill->operands[0] = Operand(var); spill->operands[1] = Operand::c32(pair.second); Block& pred = ctx.program->blocks[pred_idx]; unsigned idx = pred.instructions.size(); do { assert(idx != 0); idx--; } while (pair.first.type() == RegType::vgpr && pred.instructions[idx]->opcode != aco_opcode::p_logical_end); std::vector>::iterator it = std::next(pred.instructions.begin(), idx); pred.instructions.insert(it, std::move(spill)); ctx.spills_exit[pred.index][pair.first] = pair.second; } } /* iterate phis for which operands to reload */ for (aco_ptr& phi : instructions) { assert(phi->opcode == aco_opcode::p_phi || phi->opcode == aco_opcode::p_linear_phi); assert(!phi->definitions[0].isTemp() || !ctx.spills_entry[block_idx].count(phi->definitions[0].getTemp())); std::vector& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds; for (unsigned i = 0; i < phi->operands.size(); i++) { if (!phi->operands[i].isTemp()) continue; unsigned pred_idx = preds[i]; /* if the operand was reloaded, rename */ if (!ctx.spills_exit[pred_idx].count(phi->operands[i].getTemp())) { std::map::iterator it = ctx.renames[pred_idx].find(phi->operands[i].getTemp()); if (it != ctx.renames[pred_idx].end()) { phi->operands[i].setTemp(it->second); /* prevent the definining instruction from being DCE'd if it could be rematerialized */ } else { auto remat_it = ctx.remat.find(phi->operands[i].getTemp()); if (remat_it != ctx.remat.end()) { ctx.unused_remats.erase(remat_it->second.instr); } } continue; } Temp tmp = phi->operands[i].getTemp(); /* reload phi operand at end of predecessor block */ Temp new_name = ctx.program->allocateTmp(tmp.regClass()); Block& pred = ctx.program->blocks[pred_idx]; unsigned idx = pred.instructions.size(); do { assert(idx != 0); idx--; } while (phi->opcode == aco_opcode::p_phi && pred.instructions[idx]->opcode != aco_opcode::p_logical_end); std::vector>::iterator it = std::next(pred.instructions.begin(), idx); aco_ptr reload = do_reload(ctx, tmp, new_name, ctx.spills_exit[pred_idx][tmp]); /* reload spilled exec mask directly to exec */ if (!phi->definitions[0].isTemp()) { assert(phi->definitions[0].isFixed() && phi->definitions[0].physReg() == exec); reload->definitions[0] = phi->definitions[0]; phi->operands[i] = Operand(exec, ctx.program->lane_mask); } else { ctx.spills_exit[pred_idx].erase(tmp); ctx.renames[pred_idx][tmp] = new_name; phi->operands[i].setTemp(new_name); } pred.instructions.insert(it, std::move(reload)); } } /* iterate live variables for which to reload */ // TODO: reload at current block if variable is spilled on all predecessors for (std::pair>& pair : ctx.next_use_distances_start[block_idx]) { /* skip spilled variables */ if (ctx.spills_entry[block_idx].count(pair.first)) continue; std::vector preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds; /* variable is dead at predecessor, it must be from a phi */ bool is_dead = false; for (unsigned pred_idx : preds) { if (!ctx.next_use_distances_end[pred_idx].count(pair.first)) is_dead = true; } if (is_dead) continue; for (unsigned pred_idx : preds) { /* the variable is not spilled at the predecessor */ if (!ctx.spills_exit[pred_idx].count(pair.first)) continue; /* variable is spilled at predecessor and has to be reloaded */ Temp new_name = ctx.program->allocateTmp(pair.first.regClass()); Block& pred = ctx.program->blocks[pred_idx]; unsigned idx = pred.instructions.size(); do { assert(idx != 0); idx--; } while (pair.first.type() == RegType::vgpr && pred.instructions[idx]->opcode != aco_opcode::p_logical_end); std::vector>::iterator it = std::next(pred.instructions.begin(), idx); aco_ptr reload = do_reload(ctx, pair.first, new_name, ctx.spills_exit[pred.index][pair.first]); pred.instructions.insert(it, std::move(reload)); ctx.spills_exit[pred.index].erase(pair.first); ctx.renames[pred.index][pair.first] = new_name; } /* check if we have to create a new phi for this variable */ Temp rename = Temp(); bool is_same = true; for (unsigned pred_idx : preds) { if (!ctx.renames[pred_idx].count(pair.first)) { if (rename == Temp()) rename = pair.first; else is_same = rename == pair.first; } else { if (rename == Temp()) rename = ctx.renames[pred_idx][pair.first]; else is_same = rename == ctx.renames[pred_idx][pair.first]; } if (!is_same) break; } if (!is_same) { /* the variable was renamed differently in the predecessors: we have to create a phi */ aco_opcode opcode = pair.first.is_linear() ? aco_opcode::p_linear_phi : aco_opcode::p_phi; aco_ptr phi{ create_instruction(opcode, Format::PSEUDO, preds.size(), 1)}; rename = ctx.program->allocateTmp(pair.first.regClass()); for (unsigned i = 0; i < phi->operands.size(); i++) { Temp tmp; if (ctx.renames[preds[i]].count(pair.first)) { tmp = ctx.renames[preds[i]][pair.first]; } else if (preds[i] >= block_idx) { tmp = rename; } else { tmp = pair.first; /* prevent the definining instruction from being DCE'd if it could be rematerialized */ if (ctx.remat.count(tmp)) ctx.unused_remats.erase(ctx.remat[tmp].instr); } phi->operands[i] = Operand(tmp); } phi->definitions[0] = Definition(rename); instructions.emplace_back(std::move(phi)); } /* the variable was renamed: add new name to renames */ if (!(rename == Temp() || rename == pair.first)) ctx.renames[block_idx][pair.first] = rename; } /* combine phis with instructions */ unsigned idx = 0; while (!block->instructions[idx]) { idx++; } if (!ctx.processed[block_idx]) { assert(!(block->kind & block_kind_loop_header)); RegisterDemand demand_before = get_demand_before(ctx, block_idx, idx); ctx.register_demand[block->index].erase(ctx.register_demand[block->index].begin(), ctx.register_demand[block->index].begin() + idx); ctx.register_demand[block->index].insert(ctx.register_demand[block->index].begin(), instructions.size(), demand_before); } std::vector>::iterator start = std::next(block->instructions.begin(), idx); instructions.insert( instructions.end(), std::move_iterator>::iterator>(start), std::move_iterator>::iterator>(block->instructions.end())); block->instructions = std::move(instructions); } void process_block(spill_ctx& ctx, unsigned block_idx, Block* block, RegisterDemand spilled_registers) { assert(!ctx.processed[block_idx]); std::vector> instructions; unsigned idx = 0; /* phis are handled separetely */ while (block->instructions[idx]->opcode == aco_opcode::p_phi || block->instructions[idx]->opcode == aco_opcode::p_linear_phi) { instructions.emplace_back(std::move(block->instructions[idx++])); } if (block->register_demand.exceeds(ctx.target_pressure)) { update_local_next_uses(ctx, block, ctx.local_next_use_distance); } else { /* We won't use local_next_use_distance, so no initialization needed */ } auto& current_spills = ctx.spills_exit[block_idx]; while (idx < block->instructions.size()) { aco_ptr& instr = block->instructions[idx]; std::map> reloads; /* rename and reload operands */ for (Operand& op : instr->operands) { if (!op.isTemp()) continue; if (!current_spills.count(op.getTemp())) { /* the Operand is in register: check if it was renamed */ auto rename_it = ctx.renames[block_idx].find(op.getTemp()); if (rename_it != ctx.renames[block_idx].end()) { op.setTemp(rename_it->second); } else { /* prevent its definining instruction from being DCE'd if it could be rematerialized */ auto remat_it = ctx.remat.find(op.getTemp()); if (remat_it != ctx.remat.end()) { ctx.unused_remats.erase(remat_it->second.instr); } } continue; } /* the Operand is spilled: add it to reloads */ Temp new_tmp = ctx.program->allocateTmp(op.regClass()); ctx.renames[block_idx][op.getTemp()] = new_tmp; reloads[new_tmp] = std::make_pair(op.getTemp(), current_spills[op.getTemp()]); current_spills.erase(op.getTemp()); op.setTemp(new_tmp); spilled_registers -= new_tmp; } /* check if register demand is low enough before and after the current instruction */ if (block->register_demand.exceeds(ctx.target_pressure)) { RegisterDemand new_demand = ctx.register_demand[block_idx][idx]; new_demand.update(get_demand_before(ctx, block_idx, idx)); assert(!ctx.local_next_use_distance.empty()); /* if reg pressure is too high, spill variable with furthest next use */ while ((new_demand - spilled_registers).exceeds(ctx.target_pressure)) { unsigned distance = 0; Temp to_spill; bool do_rematerialize = false; RegType type = RegType::sgpr; if (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) type = RegType::vgpr; for (std::pair pair : ctx.local_next_use_distance[idx]) { if (pair.first.type() != type) continue; bool can_rematerialize = ctx.remat.count(pair.first); if (((pair.second > distance && can_rematerialize == do_rematerialize) || (can_rematerialize && !do_rematerialize && pair.second > idx)) && !current_spills.count(pair.first)) { to_spill = pair.first; distance = pair.second; do_rematerialize = can_rematerialize; } } assert(distance != 0 && distance > idx); uint32_t spill_id = ctx.allocate_spill_id(to_spill.regClass()); /* add interferences with currently spilled variables */ for (std::pair pair : current_spills) ctx.add_interference(spill_id, pair.second); for (std::pair>& pair : reloads) ctx.add_interference(spill_id, pair.second.second); current_spills[to_spill] = spill_id; spilled_registers += to_spill; /* rename if necessary */ if (ctx.renames[block_idx].count(to_spill)) { to_spill = ctx.renames[block_idx][to_spill]; } /* add spill to new instructions */ aco_ptr spill{ create_instruction(aco_opcode::p_spill, Format::PSEUDO, 2, 0)}; spill->operands[0] = Operand(to_spill); spill->operands[1] = Operand::c32(spill_id); instructions.emplace_back(std::move(spill)); } } /* add reloads and instruction to new instructions */ for (std::pair>& pair : reloads) { aco_ptr reload = do_reload(ctx, pair.second.first, pair.first, pair.second.second); instructions.emplace_back(std::move(reload)); } instructions.emplace_back(std::move(instr)); idx++; } block->instructions = std::move(instructions); } void spill_block(spill_ctx& ctx, unsigned block_idx) { Block* block = &ctx.program->blocks[block_idx]; /* determine set of variables which are spilled at the beginning of the block */ RegisterDemand spilled_registers = init_live_in_vars(ctx, block, block_idx); /* add interferences for spilled variables */ for (auto it = ctx.spills_entry[block_idx].begin(); it != ctx.spills_entry[block_idx].end(); ++it) { for (auto it2 = std::next(it); it2 != ctx.spills_entry[block_idx].end(); ++it2) ctx.add_interference(it->second, it2->second); } bool is_loop_header = block->loop_nest_depth && ctx.loop_header.top()->index == block_idx; if (!is_loop_header) { /* add spill/reload code on incoming control flow edges */ add_coupling_code(ctx, block, block_idx); } const auto& current_spills = ctx.spills_entry[block_idx]; /* check conditions to process this block */ bool process = (block->register_demand - spilled_registers).exceeds(ctx.target_pressure) || !ctx.renames[block_idx].empty() || ctx.unused_remats.size(); for (auto it = current_spills.begin(); !process && it != current_spills.end(); ++it) { if (ctx.next_use_distances_start[block_idx].at(it->first).first == block_idx) process = true; } assert(ctx.spills_exit[block_idx].empty()); ctx.spills_exit[block_idx] = current_spills; if (process) { process_block(ctx, block_idx, block, spilled_registers); } ctx.processed[block_idx] = true; /* check if the next block leaves the current loop */ if (block->loop_nest_depth == 0 || ctx.program->blocks[block_idx + 1].loop_nest_depth >= block->loop_nest_depth) return; Block* loop_header = ctx.loop_header.top(); /* preserve original renames at end of loop header block */ std::map renames = std::move(ctx.renames[loop_header->index]); /* add coupling code to all loop header predecessors */ add_coupling_code(ctx, loop_header, loop_header->index); /* propagate new renames through loop: i.e. repair the SSA */ renames.swap(ctx.renames[loop_header->index]); for (std::pair rename : renames) { for (unsigned idx = loop_header->index; idx <= block_idx; idx++) { Block& current = ctx.program->blocks[idx]; std::vector>::iterator instr_it = current.instructions.begin(); /* first rename phis */ while (instr_it != current.instructions.end()) { aco_ptr& phi = *instr_it; if (phi->opcode != aco_opcode::p_phi && phi->opcode != aco_opcode::p_linear_phi) break; /* no need to rename the loop header phis once again. this happened in * add_coupling_code() */ if (idx == loop_header->index) { instr_it++; continue; } for (Operand& op : phi->operands) { if (!op.isTemp()) continue; if (op.getTemp() == rename.first) op.setTemp(rename.second); } instr_it++; } /* variable is not live at beginning of this block */ if (ctx.next_use_distances_start[idx].count(rename.first) == 0) continue; /* if the variable is live at the block's exit, add rename */ if (ctx.next_use_distances_end[idx].count(rename.first) != 0) ctx.renames[idx].insert(rename); /* rename all uses in this block */ bool renamed = false; while (!renamed && instr_it != current.instructions.end()) { aco_ptr& instr = *instr_it; for (Operand& op : instr->operands) { if (!op.isTemp()) continue; if (op.getTemp() == rename.first) { op.setTemp(rename.second); /* we can stop with this block as soon as the variable is spilled */ if (instr->opcode == aco_opcode::p_spill) renamed = true; } } instr_it++; } } } /* remove loop header info from stack */ ctx.loop_header.pop(); } Temp load_scratch_resource(spill_ctx& ctx, Temp& scratch_offset, std::vector>& instructions, unsigned offset, bool is_top_level) { Builder bld(ctx.program); if (is_top_level) { bld.reset(&instructions); } else { /* find p_logical_end */ unsigned idx = instructions.size() - 1; while (instructions[idx]->opcode != aco_opcode::p_logical_end) idx--; bld.reset(&instructions, std::next(instructions.begin(), idx)); } Temp private_segment_buffer = ctx.program->private_segment_buffer; if (ctx.program->stage != compute_cs) private_segment_buffer = bld.smem(aco_opcode::s_load_dwordx2, bld.def(s2), private_segment_buffer, Operand::zero()); if (offset) scratch_offset = bld.sop2(aco_opcode::s_add_u32, bld.def(s1), bld.def(s1, scc), scratch_offset, Operand::c32(offset)); uint32_t rsrc_conf = S_008F0C_ADD_TID_ENABLE(1) | S_008F0C_INDEX_STRIDE(ctx.program->wave_size == 64 ? 3 : 2); if (ctx.program->chip_class >= GFX10) { rsrc_conf |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) | S_008F0C_RESOURCE_LEVEL(1); } else if (ctx.program->chip_class <= GFX7) { /* dfmt modifies stride on GFX8/GFX9 when ADD_TID_EN=1 */ rsrc_conf |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32); } /* older generations need element size = 4 bytes. element size removed in GFX9 */ if (ctx.program->chip_class <= GFX8) rsrc_conf |= S_008F0C_ELEMENT_SIZE(1); return bld.pseudo(aco_opcode::p_create_vector, bld.def(s4), private_segment_buffer, Operand::c32(-1u), Operand::c32(rsrc_conf)); } void add_interferences(spill_ctx& ctx, std::vector& is_assigned, std::vector& slots, std::vector& slots_used, unsigned id) { for (unsigned other : ctx.interferences[id].second) { if (!is_assigned[other]) continue; RegClass other_rc = ctx.interferences[other].first; unsigned slot = slots[other]; std::fill(slots_used.begin() + slot, slots_used.begin() + slot + other_rc.size(), true); } } unsigned find_available_slot(std::vector& used, unsigned wave_size, unsigned size, bool is_sgpr, unsigned* num_slots) { unsigned wave_size_minus_one = wave_size - 1; unsigned slot = 0; while (true) { bool available = true; for (unsigned i = 0; i < size; i++) { if (slot + i < used.size() && used[slot + i]) { available = false; break; } } if (!available) { slot++; continue; } if (is_sgpr && ((slot & wave_size_minus_one) > wave_size - size)) { slot = align(slot, wave_size); continue; } std::fill(used.begin(), used.end(), false); if (slot + size > used.size()) used.resize(slot + size); return slot; } } void assign_spill_slots_helper(spill_ctx& ctx, RegType type, std::vector& is_assigned, std::vector& slots, unsigned* num_slots) { std::vector slots_used(*num_slots); /* assign slots for ids with affinities first */ for (std::vector& vec : ctx.affinities) { if (ctx.interferences[vec[0]].first.type() != type) continue; for (unsigned id : vec) { if (!ctx.is_reloaded[id]) continue; add_interferences(ctx, is_assigned, slots, slots_used, id); } unsigned slot = find_available_slot(slots_used, ctx.wave_size, ctx.interferences[vec[0]].first.size(), type == RegType::sgpr, num_slots); for (unsigned id : vec) { assert(!is_assigned[id]); if (ctx.is_reloaded[id]) { slots[id] = slot; is_assigned[id] = true; } } } /* assign slots for ids without affinities */ for (unsigned id = 0; id < ctx.interferences.size(); id++) { if (is_assigned[id] || !ctx.is_reloaded[id] || ctx.interferences[id].first.type() != type) continue; add_interferences(ctx, is_assigned, slots, slots_used, id); unsigned slot = find_available_slot(slots_used, ctx.wave_size, ctx.interferences[id].first.size(), type == RegType::sgpr, num_slots); slots[id] = slot; is_assigned[id] = true; } *num_slots = slots_used.size(); } void assign_spill_slots(spill_ctx& ctx, unsigned spills_to_vgpr) { std::vector slots(ctx.interferences.size()); std::vector is_assigned(ctx.interferences.size()); /* first, handle affinities: just merge all interferences into both spill ids */ for (std::vector& vec : ctx.affinities) { for (unsigned i = 0; i < vec.size(); i++) { for (unsigned j = i + 1; j < vec.size(); j++) { assert(vec[i] != vec[j]); bool reloaded = ctx.is_reloaded[vec[i]] || ctx.is_reloaded[vec[j]]; ctx.is_reloaded[vec[i]] = reloaded; ctx.is_reloaded[vec[j]] = reloaded; } } } for (ASSERTED uint32_t i = 0; i < ctx.interferences.size(); i++) for (ASSERTED uint32_t id : ctx.interferences[i].second) assert(i != id); /* for each spill slot, assign as many spill ids as possible */ unsigned sgpr_spill_slots = 0, vgpr_spill_slots = 0; assign_spill_slots_helper(ctx, RegType::sgpr, is_assigned, slots, &sgpr_spill_slots); assign_spill_slots_helper(ctx, RegType::vgpr, is_assigned, slots, &vgpr_spill_slots); for (unsigned id = 0; id < is_assigned.size(); id++) assert(is_assigned[id] || !ctx.is_reloaded[id]); for (std::vector& vec : ctx.affinities) { for (unsigned i = 0; i < vec.size(); i++) { for (unsigned j = i + 1; j < vec.size(); j++) { assert(is_assigned[vec[i]] == is_assigned[vec[j]]); if (!is_assigned[vec[i]]) continue; assert(ctx.is_reloaded[vec[i]] == ctx.is_reloaded[vec[j]]); assert(ctx.interferences[vec[i]].first.type() == ctx.interferences[vec[j]].first.type()); assert(slots[vec[i]] == slots[vec[j]]); } } } /* hope, we didn't mess up */ std::vector vgpr_spill_temps((sgpr_spill_slots + ctx.wave_size - 1) / ctx.wave_size); assert(vgpr_spill_temps.size() <= spills_to_vgpr); /* replace pseudo instructions with actual hardware instructions */ Temp scratch_offset = ctx.program->scratch_offset, scratch_rsrc = Temp(); unsigned last_top_level_block_idx = 0; std::vector reload_in_loop(vgpr_spill_temps.size()); for (Block& block : ctx.program->blocks) { /* after loops, we insert a user if there was a reload inside the loop */ if (block.loop_nest_depth == 0) { int end_vgprs = 0; for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) { if (reload_in_loop[i]) end_vgprs++; } if (end_vgprs > 0) { aco_ptr destr{create_instruction( aco_opcode::p_end_linear_vgpr, Format::PSEUDO, end_vgprs, 0)}; int k = 0; for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) { if (reload_in_loop[i]) destr->operands[k++] = Operand(vgpr_spill_temps[i]); reload_in_loop[i] = false; } /* find insertion point */ std::vector>::iterator it = block.instructions.begin(); while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi) ++it; block.instructions.insert(it, std::move(destr)); } } if (block.kind & block_kind_top_level && !block.linear_preds.empty()) { last_top_level_block_idx = block.index; /* check if any spilled variables use a created linear vgpr, otherwise destroy them */ for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) { if (vgpr_spill_temps[i] == Temp()) continue; bool can_destroy = true; for (std::pair pair : ctx.spills_entry[block.index]) { if (ctx.interferences[pair.second].first.type() == RegType::sgpr && slots[pair.second] / ctx.wave_size == i) { can_destroy = false; break; } } if (can_destroy) vgpr_spill_temps[i] = Temp(); } } std::vector>::iterator it; std::vector> instructions; instructions.reserve(block.instructions.size()); Builder bld(ctx.program, &instructions); for (it = block.instructions.begin(); it != block.instructions.end(); ++it) { if ((*it)->opcode == aco_opcode::p_spill) { uint32_t spill_id = (*it)->operands[1].constantValue(); if (!ctx.is_reloaded[spill_id]) { /* never reloaded, so don't spill */ } else if (!is_assigned[spill_id]) { unreachable("No spill slot assigned for spill id"); } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) { /* spill vgpr */ ctx.program->config->spilled_vgprs += (*it)->operands[0].size(); uint32_t spill_slot = slots[spill_id]; bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096; unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size; /* check if the scratch resource descriptor already exists */ if (scratch_rsrc == Temp()) { unsigned offset = add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0; scratch_rsrc = load_scratch_resource( ctx, scratch_offset, last_top_level_block_idx == block.index ? instructions : ctx.program->blocks[last_top_level_block_idx].instructions, offset, last_top_level_block_idx == block.index); } unsigned offset = base_offset + spill_slot * 4; aco_opcode opcode = aco_opcode::buffer_store_dword; assert((*it)->operands[0].isTemp()); Temp temp = (*it)->operands[0].getTemp(); assert(temp.type() == RegType::vgpr && !temp.is_linear()); if (temp.size() > 1) { Instruction* split{create_instruction( aco_opcode::p_split_vector, Format::PSEUDO, 1, temp.size())}; split->operands[0] = Operand(temp); for (unsigned i = 0; i < temp.size(); i++) split->definitions[i] = bld.def(v1); bld.insert(split); for (unsigned i = 0; i < temp.size(); i++) { Instruction* instr = bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset, split->definitions[i].getTemp(), offset + i * 4, false, true); instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private); } } else { Instruction* instr = bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset, temp, offset, false, true); instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private); } } else { ctx.program->config->spilled_sgprs += (*it)->operands[0].size(); uint32_t spill_slot = slots[spill_id]; /* check if the linear vgpr already exists */ if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) { Temp linear_vgpr = ctx.program->allocateTmp(v1.as_linear()); vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr; aco_ptr create{create_instruction( aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)}; create->definitions[0] = Definition(linear_vgpr); /* find the right place to insert this definition */ if (last_top_level_block_idx == block.index) { /* insert right before the current instruction */ instructions.emplace_back(std::move(create)); } else { assert(last_top_level_block_idx < block.index); /* insert before the branch at last top level block */ std::vector>& block_instrs = ctx.program->blocks[last_top_level_block_idx].instructions; block_instrs.insert(std::prev(block_instrs.end()), std::move(create)); } } /* spill sgpr: just add the vgpr temp to operands */ Pseudo_instruction* spill = create_instruction(aco_opcode::p_spill, Format::PSEUDO, 3, 0); spill->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]); spill->operands[1] = Operand::c32(spill_slot % ctx.wave_size); spill->operands[2] = (*it)->operands[0]; instructions.emplace_back(aco_ptr(spill)); } } else if ((*it)->opcode == aco_opcode::p_reload) { uint32_t spill_id = (*it)->operands[0].constantValue(); assert(ctx.is_reloaded[spill_id]); if (!is_assigned[spill_id]) { unreachable("No spill slot assigned for spill id"); } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) { /* reload vgpr */ uint32_t spill_slot = slots[spill_id]; bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096; unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size; /* check if the scratch resource descriptor already exists */ if (scratch_rsrc == Temp()) { unsigned offset = add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0; scratch_rsrc = load_scratch_resource( ctx, scratch_offset, last_top_level_block_idx == block.index ? instructions : ctx.program->blocks[last_top_level_block_idx].instructions, offset, last_top_level_block_idx == block.index); } unsigned offset = base_offset + spill_slot * 4; aco_opcode opcode = aco_opcode::buffer_load_dword; Definition def = (*it)->definitions[0]; if (def.size() > 1) { Instruction* vec{create_instruction( aco_opcode::p_create_vector, Format::PSEUDO, def.size(), 1)}; vec->definitions[0] = def; for (unsigned i = 0; i < def.size(); i++) { Temp tmp = bld.tmp(v1); vec->operands[i] = Operand(tmp); Instruction* instr = bld.mubuf(opcode, Definition(tmp), scratch_rsrc, Operand(v1), scratch_offset, offset + i * 4, false, true); instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private); } bld.insert(vec); } else { Instruction* instr = bld.mubuf(opcode, def, scratch_rsrc, Operand(v1), scratch_offset, offset, false, true); instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private); } } else { uint32_t spill_slot = slots[spill_id]; reload_in_loop[spill_slot / ctx.wave_size] = block.loop_nest_depth > 0; /* check if the linear vgpr already exists */ if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) { Temp linear_vgpr = ctx.program->allocateTmp(v1.as_linear()); vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr; aco_ptr create{create_instruction( aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)}; create->definitions[0] = Definition(linear_vgpr); /* find the right place to insert this definition */ if (last_top_level_block_idx == block.index) { /* insert right before the current instruction */ instructions.emplace_back(std::move(create)); } else { assert(last_top_level_block_idx < block.index); /* insert before the branch at last top level block */ std::vector>& block_instrs = ctx.program->blocks[last_top_level_block_idx].instructions; block_instrs.insert(std::prev(block_instrs.end()), std::move(create)); } } /* reload sgpr: just add the vgpr temp to operands */ Pseudo_instruction* reload = create_instruction( aco_opcode::p_reload, Format::PSEUDO, 2, 1); reload->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]); reload->operands[1] = Operand::c32(spill_slot % ctx.wave_size); reload->definitions[0] = (*it)->definitions[0]; instructions.emplace_back(aco_ptr(reload)); } } else if (!ctx.unused_remats.count(it->get())) { instructions.emplace_back(std::move(*it)); } } block.instructions = std::move(instructions); } /* update required scratch memory */ ctx.program->config->scratch_bytes_per_wave += align(vgpr_spill_slots * 4 * ctx.program->wave_size, 1024); /* SSA elimination inserts copies for logical phis right before p_logical_end * So if a linear vgpr is used between that p_logical_end and the branch, * we need to ensure logical phis don't choose a definition which aliases * the linear vgpr. * TODO: Moving the spills and reloads to before p_logical_end might produce * slightly better code. */ for (Block& block : ctx.program->blocks) { /* loops exits are already handled */ if (block.logical_preds.size() <= 1) continue; bool has_logical_phis = false; for (aco_ptr& instr : block.instructions) { if (instr->opcode == aco_opcode::p_phi) { has_logical_phis = true; break; } else if (instr->opcode != aco_opcode::p_linear_phi) { break; } } if (!has_logical_phis) continue; std::set vgprs; for (unsigned pred_idx : block.logical_preds) { Block& pred = ctx.program->blocks[pred_idx]; for (int i = pred.instructions.size() - 1; i >= 0; i--) { aco_ptr& pred_instr = pred.instructions[i]; if (pred_instr->opcode == aco_opcode::p_logical_end) { break; } else if (pred_instr->opcode == aco_opcode::p_spill || pred_instr->opcode == aco_opcode::p_reload) { vgprs.insert(pred_instr->operands[0].getTemp()); } } } if (!vgprs.size()) continue; aco_ptr destr{create_instruction( aco_opcode::p_end_linear_vgpr, Format::PSEUDO, vgprs.size(), 0)}; int k = 0; for (Temp tmp : vgprs) { destr->operands[k++] = Operand(tmp); } /* find insertion point */ std::vector>::iterator it = block.instructions.begin(); while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi) ++it; block.instructions.insert(it, std::move(destr)); } } } /* end namespace */ void spill(Program* program, live& live_vars) { program->config->spilled_vgprs = 0; program->config->spilled_sgprs = 0; program->progress = CompilationProgress::after_spilling; /* no spilling when register pressure is low enough */ if (program->num_waves > 0) return; /* lower to CSSA before spilling to ensure correctness w.r.t. phis */ lower_to_cssa(program, live_vars); /* calculate target register demand */ const RegisterDemand demand = program->max_reg_demand; /* current max */ const uint16_t sgpr_limit = get_addr_sgpr_from_waves(program, program->min_waves); const uint16_t vgpr_limit = get_addr_vgpr_from_waves(program, program->min_waves); uint16_t extra_vgprs = 0; uint16_t extra_sgprs = 0; /* calculate extra VGPRs required for spilling SGPRs */ if (demand.sgpr > sgpr_limit) { unsigned sgpr_spills = demand.sgpr - sgpr_limit; extra_vgprs = DIV_ROUND_UP(sgpr_spills, program->wave_size) + 1; } /* add extra SGPRs required for spilling VGPRs */ if (demand.vgpr + extra_vgprs > vgpr_limit) { extra_sgprs = 5; /* scratch_resource (s4) + scratch_offset (s1) */ if (demand.sgpr + extra_sgprs > sgpr_limit) { /* re-calculate in case something has changed */ unsigned sgpr_spills = demand.sgpr + extra_sgprs - sgpr_limit; extra_vgprs = DIV_ROUND_UP(sgpr_spills, program->wave_size) + 1; } } /* the spiller has to target the following register demand */ const RegisterDemand target(vgpr_limit - extra_vgprs, sgpr_limit - extra_sgprs); /* initialize ctx */ spill_ctx ctx(target, program, live_vars.register_demand); compute_global_next_uses(ctx); get_rematerialize_info(ctx); /* create spills and reloads */ for (unsigned i = 0; i < program->blocks.size(); i++) spill_block(ctx, i); /* assign spill slots and DCE rematerialized code */ assign_spill_slots(ctx, extra_vgprs); /* update live variable information */ live_vars = live_var_analysis(program); assert(program->num_waves > 0); } } // namespace aco