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_builder.h"
27 #include "aco_ir.h"
28 
29 #include "common/sid.h"
30 
31 #include <algorithm>
32 #include <cstring>
33 #include <map>
34 #include <set>
35 #include <stack>
36 #include <unordered_map>
37 #include <unordered_set>
38 #include <vector>
39 
40 namespace std {
41 template <> struct hash<aco::Temp> {
operator ()std::hash42    size_t operator()(aco::Temp temp) const noexcept
43    {
44       uint32_t v;
45       std::memcpy(&v, &temp, sizeof(temp));
46       return std::hash<uint32_t>{}(v);
47    }
48 };
49 } // namespace std
50 
51 /*
52  * Implements the spilling algorithm on SSA-form from
53  * "Register Spilling and Live-Range Splitting for SSA-Form Programs"
54  * by Matthias Braun and Sebastian Hack.
55  */
56 
57 namespace aco {
58 
59 namespace {
60 
61 struct remat_info {
62    Instruction* instr;
63 };
64 
65 struct spill_ctx {
66    RegisterDemand target_pressure;
67    Program* program;
68    std::vector<std::vector<RegisterDemand>> register_demand;
69    std::vector<std::map<Temp, Temp>> renames;
70    std::vector<std::unordered_map<Temp, uint32_t>> spills_entry;
71    std::vector<std::unordered_map<Temp, uint32_t>> spills_exit;
72 
73    std::vector<bool> processed;
74    std::stack<Block*, std::vector<Block*>> loop_header;
75    std::vector<std::unordered_map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_start;
76    std::vector<std::unordered_map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_end;
77    std::vector<std::vector<std::pair<Temp, uint32_t>>> local_next_use_distance; /* Working buffer */
78    std::vector<std::pair<RegClass, std::unordered_set<uint32_t>>> interferences;
79    std::vector<std::vector<uint32_t>> affinities;
80    std::vector<bool> is_reloaded;
81    std::unordered_map<Temp, remat_info> remat;
82    std::set<Instruction*> unused_remats;
83    unsigned wave_size;
84 
spill_ctxaco::__anonaad5f72e0111::spill_ctx85    spill_ctx(const RegisterDemand target_pressure_, Program* program_,
86              std::vector<std::vector<RegisterDemand>> register_demand_)
87        : target_pressure(target_pressure_), program(program_),
88          register_demand(std::move(register_demand_)), renames(program->blocks.size()),
89          spills_entry(program->blocks.size()), spills_exit(program->blocks.size()),
90          processed(program->blocks.size(), false), wave_size(program->wave_size)
91    {}
92 
add_affinityaco::__anonaad5f72e0111::spill_ctx93    void add_affinity(uint32_t first, uint32_t second)
94    {
95       unsigned found_first = affinities.size();
96       unsigned found_second = affinities.size();
97       for (unsigned i = 0; i < affinities.size(); i++) {
98          std::vector<uint32_t>& vec = affinities[i];
99          for (uint32_t entry : vec) {
100             if (entry == first)
101                found_first = i;
102             else if (entry == second)
103                found_second = i;
104          }
105       }
106       if (found_first == affinities.size() && found_second == affinities.size()) {
107          affinities.emplace_back(std::vector<uint32_t>({first, second}));
108       } else if (found_first < affinities.size() && found_second == affinities.size()) {
109          affinities[found_first].push_back(second);
110       } else if (found_second < affinities.size() && found_first == affinities.size()) {
111          affinities[found_second].push_back(first);
112       } else if (found_first != found_second) {
113          /* merge second into first */
114          affinities[found_first].insert(affinities[found_first].end(),
115                                         affinities[found_second].begin(),
116                                         affinities[found_second].end());
117          affinities.erase(std::next(affinities.begin(), found_second));
118       } else {
119          assert(found_first == found_second);
120       }
121    }
122 
add_interferenceaco::__anonaad5f72e0111::spill_ctx123    void add_interference(uint32_t first, uint32_t second)
124    {
125       if (interferences[first].first.type() != interferences[second].first.type())
126          return;
127 
128       bool inserted = interferences[first].second.insert(second).second;
129       if (inserted)
130          interferences[second].second.insert(first);
131    }
132 
allocate_spill_idaco::__anonaad5f72e0111::spill_ctx133    uint32_t allocate_spill_id(RegClass rc)
134    {
135       interferences.emplace_back(rc, std::unordered_set<uint32_t>());
136       is_reloaded.push_back(false);
137       return next_spill_id++;
138    }
139 
140    uint32_t next_spill_id = 0;
141 };
142 
143 int32_t
get_dominator(int idx_a,int idx_b,Program * program,bool is_linear)144 get_dominator(int idx_a, int idx_b, Program* program, bool is_linear)
145 {
146 
147    if (idx_a == -1)
148       return idx_b;
149    if (idx_b == -1)
150       return idx_a;
151    if (is_linear) {
152       while (idx_a != idx_b) {
153          if (idx_a > idx_b)
154             idx_a = program->blocks[idx_a].linear_idom;
155          else
156             idx_b = program->blocks[idx_b].linear_idom;
157       }
158    } else {
159       while (idx_a != idx_b) {
160          if (idx_a > idx_b)
161             idx_a = program->blocks[idx_a].logical_idom;
162          else
163             idx_b = program->blocks[idx_b].logical_idom;
164       }
165    }
166    assert(idx_a != -1);
167    return idx_a;
168 }
169 
170 void
next_uses_per_block(spill_ctx & ctx,unsigned block_idx,uint32_t & worklist)171 next_uses_per_block(spill_ctx& ctx, unsigned block_idx, uint32_t& worklist)
172 {
173    Block* block = &ctx.program->blocks[block_idx];
174    ctx.next_use_distances_start[block_idx] = ctx.next_use_distances_end[block_idx];
175    auto& next_use_distances_start = ctx.next_use_distances_start[block_idx];
176 
177    /* to compute the next use distance at the beginning of the block, we have to add the block's
178     * size */
179    for (std::unordered_map<Temp, std::pair<uint32_t, uint32_t>>::iterator it =
180            next_use_distances_start.begin();
181         it != next_use_distances_start.end(); ++it)
182       it->second.second = it->second.second + block->instructions.size();
183 
184    int idx = block->instructions.size() - 1;
185    while (idx >= 0) {
186       aco_ptr<Instruction>& instr = block->instructions[idx];
187 
188       if (instr->opcode == aco_opcode::p_linear_phi || instr->opcode == aco_opcode::p_phi)
189          break;
190 
191       for (const Definition& def : instr->definitions) {
192          if (def.isTemp())
193             next_use_distances_start.erase(def.getTemp());
194       }
195 
196       for (const Operand& op : instr->operands) {
197          /* omit exec mask */
198          if (op.isFixed() && op.physReg() == exec)
199             continue;
200          if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear())
201             continue;
202          if (op.isTemp())
203             next_use_distances_start[op.getTemp()] = {block_idx, idx};
204       }
205       idx--;
206    }
207 
208    assert(block_idx != 0 || next_use_distances_start.empty());
209    std::unordered_set<Temp> phi_defs;
210    while (idx >= 0) {
211       aco_ptr<Instruction>& instr = block->instructions[idx];
212       assert(instr->opcode == aco_opcode::p_linear_phi || instr->opcode == aco_opcode::p_phi);
213 
214       std::pair<uint32_t, uint32_t> distance{block_idx, 0};
215 
216       auto it = instr->definitions[0].isTemp() ? next_use_distances_start.find(instr->definitions[0].getTemp())
217                                                : next_use_distances_start.end();
218       if (it != next_use_distances_start.end() &&
219          phi_defs.insert(instr->definitions[0].getTemp()).second) {
220          distance = it->second;
221       }
222 
223       for (unsigned i = 0; i < instr->operands.size(); i++) {
224          unsigned pred_idx =
225             instr->opcode == aco_opcode::p_phi ? block->logical_preds[i] : block->linear_preds[i];
226          if (instr->operands[i].isTemp()) {
227             auto insert_result = ctx.next_use_distances_end[pred_idx].insert(
228                std::make_pair(instr->operands[i].getTemp(), distance));
229             const bool inserted = insert_result.second;
230             std::pair<uint32_t, uint32_t>& entry_distance = insert_result.first->second;
231             if (inserted || entry_distance != distance)
232                worklist = std::max(worklist, pred_idx + 1);
233             entry_distance = distance;
234          }
235       }
236       idx--;
237    }
238 
239    /* all remaining live vars must be live-out at the predecessors */
240    for (std::pair<const Temp, std::pair<uint32_t, uint32_t>>& pair : next_use_distances_start) {
241       Temp temp = pair.first;
242       if (phi_defs.count(temp)) {
243          continue;
244       }
245       uint32_t distance = pair.second.second;
246       uint32_t dom = pair.second.first;
247       std::vector<unsigned>& preds = temp.is_linear() ? block->linear_preds : block->logical_preds;
248       for (unsigned pred_idx : preds) {
249          if (ctx.program->blocks[pred_idx].loop_nest_depth > block->loop_nest_depth)
250             distance += 0xFFFF;
251          auto insert_result = ctx.next_use_distances_end[pred_idx].insert(
252             std::make_pair(temp, std::pair<uint32_t, uint32_t>{}));
253          const bool inserted = insert_result.second;
254          std::pair<uint32_t, uint32_t>& entry_distance = insert_result.first->second;
255 
256          if (!inserted) {
257             dom = get_dominator(dom, entry_distance.first, ctx.program, temp.is_linear());
258             distance = std::min(entry_distance.second, distance);
259          }
260          if (entry_distance != std::pair<uint32_t, uint32_t>{dom, distance}) {
261             worklist = std::max(worklist, pred_idx + 1);
262             entry_distance = {dom, distance};
263          }
264       }
265    }
266 }
267 
268 void
compute_global_next_uses(spill_ctx & ctx)269 compute_global_next_uses(spill_ctx& ctx)
270 {
271    ctx.next_use_distances_start.resize(ctx.program->blocks.size());
272    ctx.next_use_distances_end.resize(ctx.program->blocks.size());
273 
274    uint32_t worklist = ctx.program->blocks.size();
275    while (worklist) {
276       unsigned block_idx = --worklist;
277       next_uses_per_block(ctx, block_idx, worklist);
278    }
279 }
280 
281 bool
should_rematerialize(aco_ptr<Instruction> & instr)282 should_rematerialize(aco_ptr<Instruction>& instr)
283 {
284    /* TODO: rematerialization is only supported for VOP1, SOP1 and PSEUDO */
285    if (instr->format != Format::VOP1 && instr->format != Format::SOP1 &&
286        instr->format != Format::PSEUDO && instr->format != Format::SOPK)
287       return false;
288    /* TODO: pseudo-instruction rematerialization is only supported for
289     * p_create_vector/p_parallelcopy */
290    if (instr->isPseudo() && instr->opcode != aco_opcode::p_create_vector &&
291        instr->opcode != aco_opcode::p_parallelcopy)
292       return false;
293    if (instr->isSOPK() && instr->opcode != aco_opcode::s_movk_i32)
294       return false;
295 
296    for (const Operand& op : instr->operands) {
297       /* TODO: rematerialization using temporaries isn't yet supported */
298       if (!op.isConstant())
299          return false;
300    }
301 
302    /* TODO: rematerialization with multiple definitions isn't yet supported */
303    if (instr->definitions.size() > 1)
304       return false;
305 
306    return true;
307 }
308 
309 aco_ptr<Instruction>
do_reload(spill_ctx & ctx,Temp tmp,Temp new_name,uint32_t spill_id)310 do_reload(spill_ctx& ctx, Temp tmp, Temp new_name, uint32_t spill_id)
311 {
312    std::unordered_map<Temp, remat_info>::iterator remat = ctx.remat.find(tmp);
313    if (remat != ctx.remat.end()) {
314       Instruction* instr = remat->second.instr;
315       assert((instr->isVOP1() || instr->isSOP1() || instr->isPseudo() || instr->isSOPK()) &&
316              "unsupported");
317       assert((instr->format != Format::PSEUDO || instr->opcode == aco_opcode::p_create_vector ||
318               instr->opcode == aco_opcode::p_parallelcopy) &&
319              "unsupported");
320       assert(instr->definitions.size() == 1 && "unsupported");
321 
322       aco_ptr<Instruction> res;
323       if (instr->isVOP1()) {
324          res.reset(create_instruction<VOP1_instruction>(
325             instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
326       } else if (instr->isSOP1()) {
327          res.reset(create_instruction<SOP1_instruction>(
328             instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
329       } else if (instr->isPseudo()) {
330          res.reset(create_instruction<Pseudo_instruction>(
331             instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
332       } else if (instr->isSOPK()) {
333          res.reset(create_instruction<SOPK_instruction>(
334             instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
335          res->sopk().imm = instr->sopk().imm;
336       }
337       for (unsigned i = 0; i < instr->operands.size(); i++) {
338          res->operands[i] = instr->operands[i];
339          if (instr->operands[i].isTemp()) {
340             assert(false && "unsupported");
341             if (ctx.remat.count(instr->operands[i].getTemp()))
342                ctx.unused_remats.erase(ctx.remat[instr->operands[i].getTemp()].instr);
343          }
344       }
345       res->definitions[0] = Definition(new_name);
346       return res;
347    } else {
348       aco_ptr<Pseudo_instruction> reload{
349          create_instruction<Pseudo_instruction>(aco_opcode::p_reload, Format::PSEUDO, 1, 1)};
350       reload->operands[0] = Operand::c32(spill_id);
351       reload->definitions[0] = Definition(new_name);
352       ctx.is_reloaded[spill_id] = true;
353       return reload;
354    }
355 }
356 
357 void
get_rematerialize_info(spill_ctx & ctx)358 get_rematerialize_info(spill_ctx& ctx)
359 {
360    for (Block& block : ctx.program->blocks) {
361       bool logical = false;
362       for (aco_ptr<Instruction>& instr : block.instructions) {
363          if (instr->opcode == aco_opcode::p_logical_start)
364             logical = true;
365          else if (instr->opcode == aco_opcode::p_logical_end)
366             logical = false;
367          if (logical && should_rematerialize(instr)) {
368             for (const Definition& def : instr->definitions) {
369                if (def.isTemp()) {
370                   ctx.remat[def.getTemp()] = remat_info{instr.get()};
371                   ctx.unused_remats.insert(instr.get());
372                }
373             }
374          }
375       }
376    }
377 }
378 
379 void
update_local_next_uses(spill_ctx & ctx,Block * block,std::vector<std::vector<std::pair<Temp,uint32_t>>> & local_next_uses)380 update_local_next_uses(spill_ctx& ctx, Block* block,
381                 std::vector<std::vector<std::pair<Temp, uint32_t>>>& local_next_uses)
382 {
383    if (local_next_uses.size() < block->instructions.size()) {
384       /* Allocate more next-use-maps. Note that by never reducing the vector size, we enable
385        * future calls to this function to re-use already allocated map memory. */
386       local_next_uses.resize(block->instructions.size());
387    }
388 
389    local_next_uses[block->instructions.size() - 1].clear();
390    for (std::pair<const Temp, std::pair<uint32_t, uint32_t>>& pair :
391         ctx.next_use_distances_end[block->index]) {
392       local_next_uses[block->instructions.size() - 1].push_back(std::make_pair<Temp, uint32_t>(
393          (Temp)pair.first, pair.second.second + block->instructions.size()));
394    }
395 
396    for (int idx = block->instructions.size() - 1; idx >= 0; idx--) {
397       aco_ptr<Instruction>& instr = block->instructions[idx];
398       if (!instr)
399          break;
400       if (instr->opcode == aco_opcode::p_phi || instr->opcode == aco_opcode::p_linear_phi)
401          break;
402 
403       if (idx != (int)block->instructions.size() - 1) {
404          local_next_uses[idx] = local_next_uses[idx + 1];
405       }
406 
407       for (const Operand& op : instr->operands) {
408          if (op.isFixed() && op.physReg() == exec)
409             continue;
410          if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear())
411             continue;
412          if (op.isTemp()) {
413             auto it = std::find_if(local_next_uses[idx].begin(), local_next_uses[idx].end(),
414                                    [op](auto& pair) { return pair.first == op.getTemp(); });
415             if (it == local_next_uses[idx].end()) {
416                local_next_uses[idx].push_back(std::make_pair<Temp, uint32_t>(op.getTemp(), idx));
417             } else {
418                it->second = idx;
419             }
420          }
421       }
422       for (const Definition& def : instr->definitions) {
423          if (def.isTemp()) {
424             auto it = std::find_if(local_next_uses[idx].begin(), local_next_uses[idx].end(),
425                                    [def](auto& pair) { return pair.first == def.getTemp(); });
426             if (it != local_next_uses[idx].end()) {
427                local_next_uses[idx].erase(it);
428             }
429          }
430       }
431    }
432 }
433 
434 RegisterDemand
get_demand_before(spill_ctx & ctx,unsigned block_idx,unsigned idx)435 get_demand_before(spill_ctx& ctx, unsigned block_idx, unsigned idx)
436 {
437    if (idx == 0) {
438       RegisterDemand demand = ctx.register_demand[block_idx][idx];
439       aco_ptr<Instruction>& instr = ctx.program->blocks[block_idx].instructions[idx];
440       aco_ptr<Instruction> instr_before(nullptr);
441       return get_demand_before(demand, instr, instr_before);
442    } else {
443       return ctx.register_demand[block_idx][idx - 1];
444    }
445 }
446 
447 RegisterDemand
get_live_in_demand(spill_ctx & ctx,unsigned block_idx)448 get_live_in_demand(spill_ctx& ctx, unsigned block_idx)
449 {
450    unsigned idx = 0;
451    RegisterDemand reg_pressure = RegisterDemand();
452    Block& block = ctx.program->blocks[block_idx];
453    for (aco_ptr<Instruction>& phi : block.instructions) {
454       if (!is_phi(phi))
455          break;
456       idx++;
457 
458       /* Killed phi definitions increase pressure in the predecessor but not
459        * the block they're in. Since the loops below are both to control
460        * pressure of the start of this block and the ends of it's
461        * predecessors, we need to count killed unspilled phi definitions here. */
462       if (phi->definitions[0].isTemp() && phi->definitions[0].isKill() &&
463           !ctx.spills_entry[block_idx].count(phi->definitions[0].getTemp()))
464          reg_pressure += phi->definitions[0].getTemp();
465    }
466 
467    reg_pressure += get_demand_before(ctx, block_idx, idx);
468 
469    /* Consider register pressure from linear predecessors. This can affect
470     * reg_pressure if the branch instructions define sgprs. */
471    for (unsigned pred : block.linear_preds)
472       reg_pressure.sgpr =
473          std::max<int16_t>(reg_pressure.sgpr, ctx.register_demand[pred].back().sgpr);
474 
475    return reg_pressure;
476 }
477 
478 RegisterDemand
init_live_in_vars(spill_ctx & ctx,Block * block,unsigned block_idx)479 init_live_in_vars(spill_ctx& ctx, Block* block, unsigned block_idx)
480 {
481    RegisterDemand spilled_registers;
482 
483    /* first block, nothing was spilled before */
484    if (block_idx == 0)
485       return {0, 0};
486 
487    /* next use distances at the beginning of the current block */
488    const auto& next_use_distances = ctx.next_use_distances_start[block_idx];
489 
490    /* loop header block */
491    if (block->loop_nest_depth > ctx.program->blocks[block_idx - 1].loop_nest_depth) {
492       assert(block->linear_preds[0] == block_idx - 1);
493       assert(block->logical_preds[0] == block_idx - 1);
494 
495       /* create new loop_info */
496       ctx.loop_header.emplace(block);
497 
498       /* check how many live-through variables should be spilled */
499       RegisterDemand reg_pressure = get_live_in_demand(ctx, block_idx);
500       RegisterDemand loop_demand = reg_pressure;
501       unsigned i = block_idx;
502       while (ctx.program->blocks[i].loop_nest_depth >= block->loop_nest_depth) {
503          assert(ctx.program->blocks.size() > i);
504          loop_demand.update(ctx.program->blocks[i++].register_demand);
505       }
506       unsigned loop_end = i;
507 
508       for (auto spilled : ctx.spills_exit[block_idx - 1]) {
509          auto it = next_use_distances.find(spilled.first);
510 
511          /* variable is not live at loop entry: probably a phi operand */
512          if (it == next_use_distances.end())
513             continue;
514 
515          /* keep constants and live-through variables spilled */
516          if (it->second.first >= loop_end || ctx.remat.count(spilled.first)) {
517             ctx.spills_entry[block_idx][spilled.first] = spilled.second;
518             spilled_registers += spilled.first;
519             loop_demand -= spilled.first;
520          }
521       }
522 
523       /* select live-through variables and constants */
524       RegType type = RegType::vgpr;
525       while (loop_demand.exceeds(ctx.target_pressure)) {
526          /* if VGPR demand is low enough, select SGPRs */
527          if (type == RegType::vgpr && loop_demand.vgpr <= ctx.target_pressure.vgpr)
528             type = RegType::sgpr;
529          /* if SGPR demand is low enough, break */
530          if (type == RegType::sgpr && loop_demand.sgpr <= ctx.target_pressure.sgpr)
531             break;
532 
533          unsigned distance = 0;
534          Temp to_spill;
535          for (const std::pair<const Temp, std::pair<uint32_t, uint32_t>>& pair :
536               next_use_distances) {
537             if (pair.first.type() == type &&
538                 (pair.second.first >= loop_end ||
539                  (ctx.remat.count(pair.first) && type == RegType::sgpr)) &&
540                 pair.second.second > distance && !ctx.spills_entry[block_idx].count(pair.first)) {
541                to_spill = pair.first;
542                distance = pair.second.second;
543             }
544          }
545 
546          /* select SGPRs or break */
547          if (distance == 0) {
548             if (type == RegType::sgpr)
549                break;
550             type = RegType::sgpr;
551             continue;
552          }
553 
554          uint32_t spill_id;
555          if (!ctx.spills_exit[block_idx - 1].count(to_spill)) {
556             spill_id = ctx.allocate_spill_id(to_spill.regClass());
557          } else {
558             spill_id = ctx.spills_exit[block_idx - 1][to_spill];
559          }
560 
561          ctx.spills_entry[block_idx][to_spill] = spill_id;
562          spilled_registers += to_spill;
563          loop_demand -= to_spill;
564       }
565 
566       /* shortcut */
567       if (!loop_demand.exceeds(ctx.target_pressure))
568          return spilled_registers;
569 
570       /* if reg pressure is too high at beginning of loop, add variables with furthest use */
571       reg_pressure -= spilled_registers;
572 
573       while (reg_pressure.exceeds(ctx.target_pressure)) {
574          unsigned distance = 0;
575          Temp to_spill;
576          type = reg_pressure.vgpr > ctx.target_pressure.vgpr ? RegType::vgpr : RegType::sgpr;
577 
578          for (const std::pair<const Temp, std::pair<uint32_t, uint32_t>>& pair :
579               next_use_distances) {
580             if (pair.first.type() == type && pair.second.second > distance &&
581                 !ctx.spills_entry[block_idx].count(pair.first)) {
582                to_spill = pair.first;
583                distance = pair.second.second;
584             }
585          }
586          assert(distance != 0);
587          ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
588          spilled_registers += to_spill;
589          reg_pressure -= to_spill;
590       }
591 
592       return spilled_registers;
593    }
594 
595    /* branch block */
596    if (block->linear_preds.size() == 1 && !(block->kind & block_kind_loop_exit)) {
597       /* keep variables spilled if they are alive and not used in the current block */
598       unsigned pred_idx = block->linear_preds[0];
599       for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
600          if (pair.first.type() != RegType::sgpr) {
601             continue;
602          }
603          auto next_use_distance_it = next_use_distances.find(pair.first);
604          if (next_use_distance_it != next_use_distances.end() &&
605              next_use_distance_it->second.first != block_idx) {
606             ctx.spills_entry[block_idx].insert(pair);
607             spilled_registers.sgpr += pair.first.size();
608          }
609       }
610       if (block->logical_preds.size() == 1) {
611          pred_idx = block->logical_preds[0];
612          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
613             if (pair.first.type() != RegType::vgpr) {
614                continue;
615             }
616             auto next_use_distance_it = next_use_distances.find(pair.first);
617             if (next_use_distance_it != next_use_distances.end() &&
618                 next_use_distance_it->second.first != block_idx) {
619                ctx.spills_entry[block_idx].insert(pair);
620                spilled_registers.vgpr += pair.first.size();
621             }
622          }
623       }
624 
625       /* if register demand is still too high, we just keep all spilled live vars
626        * and process the block */
627       if (block->register_demand.sgpr - spilled_registers.sgpr > ctx.target_pressure.sgpr) {
628          pred_idx = block->linear_preds[0];
629          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
630             if (pair.first.type() == RegType::sgpr && next_use_distances.count(pair.first) &&
631                 ctx.spills_entry[block_idx].insert(pair).second) {
632                spilled_registers.sgpr += pair.first.size();
633             }
634          }
635       }
636       if (block->register_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr &&
637           block->logical_preds.size() == 1) {
638          pred_idx = block->logical_preds[0];
639          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
640             if (pair.first.type() == RegType::vgpr && next_use_distances.count(pair.first) &&
641                 ctx.spills_entry[block_idx].insert(pair).second) {
642                spilled_registers.vgpr += pair.first.size();
643             }
644          }
645       }
646 
647       return spilled_registers;
648    }
649 
650    /* else: merge block */
651    std::set<Temp> partial_spills;
652 
653    /* keep variables spilled on all incoming paths */
654    for (const std::pair<const Temp, std::pair<uint32_t, uint32_t>>& pair : next_use_distances) {
655       std::vector<unsigned>& preds =
656          pair.first.is_linear() ? block->linear_preds : block->logical_preds;
657       /* If it can be rematerialized, keep the variable spilled if all predecessors do not reload
658        * it. Otherwise, if any predecessor reloads it, ensure it's reloaded on all other
659        * predecessors. The idea is that it's better in practice to rematerialize redundantly than to
660        * create lots of phis. */
661       /* TODO: test this idea with more than Dawn of War III shaders (the current pipeline-db
662        * doesn't seem to exercise this path much) */
663       bool remat = ctx.remat.count(pair.first);
664       bool spill = !remat;
665       uint32_t spill_id = 0;
666       for (unsigned pred_idx : preds) {
667          /* variable is not even live at the predecessor: probably from a phi */
668          if (!ctx.next_use_distances_end[pred_idx].count(pair.first)) {
669             spill = false;
670             break;
671          }
672          if (!ctx.spills_exit[pred_idx].count(pair.first)) {
673             if (!remat)
674                spill = false;
675          } else {
676             partial_spills.insert(pair.first);
677             /* it might be that on one incoming path, the variable has a different spill_id, but
678              * add_couple_code() will take care of that. */
679             spill_id = ctx.spills_exit[pred_idx][pair.first];
680             if (remat)
681                spill = true;
682          }
683       }
684       if (spill) {
685          ctx.spills_entry[block_idx][pair.first] = spill_id;
686          partial_spills.erase(pair.first);
687          spilled_registers += pair.first;
688       }
689    }
690 
691    /* same for phis */
692    for (aco_ptr<Instruction>& phi : block->instructions) {
693       if (!is_phi(phi))
694          break;
695       if (!phi->definitions[0].isTemp())
696          continue;
697 
698       std::vector<unsigned>& preds =
699          phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
700       bool spill = true;
701 
702       for (unsigned i = 0; i < phi->operands.size(); i++) {
703          /* non-temp operands can increase the register pressure */
704          if (!phi->operands[i].isTemp()) {
705             partial_spills.insert(phi->definitions[0].getTemp());
706             continue;
707          }
708 
709          if (!ctx.spills_exit[preds[i]].count(phi->operands[i].getTemp()))
710             spill = false;
711          else
712             partial_spills.insert(phi->definitions[0].getTemp());
713       }
714       if (spill) {
715          ctx.spills_entry[block_idx][phi->definitions[0].getTemp()] =
716             ctx.allocate_spill_id(phi->definitions[0].regClass());
717          partial_spills.erase(phi->definitions[0].getTemp());
718          spilled_registers += phi->definitions[0].getTemp();
719       }
720    }
721 
722    /* if reg pressure at first instruction is still too high, add partially spilled variables */
723    RegisterDemand reg_pressure = get_live_in_demand(ctx, block_idx);
724    reg_pressure -= spilled_registers;
725 
726    while (reg_pressure.exceeds(ctx.target_pressure)) {
727       assert(!partial_spills.empty());
728       std::set<Temp>::iterator it = partial_spills.begin();
729       Temp to_spill = Temp();
730       unsigned distance = 0;
731       RegType type = reg_pressure.vgpr > ctx.target_pressure.vgpr ? RegType::vgpr : RegType::sgpr;
732 
733       while (it != partial_spills.end()) {
734          assert(!ctx.spills_entry[block_idx].count(*it));
735 
736          if (it->type() == type && next_use_distances.at(*it).second > distance) {
737             distance = next_use_distances.at(*it).second;
738             to_spill = *it;
739          }
740          ++it;
741       }
742       assert(distance != 0);
743 
744       ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
745       partial_spills.erase(to_spill);
746       spilled_registers += to_spill;
747       reg_pressure -= to_spill;
748    }
749 
750    return spilled_registers;
751 }
752 
753 void
add_coupling_code(spill_ctx & ctx,Block * block,unsigned block_idx)754 add_coupling_code(spill_ctx& ctx, Block* block, unsigned block_idx)
755 {
756    /* no coupling code necessary */
757    if (block->linear_preds.size() == 0)
758       return;
759 
760    std::vector<aco_ptr<Instruction>> instructions;
761    /* branch block: TODO take other branch into consideration */
762    if (block->linear_preds.size() == 1 &&
763        !(block->kind & (block_kind_loop_exit | block_kind_loop_header))) {
764       assert(ctx.processed[block->linear_preds[0]]);
765       assert(ctx.register_demand[block_idx].size() == block->instructions.size());
766       std::vector<RegisterDemand> reg_demand;
767       unsigned insert_idx = 0;
768       RegisterDemand demand_before = get_demand_before(ctx, block_idx, 0);
769 
770       for (std::pair<const Temp, std::pair<uint32_t, uint32_t>>& live :
771            ctx.next_use_distances_start[block_idx]) {
772          const unsigned pred_idx = block->linear_preds[0];
773 
774          if (!live.first.is_linear())
775             continue;
776          /* still spilled */
777          if (ctx.spills_entry[block_idx].count(live.first))
778             continue;
779 
780          /* in register at end of predecessor */
781          auto spills_exit_it = ctx.spills_exit[pred_idx].find(live.first);
782          if (spills_exit_it == ctx.spills_exit[pred_idx].end()) {
783             std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(live.first);
784             if (it != ctx.renames[pred_idx].end())
785                ctx.renames[block_idx].insert(*it);
786             continue;
787          }
788 
789          /* variable is spilled at predecessor and live at current block: create reload instruction */
790          Temp new_name = ctx.program->allocateTmp(live.first.regClass());
791          aco_ptr<Instruction> reload = do_reload(ctx, live.first, new_name, spills_exit_it->second);
792          instructions.emplace_back(std::move(reload));
793          reg_demand.push_back(demand_before);
794          ctx.renames[block_idx][live.first] = new_name;
795       }
796 
797       if (block->logical_preds.size() == 1) {
798          do {
799             assert(insert_idx < block->instructions.size());
800             instructions.emplace_back(std::move(block->instructions[insert_idx]));
801             reg_demand.push_back(ctx.register_demand[block_idx][insert_idx]);
802             insert_idx++;
803          } while (instructions.back()->opcode != aco_opcode::p_logical_start);
804 
805          unsigned pred_idx = block->logical_preds[0];
806          for (std::pair<const Temp, std::pair<uint32_t, uint32_t>>& live :
807               ctx.next_use_distances_start[block_idx]) {
808             if (live.first.is_linear())
809                continue;
810             /* still spilled */
811             if (ctx.spills_entry[block_idx].count(live.first))
812                continue;
813 
814             /* in register at end of predecessor */
815             auto spills_exit_it = ctx.spills_exit[pred_idx].find(live.first);
816             if (spills_exit_it == ctx.spills_exit[pred_idx].end()) {
817                std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(live.first);
818                if (it != ctx.renames[pred_idx].end())
819                   ctx.renames[block_idx].insert(*it);
820                continue;
821             }
822 
823             /* variable is spilled at predecessor and live at current block:
824              * create reload instruction */
825             Temp new_name = ctx.program->allocateTmp(live.first.regClass());
826             aco_ptr<Instruction> reload =
827                do_reload(ctx, live.first, new_name, spills_exit_it->second);
828             instructions.emplace_back(std::move(reload));
829             reg_demand.emplace_back(reg_demand.back());
830             ctx.renames[block_idx][live.first] = new_name;
831          }
832       }
833 
834       /* combine new reload instructions with original block */
835       if (!instructions.empty()) {
836          reg_demand.insert(reg_demand.end(),
837                            std::next(ctx.register_demand[block->index].begin(), insert_idx),
838                            ctx.register_demand[block->index].end());
839          ctx.register_demand[block_idx] = std::move(reg_demand);
840          instructions.insert(instructions.end(),
841                              std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(
842                                 std::next(block->instructions.begin(), insert_idx)),
843                              std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(
844                                 block->instructions.end()));
845          block->instructions = std::move(instructions);
846       }
847       return;
848    }
849 
850    /* loop header and merge blocks: check if all (linear) predecessors have been processed */
851    for (ASSERTED unsigned pred : block->linear_preds)
852       assert(ctx.processed[pred]);
853 
854    /* iterate the phi nodes for which operands to spill at the predecessor */
855    for (aco_ptr<Instruction>& phi : block->instructions) {
856       if (!is_phi(phi))
857          break;
858 
859       /* if the phi is not spilled, add to instructions */
860       if (!phi->definitions[0].isTemp() ||
861           !ctx.spills_entry[block_idx].count(phi->definitions[0].getTemp())) {
862          instructions.emplace_back(std::move(phi));
863          continue;
864       }
865 
866       std::vector<unsigned>& preds =
867          phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
868       uint32_t def_spill_id = ctx.spills_entry[block_idx][phi->definitions[0].getTemp()];
869 
870       for (unsigned i = 0; i < phi->operands.size(); i++) {
871          if (phi->operands[i].isUndefined())
872             continue;
873 
874          unsigned pred_idx = preds[i];
875          Operand spill_op = phi->operands[i];
876 
877          if (spill_op.isTemp()) {
878             assert(phi->operands[i].isKill());
879             Temp var = phi->operands[i].getTemp();
880 
881             std::map<Temp, Temp>::iterator rename_it = ctx.renames[pred_idx].find(var);
882             /* prevent the definining instruction from being DCE'd if it could be rematerialized */
883             if (rename_it == ctx.renames[preds[i]].end() && ctx.remat.count(var))
884                ctx.unused_remats.erase(ctx.remat[var].instr);
885 
886             /* check if variable is already spilled at predecessor */
887             auto spilled = ctx.spills_exit[pred_idx].find(var);
888             if (spilled != ctx.spills_exit[pred_idx].end()) {
889                if (spilled->second != def_spill_id)
890                   ctx.add_affinity(def_spill_id, spilled->second);
891                continue;
892             }
893 
894             /* rename if necessary */
895             if (rename_it != ctx.renames[pred_idx].end()) {
896                spill_op.setTemp(rename_it->second);
897                ctx.renames[pred_idx].erase(rename_it);
898             }
899          }
900 
901          uint32_t spill_id = ctx.allocate_spill_id(phi->definitions[0].regClass());
902 
903          /* add interferences and affinity */
904          for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx])
905             ctx.add_interference(spill_id, pair.second);
906          ctx.add_affinity(def_spill_id, spill_id);
907 
908          aco_ptr<Pseudo_instruction> spill{
909             create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
910          spill->operands[0] = spill_op;
911          spill->operands[1] = Operand::c32(spill_id);
912          Block& pred = ctx.program->blocks[pred_idx];
913          unsigned idx = pred.instructions.size();
914          do {
915             assert(idx != 0);
916             idx--;
917          } while (phi->opcode == aco_opcode::p_phi &&
918                   pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
919          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
920          pred.instructions.insert(it, std::move(spill));
921          if (spill_op.isTemp())
922             ctx.spills_exit[pred_idx][spill_op.getTemp()] = spill_id;
923       }
924 
925       /* remove phi from instructions */
926       phi.reset();
927    }
928 
929    /* iterate all (other) spilled variables for which to spill at the predecessor */
930    // TODO: would be better to have them sorted: first vgprs and first with longest distance
931    for (std::pair<Temp, uint32_t> pair : ctx.spills_entry[block_idx]) {
932       std::vector<unsigned> preds =
933          pair.first.is_linear() ? block->linear_preds : block->logical_preds;
934 
935       for (unsigned pred_idx : preds) {
936          /* variable is already spilled at predecessor */
937          auto spilled = ctx.spills_exit[pred_idx].find(pair.first);
938          if (spilled != ctx.spills_exit[pred_idx].end()) {
939             if (spilled->second != pair.second)
940                ctx.add_affinity(pair.second, spilled->second);
941             continue;
942          }
943 
944          /* variable is dead at predecessor, it must be from a phi: this works because of CSSA form */
945          if (!ctx.next_use_distances_end[pred_idx].count(pair.first))
946             continue;
947 
948          /* add interferences between spilled variable and predecessors exit spills */
949          for (std::pair<Temp, uint32_t> exit_spill : ctx.spills_exit[pred_idx]) {
950             if (exit_spill.first == pair.first)
951                continue;
952             ctx.add_interference(exit_spill.second, pair.second);
953          }
954 
955          /* variable is in register at predecessor and has to be spilled */
956          /* rename if necessary */
957          Temp var = pair.first;
958          std::map<Temp, Temp>::iterator rename_it = ctx.renames[pred_idx].find(var);
959          if (rename_it != ctx.renames[pred_idx].end()) {
960             var = rename_it->second;
961             ctx.renames[pred_idx].erase(rename_it);
962          }
963 
964          aco_ptr<Pseudo_instruction> spill{
965             create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
966          spill->operands[0] = Operand(var);
967          spill->operands[1] = Operand::c32(pair.second);
968          Block& pred = ctx.program->blocks[pred_idx];
969          unsigned idx = pred.instructions.size();
970          do {
971             assert(idx != 0);
972             idx--;
973          } while (pair.first.type() == RegType::vgpr &&
974                   pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
975          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
976          pred.instructions.insert(it, std::move(spill));
977          ctx.spills_exit[pred.index][pair.first] = pair.second;
978       }
979    }
980 
981    /* iterate phis for which operands to reload */
982    for (aco_ptr<Instruction>& phi : instructions) {
983       assert(phi->opcode == aco_opcode::p_phi || phi->opcode == aco_opcode::p_linear_phi);
984       assert(!phi->definitions[0].isTemp() ||
985              !ctx.spills_entry[block_idx].count(phi->definitions[0].getTemp()));
986 
987       std::vector<unsigned>& preds =
988          phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
989       for (unsigned i = 0; i < phi->operands.size(); i++) {
990          if (!phi->operands[i].isTemp())
991             continue;
992          unsigned pred_idx = preds[i];
993 
994          /* if the operand was reloaded, rename */
995          if (!ctx.spills_exit[pred_idx].count(phi->operands[i].getTemp())) {
996             std::map<Temp, Temp>::iterator it =
997                ctx.renames[pred_idx].find(phi->operands[i].getTemp());
998             if (it != ctx.renames[pred_idx].end()) {
999                phi->operands[i].setTemp(it->second);
1000             /* prevent the definining instruction from being DCE'd if it could be rematerialized */
1001             } else {
1002                auto remat_it = ctx.remat.find(phi->operands[i].getTemp());
1003                if (remat_it != ctx.remat.end()) {
1004                   ctx.unused_remats.erase(remat_it->second.instr);
1005                }
1006             }
1007             continue;
1008          }
1009 
1010          Temp tmp = phi->operands[i].getTemp();
1011 
1012          /* reload phi operand at end of predecessor block */
1013          Temp new_name = ctx.program->allocateTmp(tmp.regClass());
1014          Block& pred = ctx.program->blocks[pred_idx];
1015          unsigned idx = pred.instructions.size();
1016          do {
1017             assert(idx != 0);
1018             idx--;
1019          } while (phi->opcode == aco_opcode::p_phi &&
1020                   pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
1021          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
1022          aco_ptr<Instruction> reload =
1023             do_reload(ctx, tmp, new_name, ctx.spills_exit[pred_idx][tmp]);
1024 
1025          /* reload spilled exec mask directly to exec */
1026          if (!phi->definitions[0].isTemp()) {
1027             assert(phi->definitions[0].isFixed() && phi->definitions[0].physReg() == exec);
1028             reload->definitions[0] = phi->definitions[0];
1029             phi->operands[i] = Operand(exec, ctx.program->lane_mask);
1030          } else {
1031             ctx.spills_exit[pred_idx].erase(tmp);
1032             ctx.renames[pred_idx][tmp] = new_name;
1033             phi->operands[i].setTemp(new_name);
1034          }
1035 
1036          pred.instructions.insert(it, std::move(reload));
1037       }
1038    }
1039 
1040    /* iterate live variables for which to reload */
1041    // TODO: reload at current block if variable is spilled on all predecessors
1042    for (std::pair<const Temp, std::pair<uint32_t, uint32_t>>& pair :
1043         ctx.next_use_distances_start[block_idx]) {
1044       /* skip spilled variables */
1045       if (ctx.spills_entry[block_idx].count(pair.first))
1046          continue;
1047       std::vector<unsigned> preds =
1048          pair.first.is_linear() ? block->linear_preds : block->logical_preds;
1049 
1050       /* variable is dead at predecessor, it must be from a phi */
1051       bool is_dead = false;
1052       for (unsigned pred_idx : preds) {
1053          if (!ctx.next_use_distances_end[pred_idx].count(pair.first))
1054             is_dead = true;
1055       }
1056       if (is_dead)
1057          continue;
1058       for (unsigned pred_idx : preds) {
1059          /* the variable is not spilled at the predecessor */
1060          if (!ctx.spills_exit[pred_idx].count(pair.first))
1061             continue;
1062 
1063          /* variable is spilled at predecessor and has to be reloaded */
1064          Temp new_name = ctx.program->allocateTmp(pair.first.regClass());
1065          Block& pred = ctx.program->blocks[pred_idx];
1066          unsigned idx = pred.instructions.size();
1067          do {
1068             assert(idx != 0);
1069             idx--;
1070          } while (pair.first.type() == RegType::vgpr &&
1071                   pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
1072          std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
1073 
1074          aco_ptr<Instruction> reload =
1075             do_reload(ctx, pair.first, new_name, ctx.spills_exit[pred.index][pair.first]);
1076          pred.instructions.insert(it, std::move(reload));
1077 
1078          ctx.spills_exit[pred.index].erase(pair.first);
1079          ctx.renames[pred.index][pair.first] = new_name;
1080       }
1081 
1082       /* check if we have to create a new phi for this variable */
1083       Temp rename = Temp();
1084       bool is_same = true;
1085       for (unsigned pred_idx : preds) {
1086          if (!ctx.renames[pred_idx].count(pair.first)) {
1087             if (rename == Temp())
1088                rename = pair.first;
1089             else
1090                is_same = rename == pair.first;
1091          } else {
1092             if (rename == Temp())
1093                rename = ctx.renames[pred_idx][pair.first];
1094             else
1095                is_same = rename == ctx.renames[pred_idx][pair.first];
1096          }
1097 
1098          if (!is_same)
1099             break;
1100       }
1101 
1102       if (!is_same) {
1103          /* the variable was renamed differently in the predecessors: we have to create a phi */
1104          aco_opcode opcode = pair.first.is_linear() ? aco_opcode::p_linear_phi : aco_opcode::p_phi;
1105          aco_ptr<Pseudo_instruction> phi{
1106             create_instruction<Pseudo_instruction>(opcode, Format::PSEUDO, preds.size(), 1)};
1107          rename = ctx.program->allocateTmp(pair.first.regClass());
1108          for (unsigned i = 0; i < phi->operands.size(); i++) {
1109             Temp tmp;
1110             if (ctx.renames[preds[i]].count(pair.first)) {
1111                tmp = ctx.renames[preds[i]][pair.first];
1112             } else if (preds[i] >= block_idx) {
1113                tmp = rename;
1114             } else {
1115                tmp = pair.first;
1116                /* prevent the definining instruction from being DCE'd if it could be rematerialized */
1117                if (ctx.remat.count(tmp))
1118                   ctx.unused_remats.erase(ctx.remat[tmp].instr);
1119             }
1120             phi->operands[i] = Operand(tmp);
1121          }
1122          phi->definitions[0] = Definition(rename);
1123          instructions.emplace_back(std::move(phi));
1124       }
1125 
1126       /* the variable was renamed: add new name to renames */
1127       if (!(rename == Temp() || rename == pair.first))
1128          ctx.renames[block_idx][pair.first] = rename;
1129    }
1130 
1131    /* combine phis with instructions */
1132    unsigned idx = 0;
1133    while (!block->instructions[idx]) {
1134       idx++;
1135    }
1136 
1137    if (!ctx.processed[block_idx]) {
1138       assert(!(block->kind & block_kind_loop_header));
1139       RegisterDemand demand_before = get_demand_before(ctx, block_idx, idx);
1140       ctx.register_demand[block->index].erase(ctx.register_demand[block->index].begin(),
1141                                               ctx.register_demand[block->index].begin() + idx);
1142       ctx.register_demand[block->index].insert(ctx.register_demand[block->index].begin(),
1143                                                instructions.size(), demand_before);
1144    }
1145 
1146    std::vector<aco_ptr<Instruction>>::iterator start = std::next(block->instructions.begin(), idx);
1147    instructions.insert(
1148       instructions.end(), std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(start),
1149       std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
1150    block->instructions = std::move(instructions);
1151 }
1152 
1153 void
process_block(spill_ctx & ctx,unsigned block_idx,Block * block,RegisterDemand spilled_registers)1154 process_block(spill_ctx& ctx, unsigned block_idx, Block* block, RegisterDemand spilled_registers)
1155 {
1156    assert(!ctx.processed[block_idx]);
1157 
1158    std::vector<aco_ptr<Instruction>> instructions;
1159    unsigned idx = 0;
1160 
1161    /* phis are handled separetely */
1162    while (block->instructions[idx]->opcode == aco_opcode::p_phi ||
1163           block->instructions[idx]->opcode == aco_opcode::p_linear_phi) {
1164       instructions.emplace_back(std::move(block->instructions[idx++]));
1165    }
1166 
1167    if (block->register_demand.exceeds(ctx.target_pressure)) {
1168       update_local_next_uses(ctx, block, ctx.local_next_use_distance);
1169    } else {
1170       /* We won't use local_next_use_distance, so no initialization needed */
1171    }
1172 
1173    auto& current_spills = ctx.spills_exit[block_idx];
1174 
1175    while (idx < block->instructions.size()) {
1176       aco_ptr<Instruction>& instr = block->instructions[idx];
1177 
1178       std::map<Temp, std::pair<Temp, uint32_t>> reloads;
1179 
1180       /* rename and reload operands */
1181       for (Operand& op : instr->operands) {
1182          if (!op.isTemp())
1183             continue;
1184          if (!current_spills.count(op.getTemp())) {
1185             /* the Operand is in register: check if it was renamed */
1186             auto rename_it = ctx.renames[block_idx].find(op.getTemp());
1187             if (rename_it != ctx.renames[block_idx].end()) {
1188                op.setTemp(rename_it->second);
1189             } else {
1190                /* prevent its definining instruction from being DCE'd if it could be rematerialized */
1191                auto remat_it = ctx.remat.find(op.getTemp());
1192                if (remat_it != ctx.remat.end()) {
1193                   ctx.unused_remats.erase(remat_it->second.instr);
1194                }
1195             }
1196             continue;
1197          }
1198          /* the Operand is spilled: add it to reloads */
1199          Temp new_tmp = ctx.program->allocateTmp(op.regClass());
1200          ctx.renames[block_idx][op.getTemp()] = new_tmp;
1201          reloads[new_tmp] = std::make_pair(op.getTemp(), current_spills[op.getTemp()]);
1202          current_spills.erase(op.getTemp());
1203          op.setTemp(new_tmp);
1204          spilled_registers -= new_tmp;
1205       }
1206 
1207       /* check if register demand is low enough before and after the current instruction */
1208       if (block->register_demand.exceeds(ctx.target_pressure)) {
1209 
1210          RegisterDemand new_demand = ctx.register_demand[block_idx][idx];
1211          new_demand.update(get_demand_before(ctx, block_idx, idx));
1212 
1213          assert(!ctx.local_next_use_distance.empty());
1214 
1215          /* if reg pressure is too high, spill variable with furthest next use */
1216          while ((new_demand - spilled_registers).exceeds(ctx.target_pressure)) {
1217             unsigned distance = 0;
1218             Temp to_spill;
1219             bool do_rematerialize = false;
1220             RegType type = RegType::sgpr;
1221             if (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr)
1222                type = RegType::vgpr;
1223 
1224             for (std::pair<Temp, uint32_t> pair : ctx.local_next_use_distance[idx]) {
1225                if (pair.first.type() != type)
1226                   continue;
1227                bool can_rematerialize = ctx.remat.count(pair.first);
1228                if (((pair.second > distance && can_rematerialize == do_rematerialize) ||
1229                     (can_rematerialize && !do_rematerialize && pair.second > idx)) &&
1230                    !current_spills.count(pair.first)) {
1231                   to_spill = pair.first;
1232                   distance = pair.second;
1233                   do_rematerialize = can_rematerialize;
1234                }
1235             }
1236 
1237             assert(distance != 0 && distance > idx);
1238             uint32_t spill_id = ctx.allocate_spill_id(to_spill.regClass());
1239 
1240             /* add interferences with currently spilled variables */
1241             for (std::pair<Temp, uint32_t> pair : current_spills)
1242                ctx.add_interference(spill_id, pair.second);
1243             for (std::pair<const Temp, std::pair<Temp, uint32_t>>& pair : reloads)
1244                ctx.add_interference(spill_id, pair.second.second);
1245 
1246             current_spills[to_spill] = spill_id;
1247             spilled_registers += to_spill;
1248 
1249             /* rename if necessary */
1250             if (ctx.renames[block_idx].count(to_spill)) {
1251                to_spill = ctx.renames[block_idx][to_spill];
1252             }
1253 
1254             /* add spill to new instructions */
1255             aco_ptr<Pseudo_instruction> spill{
1256                create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
1257             spill->operands[0] = Operand(to_spill);
1258             spill->operands[1] = Operand::c32(spill_id);
1259             instructions.emplace_back(std::move(spill));
1260          }
1261       }
1262 
1263       /* add reloads and instruction to new instructions */
1264       for (std::pair<const Temp, std::pair<Temp, uint32_t>>& pair : reloads) {
1265          aco_ptr<Instruction> reload =
1266             do_reload(ctx, pair.second.first, pair.first, pair.second.second);
1267          instructions.emplace_back(std::move(reload));
1268       }
1269       instructions.emplace_back(std::move(instr));
1270       idx++;
1271    }
1272 
1273    block->instructions = std::move(instructions);
1274 }
1275 
1276 void
spill_block(spill_ctx & ctx,unsigned block_idx)1277 spill_block(spill_ctx& ctx, unsigned block_idx)
1278 {
1279    Block* block = &ctx.program->blocks[block_idx];
1280 
1281    /* determine set of variables which are spilled at the beginning of the block */
1282    RegisterDemand spilled_registers = init_live_in_vars(ctx, block, block_idx);
1283 
1284    /* add interferences for spilled variables */
1285    for (auto it = ctx.spills_entry[block_idx].begin(); it != ctx.spills_entry[block_idx].end();
1286         ++it) {
1287       for (auto it2 = std::next(it); it2 != ctx.spills_entry[block_idx].end(); ++it2)
1288          ctx.add_interference(it->second, it2->second);
1289    }
1290 
1291    bool is_loop_header = block->loop_nest_depth && ctx.loop_header.top()->index == block_idx;
1292    if (!is_loop_header) {
1293       /* add spill/reload code on incoming control flow edges */
1294       add_coupling_code(ctx, block, block_idx);
1295    }
1296 
1297    const auto& current_spills = ctx.spills_entry[block_idx];
1298 
1299    /* check conditions to process this block */
1300    bool process = (block->register_demand - spilled_registers).exceeds(ctx.target_pressure) ||
1301                   !ctx.renames[block_idx].empty() || ctx.unused_remats.size();
1302 
1303    for (auto it = current_spills.begin(); !process && it != current_spills.end(); ++it) {
1304       if (ctx.next_use_distances_start[block_idx].at(it->first).first == block_idx)
1305          process = true;
1306    }
1307 
1308    assert(ctx.spills_exit[block_idx].empty());
1309    ctx.spills_exit[block_idx] = current_spills;
1310    if (process) {
1311       process_block(ctx, block_idx, block, spilled_registers);
1312    }
1313 
1314    ctx.processed[block_idx] = true;
1315 
1316    /* check if the next block leaves the current loop */
1317    if (block->loop_nest_depth == 0 ||
1318        ctx.program->blocks[block_idx + 1].loop_nest_depth >= block->loop_nest_depth)
1319       return;
1320 
1321    Block* loop_header = ctx.loop_header.top();
1322 
1323    /* preserve original renames at end of loop header block */
1324    std::map<Temp, Temp> renames = std::move(ctx.renames[loop_header->index]);
1325 
1326    /* add coupling code to all loop header predecessors */
1327    add_coupling_code(ctx, loop_header, loop_header->index);
1328 
1329    /* propagate new renames through loop: i.e. repair the SSA */
1330    renames.swap(ctx.renames[loop_header->index]);
1331    for (std::pair<Temp, Temp> rename : renames) {
1332       for (unsigned idx = loop_header->index; idx <= block_idx; idx++) {
1333          Block& current = ctx.program->blocks[idx];
1334          std::vector<aco_ptr<Instruction>>::iterator instr_it = current.instructions.begin();
1335 
1336          /* first rename phis */
1337          while (instr_it != current.instructions.end()) {
1338             aco_ptr<Instruction>& phi = *instr_it;
1339             if (phi->opcode != aco_opcode::p_phi && phi->opcode != aco_opcode::p_linear_phi)
1340                break;
1341             /* no need to rename the loop header phis once again. this happened in
1342              * add_coupling_code() */
1343             if (idx == loop_header->index) {
1344                instr_it++;
1345                continue;
1346             }
1347 
1348             for (Operand& op : phi->operands) {
1349                if (!op.isTemp())
1350                   continue;
1351                if (op.getTemp() == rename.first)
1352                   op.setTemp(rename.second);
1353             }
1354             instr_it++;
1355          }
1356 
1357          /* variable is not live at beginning of this block */
1358          if (ctx.next_use_distances_start[idx].count(rename.first) == 0)
1359             continue;
1360 
1361          /* if the variable is live at the block's exit, add rename */
1362          if (ctx.next_use_distances_end[idx].count(rename.first) != 0)
1363             ctx.renames[idx].insert(rename);
1364 
1365          /* rename all uses in this block */
1366          bool renamed = false;
1367          while (!renamed && instr_it != current.instructions.end()) {
1368             aco_ptr<Instruction>& instr = *instr_it;
1369             for (Operand& op : instr->operands) {
1370                if (!op.isTemp())
1371                   continue;
1372                if (op.getTemp() == rename.first) {
1373                   op.setTemp(rename.second);
1374                   /* we can stop with this block as soon as the variable is spilled */
1375                   if (instr->opcode == aco_opcode::p_spill)
1376                      renamed = true;
1377                }
1378             }
1379             instr_it++;
1380          }
1381       }
1382    }
1383 
1384    /* remove loop header info from stack */
1385    ctx.loop_header.pop();
1386 }
1387 
1388 Temp
load_scratch_resource(spill_ctx & ctx,Temp & scratch_offset,std::vector<aco_ptr<Instruction>> & instructions,unsigned offset,bool is_top_level)1389 load_scratch_resource(spill_ctx& ctx, Temp& scratch_offset,
1390                       std::vector<aco_ptr<Instruction>>& instructions, unsigned offset,
1391                       bool is_top_level)
1392 {
1393    Builder bld(ctx.program);
1394    if (is_top_level) {
1395       bld.reset(&instructions);
1396    } else {
1397       /* find p_logical_end */
1398       unsigned idx = instructions.size() - 1;
1399       while (instructions[idx]->opcode != aco_opcode::p_logical_end)
1400          idx--;
1401       bld.reset(&instructions, std::next(instructions.begin(), idx));
1402    }
1403 
1404    Temp private_segment_buffer = ctx.program->private_segment_buffer;
1405    if (ctx.program->stage != compute_cs)
1406       private_segment_buffer =
1407          bld.smem(aco_opcode::s_load_dwordx2, bld.def(s2), private_segment_buffer, Operand::zero());
1408 
1409    if (offset)
1410       scratch_offset = bld.sop2(aco_opcode::s_add_u32, bld.def(s1), bld.def(s1, scc),
1411                                 scratch_offset, Operand::c32(offset));
1412 
1413    uint32_t rsrc_conf =
1414       S_008F0C_ADD_TID_ENABLE(1) | S_008F0C_INDEX_STRIDE(ctx.program->wave_size == 64 ? 3 : 2);
1415 
1416    if (ctx.program->chip_class >= GFX10) {
1417       rsrc_conf |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
1418                    S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) | S_008F0C_RESOURCE_LEVEL(1);
1419    } else if (ctx.program->chip_class <= GFX7) {
1420       /* dfmt modifies stride on GFX8/GFX9 when ADD_TID_EN=1 */
1421       rsrc_conf |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
1422                    S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
1423    }
1424    /* older generations need element size = 4 bytes. element size removed in GFX9 */
1425    if (ctx.program->chip_class <= GFX8)
1426       rsrc_conf |= S_008F0C_ELEMENT_SIZE(1);
1427 
1428    return bld.pseudo(aco_opcode::p_create_vector, bld.def(s4), private_segment_buffer,
1429                      Operand::c32(-1u), Operand::c32(rsrc_conf));
1430 }
1431 
1432 void
add_interferences(spill_ctx & ctx,std::vector<bool> & is_assigned,std::vector<uint32_t> & slots,std::vector<bool> & slots_used,unsigned id)1433 add_interferences(spill_ctx& ctx, std::vector<bool>& is_assigned, std::vector<uint32_t>& slots,
1434                   std::vector<bool>& slots_used, unsigned id)
1435 {
1436    for (unsigned other : ctx.interferences[id].second) {
1437       if (!is_assigned[other])
1438          continue;
1439 
1440       RegClass other_rc = ctx.interferences[other].first;
1441       unsigned slot = slots[other];
1442       std::fill(slots_used.begin() + slot, slots_used.begin() + slot + other_rc.size(), true);
1443    }
1444 }
1445 
1446 unsigned
find_available_slot(std::vector<bool> & used,unsigned wave_size,unsigned size,bool is_sgpr,unsigned * num_slots)1447 find_available_slot(std::vector<bool>& used, unsigned wave_size, unsigned size, bool is_sgpr,
1448                     unsigned* num_slots)
1449 {
1450    unsigned wave_size_minus_one = wave_size - 1;
1451    unsigned slot = 0;
1452 
1453    while (true) {
1454       bool available = true;
1455       for (unsigned i = 0; i < size; i++) {
1456          if (slot + i < used.size() && used[slot + i]) {
1457             available = false;
1458             break;
1459          }
1460       }
1461       if (!available) {
1462          slot++;
1463          continue;
1464       }
1465 
1466       if (is_sgpr && ((slot & wave_size_minus_one) > wave_size - size)) {
1467          slot = align(slot, wave_size);
1468          continue;
1469       }
1470 
1471       std::fill(used.begin(), used.end(), false);
1472 
1473       if (slot + size > used.size())
1474          used.resize(slot + size);
1475 
1476       return slot;
1477    }
1478 }
1479 
1480 void
assign_spill_slots_helper(spill_ctx & ctx,RegType type,std::vector<bool> & is_assigned,std::vector<uint32_t> & slots,unsigned * num_slots)1481 assign_spill_slots_helper(spill_ctx& ctx, RegType type, std::vector<bool>& is_assigned,
1482                           std::vector<uint32_t>& slots, unsigned* num_slots)
1483 {
1484    std::vector<bool> slots_used(*num_slots);
1485 
1486    /* assign slots for ids with affinities first */
1487    for (std::vector<uint32_t>& vec : ctx.affinities) {
1488       if (ctx.interferences[vec[0]].first.type() != type)
1489          continue;
1490 
1491       for (unsigned id : vec) {
1492          if (!ctx.is_reloaded[id])
1493             continue;
1494 
1495          add_interferences(ctx, is_assigned, slots, slots_used, id);
1496       }
1497 
1498       unsigned slot =
1499          find_available_slot(slots_used, ctx.wave_size, ctx.interferences[vec[0]].first.size(),
1500                              type == RegType::sgpr, num_slots);
1501 
1502       for (unsigned id : vec) {
1503          assert(!is_assigned[id]);
1504 
1505          if (ctx.is_reloaded[id]) {
1506             slots[id] = slot;
1507             is_assigned[id] = true;
1508          }
1509       }
1510    }
1511 
1512    /* assign slots for ids without affinities */
1513    for (unsigned id = 0; id < ctx.interferences.size(); id++) {
1514       if (is_assigned[id] || !ctx.is_reloaded[id] || ctx.interferences[id].first.type() != type)
1515          continue;
1516 
1517       add_interferences(ctx, is_assigned, slots, slots_used, id);
1518 
1519       unsigned slot =
1520          find_available_slot(slots_used, ctx.wave_size, ctx.interferences[id].first.size(),
1521                              type == RegType::sgpr, num_slots);
1522 
1523       slots[id] = slot;
1524       is_assigned[id] = true;
1525    }
1526 
1527    *num_slots = slots_used.size();
1528 }
1529 
1530 void
assign_spill_slots(spill_ctx & ctx,unsigned spills_to_vgpr)1531 assign_spill_slots(spill_ctx& ctx, unsigned spills_to_vgpr)
1532 {
1533    std::vector<uint32_t> slots(ctx.interferences.size());
1534    std::vector<bool> is_assigned(ctx.interferences.size());
1535 
1536    /* first, handle affinities: just merge all interferences into both spill ids */
1537    for (std::vector<uint32_t>& vec : ctx.affinities) {
1538       for (unsigned i = 0; i < vec.size(); i++) {
1539          for (unsigned j = i + 1; j < vec.size(); j++) {
1540             assert(vec[i] != vec[j]);
1541             bool reloaded = ctx.is_reloaded[vec[i]] || ctx.is_reloaded[vec[j]];
1542             ctx.is_reloaded[vec[i]] = reloaded;
1543             ctx.is_reloaded[vec[j]] = reloaded;
1544          }
1545       }
1546    }
1547    for (ASSERTED uint32_t i = 0; i < ctx.interferences.size(); i++)
1548       for (ASSERTED uint32_t id : ctx.interferences[i].second)
1549          assert(i != id);
1550 
1551    /* for each spill slot, assign as many spill ids as possible */
1552    unsigned sgpr_spill_slots = 0, vgpr_spill_slots = 0;
1553    assign_spill_slots_helper(ctx, RegType::sgpr, is_assigned, slots, &sgpr_spill_slots);
1554    assign_spill_slots_helper(ctx, RegType::vgpr, is_assigned, slots, &vgpr_spill_slots);
1555 
1556    for (unsigned id = 0; id < is_assigned.size(); id++)
1557       assert(is_assigned[id] || !ctx.is_reloaded[id]);
1558 
1559    for (std::vector<uint32_t>& vec : ctx.affinities) {
1560       for (unsigned i = 0; i < vec.size(); i++) {
1561          for (unsigned j = i + 1; j < vec.size(); j++) {
1562             assert(is_assigned[vec[i]] == is_assigned[vec[j]]);
1563             if (!is_assigned[vec[i]])
1564                continue;
1565             assert(ctx.is_reloaded[vec[i]] == ctx.is_reloaded[vec[j]]);
1566             assert(ctx.interferences[vec[i]].first.type() ==
1567                    ctx.interferences[vec[j]].first.type());
1568             assert(slots[vec[i]] == slots[vec[j]]);
1569          }
1570       }
1571    }
1572 
1573    /* hope, we didn't mess up */
1574    std::vector<Temp> vgpr_spill_temps((sgpr_spill_slots + ctx.wave_size - 1) / ctx.wave_size);
1575    assert(vgpr_spill_temps.size() <= spills_to_vgpr);
1576 
1577    /* replace pseudo instructions with actual hardware instructions */
1578    Temp scratch_offset = ctx.program->scratch_offset, scratch_rsrc = Temp();
1579    unsigned last_top_level_block_idx = 0;
1580    std::vector<bool> reload_in_loop(vgpr_spill_temps.size());
1581    for (Block& block : ctx.program->blocks) {
1582 
1583       /* after loops, we insert a user if there was a reload inside the loop */
1584       if (block.loop_nest_depth == 0) {
1585          int end_vgprs = 0;
1586          for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1587             if (reload_in_loop[i])
1588                end_vgprs++;
1589          }
1590 
1591          if (end_vgprs > 0) {
1592             aco_ptr<Instruction> destr{create_instruction<Pseudo_instruction>(
1593                aco_opcode::p_end_linear_vgpr, Format::PSEUDO, end_vgprs, 0)};
1594             int k = 0;
1595             for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1596                if (reload_in_loop[i])
1597                   destr->operands[k++] = Operand(vgpr_spill_temps[i]);
1598                reload_in_loop[i] = false;
1599             }
1600             /* find insertion point */
1601             std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
1602             while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
1603                ++it;
1604             block.instructions.insert(it, std::move(destr));
1605          }
1606       }
1607 
1608       if (block.kind & block_kind_top_level && !block.linear_preds.empty()) {
1609          last_top_level_block_idx = block.index;
1610 
1611          /* check if any spilled variables use a created linear vgpr, otherwise destroy them */
1612          for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1613             if (vgpr_spill_temps[i] == Temp())
1614                continue;
1615 
1616             bool can_destroy = true;
1617             for (std::pair<Temp, uint32_t> pair : ctx.spills_entry[block.index]) {
1618 
1619                if (ctx.interferences[pair.second].first.type() == RegType::sgpr &&
1620                    slots[pair.second] / ctx.wave_size == i) {
1621                   can_destroy = false;
1622                   break;
1623                }
1624             }
1625             if (can_destroy)
1626                vgpr_spill_temps[i] = Temp();
1627          }
1628       }
1629 
1630       std::vector<aco_ptr<Instruction>>::iterator it;
1631       std::vector<aco_ptr<Instruction>> instructions;
1632       instructions.reserve(block.instructions.size());
1633       Builder bld(ctx.program, &instructions);
1634       for (it = block.instructions.begin(); it != block.instructions.end(); ++it) {
1635 
1636          if ((*it)->opcode == aco_opcode::p_spill) {
1637             uint32_t spill_id = (*it)->operands[1].constantValue();
1638 
1639             if (!ctx.is_reloaded[spill_id]) {
1640                /* never reloaded, so don't spill */
1641             } else if (!is_assigned[spill_id]) {
1642                unreachable("No spill slot assigned for spill id");
1643             } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) {
1644                /* spill vgpr */
1645                ctx.program->config->spilled_vgprs += (*it)->operands[0].size();
1646                uint32_t spill_slot = slots[spill_id];
1647                bool add_offset_to_sgpr =
1648                   ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size +
1649                      vgpr_spill_slots * 4 >
1650                   4096;
1651                unsigned base_offset =
1652                   add_offset_to_sgpr
1653                      ? 0
1654                      : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
1655 
1656                /* check if the scratch resource descriptor already exists */
1657                if (scratch_rsrc == Temp()) {
1658                   unsigned offset =
1659                      add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
1660                   scratch_rsrc = load_scratch_resource(
1661                      ctx, scratch_offset,
1662                      last_top_level_block_idx == block.index
1663                         ? instructions
1664                         : ctx.program->blocks[last_top_level_block_idx].instructions,
1665                      offset, last_top_level_block_idx == block.index);
1666                }
1667 
1668                unsigned offset = base_offset + spill_slot * 4;
1669                aco_opcode opcode = aco_opcode::buffer_store_dword;
1670                assert((*it)->operands[0].isTemp());
1671                Temp temp = (*it)->operands[0].getTemp();
1672                assert(temp.type() == RegType::vgpr && !temp.is_linear());
1673                if (temp.size() > 1) {
1674                   Instruction* split{create_instruction<Pseudo_instruction>(
1675                      aco_opcode::p_split_vector, Format::PSEUDO, 1, temp.size())};
1676                   split->operands[0] = Operand(temp);
1677                   for (unsigned i = 0; i < temp.size(); i++)
1678                      split->definitions[i] = bld.def(v1);
1679                   bld.insert(split);
1680                   for (unsigned i = 0; i < temp.size(); i++) {
1681                      Instruction* instr =
1682                         bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset,
1683                                   split->definitions[i].getTemp(), offset + i * 4, false, true);
1684                      instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1685                   }
1686                } else {
1687                   Instruction* instr = bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset,
1688                                                  temp, offset, false, true);
1689                   instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1690                }
1691             } else {
1692                ctx.program->config->spilled_sgprs += (*it)->operands[0].size();
1693 
1694                uint32_t spill_slot = slots[spill_id];
1695 
1696                /* check if the linear vgpr already exists */
1697                if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) {
1698                   Temp linear_vgpr = ctx.program->allocateTmp(v1.as_linear());
1699                   vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr;
1700                   aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(
1701                      aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
1702                   create->definitions[0] = Definition(linear_vgpr);
1703                   /* find the right place to insert this definition */
1704                   if (last_top_level_block_idx == block.index) {
1705                      /* insert right before the current instruction */
1706                      instructions.emplace_back(std::move(create));
1707                   } else {
1708                      assert(last_top_level_block_idx < block.index);
1709                      /* insert before the branch at last top level block */
1710                      std::vector<aco_ptr<Instruction>>& block_instrs =
1711                         ctx.program->blocks[last_top_level_block_idx].instructions;
1712                      block_instrs.insert(std::prev(block_instrs.end()), std::move(create));
1713                   }
1714                }
1715 
1716                /* spill sgpr: just add the vgpr temp to operands */
1717                Pseudo_instruction* spill =
1718                   create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 3, 0);
1719                spill->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]);
1720                spill->operands[1] = Operand::c32(spill_slot % ctx.wave_size);
1721                spill->operands[2] = (*it)->operands[0];
1722                instructions.emplace_back(aco_ptr<Instruction>(spill));
1723             }
1724 
1725          } else if ((*it)->opcode == aco_opcode::p_reload) {
1726             uint32_t spill_id = (*it)->operands[0].constantValue();
1727             assert(ctx.is_reloaded[spill_id]);
1728 
1729             if (!is_assigned[spill_id]) {
1730                unreachable("No spill slot assigned for spill id");
1731             } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) {
1732                /* reload vgpr */
1733                uint32_t spill_slot = slots[spill_id];
1734                bool add_offset_to_sgpr =
1735                   ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size +
1736                      vgpr_spill_slots * 4 >
1737                   4096;
1738                unsigned base_offset =
1739                   add_offset_to_sgpr
1740                      ? 0
1741                      : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
1742 
1743                /* check if the scratch resource descriptor already exists */
1744                if (scratch_rsrc == Temp()) {
1745                   unsigned offset =
1746                      add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
1747                   scratch_rsrc = load_scratch_resource(
1748                      ctx, scratch_offset,
1749                      last_top_level_block_idx == block.index
1750                         ? instructions
1751                         : ctx.program->blocks[last_top_level_block_idx].instructions,
1752                      offset, last_top_level_block_idx == block.index);
1753                }
1754 
1755                unsigned offset = base_offset + spill_slot * 4;
1756                aco_opcode opcode = aco_opcode::buffer_load_dword;
1757                Definition def = (*it)->definitions[0];
1758                if (def.size() > 1) {
1759                   Instruction* vec{create_instruction<Pseudo_instruction>(
1760                      aco_opcode::p_create_vector, Format::PSEUDO, def.size(), 1)};
1761                   vec->definitions[0] = def;
1762                   for (unsigned i = 0; i < def.size(); i++) {
1763                      Temp tmp = bld.tmp(v1);
1764                      vec->operands[i] = Operand(tmp);
1765                      Instruction* instr =
1766                         bld.mubuf(opcode, Definition(tmp), scratch_rsrc, Operand(v1),
1767                                   scratch_offset, offset + i * 4, false, true);
1768                      instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1769                   }
1770                   bld.insert(vec);
1771                } else {
1772                   Instruction* instr = bld.mubuf(opcode, def, scratch_rsrc, Operand(v1),
1773                                                  scratch_offset, offset, false, true);
1774                   instr->mubuf().sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1775                }
1776             } else {
1777                uint32_t spill_slot = slots[spill_id];
1778                reload_in_loop[spill_slot / ctx.wave_size] = block.loop_nest_depth > 0;
1779 
1780                /* check if the linear vgpr already exists */
1781                if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) {
1782                   Temp linear_vgpr = ctx.program->allocateTmp(v1.as_linear());
1783                   vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr;
1784                   aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(
1785                      aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
1786                   create->definitions[0] = Definition(linear_vgpr);
1787                   /* find the right place to insert this definition */
1788                   if (last_top_level_block_idx == block.index) {
1789                      /* insert right before the current instruction */
1790                      instructions.emplace_back(std::move(create));
1791                   } else {
1792                      assert(last_top_level_block_idx < block.index);
1793                      /* insert before the branch at last top level block */
1794                      std::vector<aco_ptr<Instruction>>& block_instrs =
1795                         ctx.program->blocks[last_top_level_block_idx].instructions;
1796                      block_instrs.insert(std::prev(block_instrs.end()), std::move(create));
1797                   }
1798                }
1799 
1800                /* reload sgpr: just add the vgpr temp to operands */
1801                Pseudo_instruction* reload = create_instruction<Pseudo_instruction>(
1802                   aco_opcode::p_reload, Format::PSEUDO, 2, 1);
1803                reload->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]);
1804                reload->operands[1] = Operand::c32(spill_slot % ctx.wave_size);
1805                reload->definitions[0] = (*it)->definitions[0];
1806                instructions.emplace_back(aco_ptr<Instruction>(reload));
1807             }
1808          } else if (!ctx.unused_remats.count(it->get())) {
1809             instructions.emplace_back(std::move(*it));
1810          }
1811       }
1812       block.instructions = std::move(instructions);
1813    }
1814 
1815    /* update required scratch memory */
1816    ctx.program->config->scratch_bytes_per_wave +=
1817       align(vgpr_spill_slots * 4 * ctx.program->wave_size, 1024);
1818 
1819    /* SSA elimination inserts copies for logical phis right before p_logical_end
1820     * So if a linear vgpr is used between that p_logical_end and the branch,
1821     * we need to ensure logical phis don't choose a definition which aliases
1822     * the linear vgpr.
1823     * TODO: Moving the spills and reloads to before p_logical_end might produce
1824     *       slightly better code. */
1825    for (Block& block : ctx.program->blocks) {
1826       /* loops exits are already handled */
1827       if (block.logical_preds.size() <= 1)
1828          continue;
1829 
1830       bool has_logical_phis = false;
1831       for (aco_ptr<Instruction>& instr : block.instructions) {
1832          if (instr->opcode == aco_opcode::p_phi) {
1833             has_logical_phis = true;
1834             break;
1835          } else if (instr->opcode != aco_opcode::p_linear_phi) {
1836             break;
1837          }
1838       }
1839       if (!has_logical_phis)
1840          continue;
1841 
1842       std::set<Temp> vgprs;
1843       for (unsigned pred_idx : block.logical_preds) {
1844          Block& pred = ctx.program->blocks[pred_idx];
1845          for (int i = pred.instructions.size() - 1; i >= 0; i--) {
1846             aco_ptr<Instruction>& pred_instr = pred.instructions[i];
1847             if (pred_instr->opcode == aco_opcode::p_logical_end) {
1848                break;
1849             } else if (pred_instr->opcode == aco_opcode::p_spill ||
1850                        pred_instr->opcode == aco_opcode::p_reload) {
1851                vgprs.insert(pred_instr->operands[0].getTemp());
1852             }
1853          }
1854       }
1855       if (!vgprs.size())
1856          continue;
1857 
1858       aco_ptr<Instruction> destr{create_instruction<Pseudo_instruction>(
1859          aco_opcode::p_end_linear_vgpr, Format::PSEUDO, vgprs.size(), 0)};
1860       int k = 0;
1861       for (Temp tmp : vgprs) {
1862          destr->operands[k++] = Operand(tmp);
1863       }
1864       /* find insertion point */
1865       std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
1866       while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
1867          ++it;
1868       block.instructions.insert(it, std::move(destr));
1869    }
1870 }
1871 
1872 } /* end namespace */
1873 
1874 void
spill(Program * program,live & live_vars)1875 spill(Program* program, live& live_vars)
1876 {
1877    program->config->spilled_vgprs = 0;
1878    program->config->spilled_sgprs = 0;
1879 
1880    program->progress = CompilationProgress::after_spilling;
1881 
1882    /* no spilling when register pressure is low enough */
1883    if (program->num_waves > 0)
1884       return;
1885 
1886    /* lower to CSSA before spilling to ensure correctness w.r.t. phis */
1887    lower_to_cssa(program, live_vars);
1888 
1889    /* calculate target register demand */
1890    const RegisterDemand demand = program->max_reg_demand; /* current max */
1891    const uint16_t sgpr_limit = get_addr_sgpr_from_waves(program, program->min_waves);
1892    const uint16_t vgpr_limit = get_addr_vgpr_from_waves(program, program->min_waves);
1893    uint16_t extra_vgprs = 0;
1894    uint16_t extra_sgprs = 0;
1895 
1896    /* calculate extra VGPRs required for spilling SGPRs */
1897    if (demand.sgpr > sgpr_limit) {
1898       unsigned sgpr_spills = demand.sgpr - sgpr_limit;
1899       extra_vgprs = DIV_ROUND_UP(sgpr_spills, program->wave_size) + 1;
1900    }
1901    /* add extra SGPRs required for spilling VGPRs */
1902    if (demand.vgpr + extra_vgprs > vgpr_limit) {
1903       extra_sgprs = 5; /* scratch_resource (s4) + scratch_offset (s1) */
1904       if (demand.sgpr + extra_sgprs > sgpr_limit) {
1905          /* re-calculate in case something has changed */
1906          unsigned sgpr_spills = demand.sgpr + extra_sgprs - sgpr_limit;
1907          extra_vgprs = DIV_ROUND_UP(sgpr_spills, program->wave_size) + 1;
1908       }
1909    }
1910    /* the spiller has to target the following register demand */
1911    const RegisterDemand target(vgpr_limit - extra_vgprs, sgpr_limit - extra_sgprs);
1912 
1913    /* initialize ctx */
1914    spill_ctx ctx(target, program, live_vars.register_demand);
1915    compute_global_next_uses(ctx);
1916    get_rematerialize_info(ctx);
1917 
1918    /* create spills and reloads */
1919    for (unsigned i = 0; i < program->blocks.size(); i++)
1920       spill_block(ctx, i);
1921 
1922    /* assign spill slots and DCE rematerialized code */
1923    assign_spill_slots(ctx, extra_vgprs);
1924 
1925    /* update live variable information */
1926    live_vars = live_var_analysis(program);
1927 
1928    assert(program->num_waves > 0);
1929 }
1930 
1931 } // namespace aco
1932