1 /*
2 * Copyright © 2018 Valve Corporation
3 * Copyright © 2018 Google
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 *
24 */
25
26 #include "aco_ir.h"
27 #include "aco_builder.h"
28 #include "sid.h"
29
30 #include <map>
31 #include <set>
32 #include <stack>
33
34 /*
35 * Implements the spilling algorithm on SSA-form from
36 * "Register Spilling and Live-Range Splitting for SSA-Form Programs"
37 * by Matthias Braun and Sebastian Hack.
38 */
39
40 namespace aco {
41
42 namespace {
43
44 struct remat_info {
45 Instruction *instr;
46 };
47
48 struct spill_ctx {
49 RegisterDemand target_pressure;
50 Program* program;
51 std::vector<std::vector<RegisterDemand>> register_demand;
52 std::vector<std::map<Temp, Temp>> renames;
53 std::vector<std::map<Temp, uint32_t>> spills_entry;
54 std::vector<std::map<Temp, uint32_t>> spills_exit;
55 std::vector<bool> processed;
56 std::stack<Block*> loop_header;
57 std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_start;
58 std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_end;
59 std::vector<std::pair<RegClass, std::unordered_set<uint32_t>>> interferences;
60 std::vector<std::vector<uint32_t>> affinities;
61 std::vector<bool> is_reloaded;
62 std::map<Temp, remat_info> remat;
63 std::map<Instruction *, bool> remat_used;
64 unsigned wave_size;
65
spill_ctxaco::__anonb331a64d0111::spill_ctx66 spill_ctx(const RegisterDemand target_pressure, Program* program,
67 std::vector<std::vector<RegisterDemand>> register_demand)
68 : target_pressure(target_pressure), program(program),
69 register_demand(std::move(register_demand)), renames(program->blocks.size()),
70 spills_entry(program->blocks.size()), spills_exit(program->blocks.size()),
71 processed(program->blocks.size(), false), wave_size(program->wave_size) {}
72
add_affinityaco::__anonb331a64d0111::spill_ctx73 void add_affinity(uint32_t first, uint32_t second)
74 {
75 unsigned found_first = affinities.size();
76 unsigned found_second = affinities.size();
77 for (unsigned i = 0; i < affinities.size(); i++) {
78 std::vector<uint32_t>& vec = affinities[i];
79 for (uint32_t entry : vec) {
80 if (entry == first)
81 found_first = i;
82 else if (entry == second)
83 found_second = i;
84 }
85 }
86 if (found_first == affinities.size() && found_second == affinities.size()) {
87 affinities.emplace_back(std::vector<uint32_t>({first, second}));
88 } else if (found_first < affinities.size() && found_second == affinities.size()) {
89 affinities[found_first].push_back(second);
90 } else if (found_second < affinities.size() && found_first == affinities.size()) {
91 affinities[found_second].push_back(first);
92 } else if (found_first != found_second) {
93 /* merge second into first */
94 affinities[found_first].insert(affinities[found_first].end(), affinities[found_second].begin(), affinities[found_second].end());
95 affinities.erase(std::next(affinities.begin(), found_second));
96 } else {
97 assert(found_first == found_second);
98 }
99 }
100
add_interferenceaco::__anonb331a64d0111::spill_ctx101 void add_interference(uint32_t first, uint32_t second)
102 {
103 if (interferences[first].first.type() != interferences[second].first.type())
104 return;
105
106 bool inserted = interferences[first].second.insert(second).second;
107 if (inserted)
108 interferences[second].second.insert(first);
109 }
110
allocate_spill_idaco::__anonb331a64d0111::spill_ctx111 uint32_t allocate_spill_id(RegClass rc)
112 {
113 interferences.emplace_back(rc, std::unordered_set<uint32_t>());
114 is_reloaded.push_back(false);
115 return next_spill_id++;
116 }
117
118 uint32_t next_spill_id = 0;
119 };
120
get_dominator(int idx_a,int idx_b,Program * program,bool is_linear)121 int32_t get_dominator(int idx_a, int idx_b, Program* program, bool is_linear)
122 {
123
124 if (idx_a == -1)
125 return idx_b;
126 if (idx_b == -1)
127 return idx_a;
128 if (is_linear) {
129 while (idx_a != idx_b) {
130 if (idx_a > idx_b)
131 idx_a = program->blocks[idx_a].linear_idom;
132 else
133 idx_b = program->blocks[idx_b].linear_idom;
134 }
135 } else {
136 while (idx_a != idx_b) {
137 if (idx_a > idx_b)
138 idx_a = program->blocks[idx_a].logical_idom;
139 else
140 idx_b = program->blocks[idx_b].logical_idom;
141 }
142 }
143 assert(idx_a != -1);
144 return idx_a;
145 }
146
next_uses_per_block(spill_ctx & ctx,unsigned block_idx,std::set<uint32_t> & worklist)147 void next_uses_per_block(spill_ctx& ctx, unsigned block_idx, std::set<uint32_t>& worklist)
148 {
149 Block* block = &ctx.program->blocks[block_idx];
150 std::map<Temp, std::pair<uint32_t, uint32_t>> next_uses = ctx.next_use_distances_end[block_idx];
151
152 /* to compute the next use distance at the beginning of the block, we have to add the block's size */
153 for (std::map<Temp, std::pair<uint32_t, uint32_t>>::iterator it = next_uses.begin(); it != next_uses.end(); ++it)
154 it->second.second = it->second.second + block->instructions.size();
155
156 int idx = block->instructions.size() - 1;
157 while (idx >= 0) {
158 aco_ptr<Instruction>& instr = block->instructions[idx];
159
160 if (instr->opcode == aco_opcode::p_linear_phi ||
161 instr->opcode == aco_opcode::p_phi)
162 break;
163
164 for (const Definition& def : instr->definitions) {
165 if (def.isTemp())
166 next_uses.erase(def.getTemp());
167 }
168
169 for (const Operand& op : instr->operands) {
170 /* omit exec mask */
171 if (op.isFixed() && op.physReg() == exec)
172 continue;
173 if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear())
174 continue;
175 if (op.isTemp())
176 next_uses[op.getTemp()] = {block_idx, idx};
177 }
178 idx--;
179 }
180
181 assert(block_idx != 0 || next_uses.empty());
182 ctx.next_use_distances_start[block_idx] = next_uses;
183 while (idx >= 0) {
184 aco_ptr<Instruction>& instr = block->instructions[idx];
185 assert(instr->opcode == aco_opcode::p_linear_phi || instr->opcode == aco_opcode::p_phi);
186
187 for (unsigned i = 0; i < instr->operands.size(); i++) {
188 unsigned pred_idx = instr->opcode == aco_opcode::p_phi ?
189 block->logical_preds[i] :
190 block->linear_preds[i];
191 if (instr->operands[i].isTemp()) {
192 if (instr->operands[i].getTemp() == ctx.program->blocks[pred_idx].live_out_exec)
193 continue;
194 if (ctx.next_use_distances_end[pred_idx].find(instr->operands[i].getTemp()) == ctx.next_use_distances_end[pred_idx].end() ||
195 ctx.next_use_distances_end[pred_idx][instr->operands[i].getTemp()] != std::pair<uint32_t, uint32_t>{block_idx, 0})
196 worklist.insert(pred_idx);
197 ctx.next_use_distances_end[pred_idx][instr->operands[i].getTemp()] = {block_idx, 0};
198 }
199 }
200 next_uses.erase(instr->definitions[0].getTemp());
201 idx--;
202 }
203
204 /* all remaining live vars must be live-out at the predecessors */
205 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : next_uses) {
206 Temp temp = pair.first;
207 uint32_t distance = pair.second.second;
208 uint32_t dom = pair.second.first;
209 std::vector<unsigned>& preds = temp.is_linear() ? block->linear_preds : block->logical_preds;
210 for (unsigned pred_idx : preds) {
211 if (temp == ctx.program->blocks[pred_idx].live_out_exec)
212 continue;
213 if (ctx.program->blocks[pred_idx].loop_nest_depth > block->loop_nest_depth)
214 distance += 0xFFFF;
215 if (ctx.next_use_distances_end[pred_idx].find(temp) != ctx.next_use_distances_end[pred_idx].end()) {
216 dom = get_dominator(dom, ctx.next_use_distances_end[pred_idx][temp].first, ctx.program, temp.is_linear());
217 distance = std::min(ctx.next_use_distances_end[pred_idx][temp].second, distance);
218 }
219 if (ctx.next_use_distances_end[pred_idx][temp] != std::pair<uint32_t, uint32_t>{dom, distance})
220 worklist.insert(pred_idx);
221 ctx.next_use_distances_end[pred_idx][temp] = {dom, distance};
222 }
223 }
224
225 }
226
compute_global_next_uses(spill_ctx & ctx)227 void compute_global_next_uses(spill_ctx& ctx)
228 {
229 ctx.next_use_distances_start.resize(ctx.program->blocks.size());
230 ctx.next_use_distances_end.resize(ctx.program->blocks.size());
231 std::set<uint32_t> worklist;
232 for (Block& block : ctx.program->blocks)
233 worklist.insert(block.index);
234
235 while (!worklist.empty()) {
236 std::set<unsigned>::reverse_iterator b_it = worklist.rbegin();
237 unsigned block_idx = *b_it;
238 worklist.erase(block_idx);
239 next_uses_per_block(ctx, block_idx, worklist);
240 }
241 }
242
should_rematerialize(aco_ptr<Instruction> & instr)243 bool should_rematerialize(aco_ptr<Instruction>& instr)
244 {
245 /* TODO: rematerialization is only supported for VOP1, SOP1 and PSEUDO */
246 if (instr->format != Format::VOP1 && instr->format != Format::SOP1 && instr->format != Format::PSEUDO && instr->format != Format::SOPK)
247 return false;
248 /* TODO: pseudo-instruction rematerialization is only supported for p_create_vector/p_parallelcopy */
249 if (instr->format == Format::PSEUDO && instr->opcode != aco_opcode::p_create_vector &&
250 instr->opcode != aco_opcode::p_parallelcopy)
251 return false;
252 if (instr->format == Format::SOPK && instr->opcode != aco_opcode::s_movk_i32)
253 return false;
254
255 for (const Operand& op : instr->operands) {
256 /* TODO: rematerialization using temporaries isn't yet supported */
257 if (op.isTemp())
258 return false;
259 }
260
261 /* TODO: rematerialization with multiple definitions isn't yet supported */
262 if (instr->definitions.size() > 1)
263 return false;
264
265 return true;
266 }
267
do_reload(spill_ctx & ctx,Temp tmp,Temp new_name,uint32_t spill_id)268 aco_ptr<Instruction> do_reload(spill_ctx& ctx, Temp tmp, Temp new_name, uint32_t spill_id)
269 {
270 std::map<Temp, remat_info>::iterator remat = ctx.remat.find(tmp);
271 if (remat != ctx.remat.end()) {
272 Instruction *instr = remat->second.instr;
273 assert((instr->format == Format::VOP1 || instr->format == Format::SOP1 || instr->format == Format::PSEUDO || instr->format == Format::SOPK) && "unsupported");
274 assert((instr->format != Format::PSEUDO || instr->opcode == aco_opcode::p_create_vector || instr->opcode == aco_opcode::p_parallelcopy) && "unsupported");
275 assert(instr->definitions.size() == 1 && "unsupported");
276
277 aco_ptr<Instruction> res;
278 if (instr->format == Format::VOP1) {
279 res.reset(create_instruction<VOP1_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
280 } else if (instr->format == Format::SOP1) {
281 res.reset(create_instruction<SOP1_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
282 } else if (instr->format == Format::PSEUDO) {
283 res.reset(create_instruction<Pseudo_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
284 } else if (instr->format == Format::SOPK) {
285 res.reset(create_instruction<SOPK_instruction>(instr->opcode, instr->format, instr->operands.size(), instr->definitions.size()));
286 static_cast<SOPK_instruction*>(res.get())->imm = static_cast<SOPK_instruction*>(instr)->imm;
287 }
288 for (unsigned i = 0; i < instr->operands.size(); i++) {
289 res->operands[i] = instr->operands[i];
290 if (instr->operands[i].isTemp()) {
291 assert(false && "unsupported");
292 if (ctx.remat.count(instr->operands[i].getTemp()))
293 ctx.remat_used[ctx.remat[instr->operands[i].getTemp()].instr] = true;
294 }
295 }
296 res->definitions[0] = Definition(new_name);
297 return res;
298 } else {
299 aco_ptr<Pseudo_instruction> reload{create_instruction<Pseudo_instruction>(aco_opcode::p_reload, Format::PSEUDO, 1, 1)};
300 reload->operands[0] = Operand(spill_id);
301 reload->definitions[0] = Definition(new_name);
302 ctx.is_reloaded[spill_id] = true;
303 return reload;
304 }
305 }
306
get_rematerialize_info(spill_ctx & ctx)307 void get_rematerialize_info(spill_ctx& ctx)
308 {
309 for (Block& block : ctx.program->blocks) {
310 bool logical = false;
311 for (aco_ptr<Instruction>& instr : block.instructions) {
312 if (instr->opcode == aco_opcode::p_logical_start)
313 logical = true;
314 else if (instr->opcode == aco_opcode::p_logical_end)
315 logical = false;
316 if (logical && should_rematerialize(instr)) {
317 for (const Definition& def : instr->definitions) {
318 if (def.isTemp()) {
319 ctx.remat[def.getTemp()] = (remat_info){instr.get()};
320 ctx.remat_used[instr.get()] = false;
321 }
322 }
323 }
324 }
325 }
326 }
327
local_next_uses(spill_ctx & ctx,Block * block)328 std::vector<std::map<Temp, uint32_t>> local_next_uses(spill_ctx& ctx, Block* block)
329 {
330 std::vector<std::map<Temp, uint32_t>> local_next_uses(block->instructions.size());
331
332 std::map<Temp, uint32_t> next_uses;
333 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block->index])
334 next_uses[pair.first] = pair.second.second + block->instructions.size();
335
336 for (int idx = block->instructions.size() - 1; idx >= 0; idx--) {
337 aco_ptr<Instruction>& instr = block->instructions[idx];
338 if (!instr)
339 break;
340 if (instr->opcode == aco_opcode::p_phi || instr->opcode == aco_opcode::p_linear_phi)
341 break;
342
343 for (const Operand& op : instr->operands) {
344 if (op.isFixed() && op.physReg() == exec)
345 continue;
346 if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear())
347 continue;
348 if (op.isTemp())
349 next_uses[op.getTemp()] = idx;
350 }
351 for (const Definition& def : instr->definitions) {
352 if (def.isTemp())
353 next_uses.erase(def.getTemp());
354 }
355 local_next_uses[idx] = next_uses;
356 }
357 return local_next_uses;
358 }
359
360
init_live_in_vars(spill_ctx & ctx,Block * block,unsigned block_idx)361 RegisterDemand init_live_in_vars(spill_ctx& ctx, Block* block, unsigned block_idx)
362 {
363 RegisterDemand spilled_registers;
364
365 /* first block, nothing was spilled before */
366 if (block_idx == 0)
367 return {0, 0};
368
369 /* loop header block */
370 if (block->loop_nest_depth > ctx.program->blocks[block_idx - 1].loop_nest_depth) {
371 assert(block->linear_preds[0] == block_idx - 1);
372 assert(block->logical_preds[0] == block_idx - 1);
373
374 /* create new loop_info */
375 ctx.loop_header.emplace(block);
376
377 /* check how many live-through variables should be spilled */
378 RegisterDemand new_demand;
379 unsigned i = block_idx;
380 while (ctx.program->blocks[i].loop_nest_depth >= block->loop_nest_depth) {
381 assert(ctx.program->blocks.size() > i);
382 new_demand.update(ctx.program->blocks[i].register_demand);
383 i++;
384 }
385 unsigned loop_end = i;
386
387 /* keep live-through spilled */
388 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
389 if (pair.second.first < loop_end)
390 continue;
391
392 Temp to_spill = pair.first;
393 auto it = ctx.spills_exit[block_idx - 1].find(to_spill);
394 if (it == ctx.spills_exit[block_idx - 1].end())
395 continue;
396
397 ctx.spills_entry[block_idx][to_spill] = it->second;
398 spilled_registers += to_spill;
399 }
400
401 /* select live-through vgpr variables */
402 while (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) {
403 unsigned distance = 0;
404 Temp to_spill;
405 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
406 if (pair.first.type() == RegType::vgpr &&
407 pair.second.first >= loop_end &&
408 pair.second.second > distance &&
409 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
410 to_spill = pair.first;
411 distance = pair.second.second;
412 }
413 }
414 if (distance == 0)
415 break;
416
417 uint32_t spill_id;
418 if (ctx.spills_exit[block_idx - 1].find(to_spill) == ctx.spills_exit[block_idx - 1].end()) {
419 spill_id = ctx.allocate_spill_id(to_spill.regClass());
420 } else {
421 spill_id = ctx.spills_exit[block_idx - 1][to_spill];
422 }
423
424 ctx.spills_entry[block_idx][to_spill] = spill_id;
425 spilled_registers.vgpr += to_spill.size();
426 }
427
428 /* select live-through sgpr variables */
429 while (new_demand.sgpr - spilled_registers.sgpr > ctx.target_pressure.sgpr) {
430 unsigned distance = 0;
431 Temp to_spill;
432 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
433 if (pair.first.type() == RegType::sgpr &&
434 pair.second.first >= loop_end &&
435 pair.second.second > distance &&
436 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
437 to_spill = pair.first;
438 distance = pair.second.second;
439 }
440 }
441 if (distance == 0)
442 break;
443
444 uint32_t spill_id;
445 if (ctx.spills_exit[block_idx - 1].find(to_spill) == ctx.spills_exit[block_idx - 1].end()) {
446 spill_id = ctx.allocate_spill_id(to_spill.regClass());
447 } else {
448 spill_id = ctx.spills_exit[block_idx - 1][to_spill];
449 }
450
451 ctx.spills_entry[block_idx][to_spill] = spill_id;
452 spilled_registers.sgpr += to_spill.size();
453 }
454
455
456
457 /* shortcut */
458 if (!RegisterDemand(new_demand - spilled_registers).exceeds(ctx.target_pressure))
459 return spilled_registers;
460
461 /* if reg pressure is too high at beginning of loop, add variables with furthest use */
462 unsigned idx = 0;
463 while (block->instructions[idx]->opcode == aco_opcode::p_phi || block->instructions[idx]->opcode == aco_opcode::p_linear_phi)
464 idx++;
465
466 assert(idx != 0 && "loop without phis: TODO");
467 idx--;
468 RegisterDemand reg_pressure = ctx.register_demand[block_idx][idx] - spilled_registers;
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 = std::max<int16_t>(
473 reg_pressure.sgpr, ctx.register_demand[pred].back().sgpr - spilled_registers.sgpr);
474 }
475
476 while (reg_pressure.sgpr > ctx.target_pressure.sgpr) {
477 unsigned distance = 0;
478 Temp to_spill;
479 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
480 if (pair.first.type() == RegType::sgpr &&
481 pair.second.second > distance &&
482 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
483 to_spill = pair.first;
484 distance = pair.second.second;
485 }
486 }
487 assert(distance != 0);
488
489 ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
490 spilled_registers.sgpr += to_spill.size();
491 reg_pressure.sgpr -= to_spill.size();
492 }
493 while (reg_pressure.vgpr > ctx.target_pressure.vgpr) {
494 unsigned distance = 0;
495 Temp to_spill;
496 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
497 if (pair.first.type() == RegType::vgpr &&
498 pair.second.second > distance &&
499 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
500 to_spill = pair.first;
501 distance = pair.second.second;
502 }
503 }
504 assert(distance != 0);
505 ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
506 spilled_registers.vgpr += to_spill.size();
507 reg_pressure.vgpr -= to_spill.size();
508 }
509
510 return spilled_registers;
511 }
512
513 /* branch block */
514 if (block->linear_preds.size() == 1 && !(block->kind & block_kind_loop_exit)) {
515 /* keep variables spilled if they are alive and not used in the current block */
516 unsigned pred_idx = block->linear_preds[0];
517 for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
518 if (pair.first.type() == RegType::sgpr &&
519 ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
520 ctx.next_use_distances_start[block_idx][pair.first].first != block_idx) {
521 ctx.spills_entry[block_idx].insert(pair);
522 spilled_registers.sgpr += pair.first.size();
523 }
524 }
525 if (block->logical_preds.size() == 1) {
526 pred_idx = block->logical_preds[0];
527 for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
528 if (pair.first.type() == RegType::vgpr &&
529 ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
530 ctx.next_use_distances_start[block_idx][pair.first].first != block_idx) {
531 ctx.spills_entry[block_idx].insert(pair);
532 spilled_registers.vgpr += pair.first.size();
533 }
534 }
535 }
536
537 /* if register demand is still too high, we just keep all spilled live vars and process the block */
538 if (block->register_demand.sgpr - spilled_registers.sgpr > ctx.target_pressure.sgpr) {
539 pred_idx = block->linear_preds[0];
540 for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
541 if (pair.first.type() == RegType::sgpr &&
542 ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
543 ctx.spills_entry[block_idx].insert(pair).second) {
544 spilled_registers.sgpr += pair.first.size();
545 }
546 }
547 }
548 if (block->register_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr && block->logical_preds.size() == 1) {
549 pred_idx = block->logical_preds[0];
550 for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
551 if (pair.first.type() == RegType::vgpr &&
552 ctx.next_use_distances_start[block_idx].find(pair.first) != ctx.next_use_distances_start[block_idx].end() &&
553 ctx.spills_entry[block_idx].insert(pair).second) {
554 spilled_registers.vgpr += pair.first.size();
555 }
556 }
557 }
558
559 return spilled_registers;
560 }
561
562 /* else: merge block */
563 std::set<Temp> partial_spills;
564
565 /* keep variables spilled on all incoming paths */
566 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
567 std::vector<unsigned>& preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds;
568 /* If it can be rematerialized, keep the variable spilled if all predecessors do not reload it.
569 * Otherwise, if any predecessor reloads it, ensure it's reloaded on all other predecessors.
570 * The idea is that it's better in practice to rematerialize redundantly than to create lots of phis. */
571 /* TODO: test this idea with more than Dawn of War III shaders (the current pipeline-db doesn't seem to exercise this path much) */
572 bool remat = ctx.remat.count(pair.first);
573 bool spill = !remat;
574 uint32_t spill_id = 0;
575 for (unsigned pred_idx : preds) {
576 /* variable is not even live at the predecessor: probably from a phi */
577 if (ctx.next_use_distances_end[pred_idx].find(pair.first) == ctx.next_use_distances_end[pred_idx].end()) {
578 spill = false;
579 break;
580 }
581 if (ctx.spills_exit[pred_idx].find(pair.first) == ctx.spills_exit[pred_idx].end()) {
582 if (!remat)
583 spill = false;
584 } else {
585 partial_spills.insert(pair.first);
586 /* it might be that on one incoming path, the variable has a different spill_id, but add_couple_code() will take care of that. */
587 spill_id = ctx.spills_exit[pred_idx][pair.first];
588 if (remat)
589 spill = true;
590 }
591 }
592 if (spill) {
593 ctx.spills_entry[block_idx][pair.first] = spill_id;
594 partial_spills.erase(pair.first);
595 spilled_registers += pair.first;
596 }
597 }
598
599 /* same for phis */
600 unsigned idx = 0;
601 while (block->instructions[idx]->opcode == aco_opcode::p_linear_phi ||
602 block->instructions[idx]->opcode == aco_opcode::p_phi) {
603 aco_ptr<Instruction>& phi = block->instructions[idx];
604 std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
605 bool spill = true;
606
607 for (unsigned i = 0; i < phi->operands.size(); i++) {
608 if (phi->operands[i].isUndefined())
609 continue;
610 assert(phi->operands[i].isTemp());
611 if (ctx.spills_exit[preds[i]].find(phi->operands[i].getTemp()) == ctx.spills_exit[preds[i]].end())
612 spill = false;
613 else
614 partial_spills.insert(phi->definitions[0].getTemp());
615 }
616 if (spill) {
617 ctx.spills_entry[block_idx][phi->definitions[0].getTemp()] = ctx.allocate_spill_id(phi->definitions[0].regClass());
618 partial_spills.erase(phi->definitions[0].getTemp());
619 spilled_registers += phi->definitions[0].getTemp();
620 }
621
622 idx++;
623 }
624
625 /* if reg pressure at first instruction is still too high, add partially spilled variables */
626 RegisterDemand reg_pressure;
627 if (idx == 0) {
628 for (const Definition& def : block->instructions[idx]->definitions) {
629 if (def.isTemp()) {
630 reg_pressure -= def.getTemp();
631 }
632 }
633 for (const Operand& op : block->instructions[idx]->operands) {
634 if (op.isTemp() && op.isFirstKill()) {
635 reg_pressure += op.getTemp();
636 }
637 }
638 } else {
639 for (unsigned i = 0; i < idx; i++) {
640 aco_ptr<Instruction>& instr = block->instructions[i];
641 assert(is_phi(instr));
642 /* Killed phi definitions increase pressure in the predecessor but not
643 * the block they're in. Since the loops below are both to control
644 * pressure of the start of this block and the ends of it's
645 * predecessors, we need to count killed unspilled phi definitions here. */
646 if (instr->definitions[0].isKill() &&
647 !ctx.spills_entry[block_idx].count(instr->definitions[0].getTemp()))
648 reg_pressure += instr->definitions[0].getTemp();
649 }
650 idx--;
651 }
652 reg_pressure += ctx.register_demand[block_idx][idx] - spilled_registers;
653
654 /* Consider register pressure from linear predecessors. This can affect
655 * reg_pressure if the branch instructions define sgprs. */
656 for (unsigned pred : block->linear_preds) {
657 reg_pressure.sgpr = std::max<int16_t>(
658 reg_pressure.sgpr, ctx.register_demand[pred].back().sgpr - spilled_registers.sgpr);
659 }
660
661 while (reg_pressure.sgpr > ctx.target_pressure.sgpr) {
662 assert(!partial_spills.empty());
663
664 std::set<Temp>::iterator it = partial_spills.begin();
665 Temp to_spill = Temp();
666 unsigned distance = 0;
667 while (it != partial_spills.end()) {
668 assert(ctx.spills_entry[block_idx].find(*it) == ctx.spills_entry[block_idx].end());
669
670 if (it->type() == RegType::sgpr && ctx.next_use_distances_start[block_idx][*it].second > distance) {
671 distance = ctx.next_use_distances_start[block_idx][*it].second;
672 to_spill = *it;
673 }
674 ++it;
675 }
676 assert(distance != 0);
677
678 ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
679 partial_spills.erase(to_spill);
680 spilled_registers.sgpr += to_spill.size();
681 reg_pressure.sgpr -= to_spill.size();
682 }
683
684 while (reg_pressure.vgpr > ctx.target_pressure.vgpr) {
685 assert(!partial_spills.empty());
686
687 std::set<Temp>::iterator it = partial_spills.begin();
688 Temp to_spill = Temp();
689 unsigned distance = 0;
690 while (it != partial_spills.end()) {
691 assert(ctx.spills_entry[block_idx].find(*it) == ctx.spills_entry[block_idx].end());
692
693 if (it->type() == RegType::vgpr && ctx.next_use_distances_start[block_idx][*it].second > distance) {
694 distance = ctx.next_use_distances_start[block_idx][*it].second;
695 to_spill = *it;
696 }
697 ++it;
698 }
699 assert(distance != 0);
700
701 ctx.spills_entry[block_idx][to_spill] = ctx.allocate_spill_id(to_spill.regClass());
702 partial_spills.erase(to_spill);
703 spilled_registers.vgpr += to_spill.size();
704 reg_pressure.vgpr -= to_spill.size();
705 }
706
707 return spilled_registers;
708 }
709
710
get_demand_before(spill_ctx & ctx,unsigned block_idx,unsigned idx)711 RegisterDemand get_demand_before(spill_ctx& ctx, unsigned block_idx, unsigned idx)
712 {
713 if (idx == 0) {
714 RegisterDemand demand = ctx.register_demand[block_idx][idx];
715 aco_ptr<Instruction>& instr = ctx.program->blocks[block_idx].instructions[idx];
716 aco_ptr<Instruction> instr_before(nullptr);
717 return get_demand_before(demand, instr, instr_before);
718 } else {
719 return ctx.register_demand[block_idx][idx - 1];
720 }
721 }
722
add_coupling_code(spill_ctx & ctx,Block * block,unsigned block_idx)723 void add_coupling_code(spill_ctx& ctx, Block* block, unsigned block_idx)
724 {
725 /* no coupling code necessary */
726 if (block->linear_preds.size() == 0)
727 return;
728
729 std::vector<aco_ptr<Instruction>> instructions;
730 /* branch block: TODO take other branch into consideration */
731 if (block->linear_preds.size() == 1 && !(block->kind & (block_kind_loop_exit | block_kind_loop_header))) {
732 assert(ctx.processed[block->linear_preds[0]]);
733 assert(ctx.register_demand[block_idx].size() == block->instructions.size());
734 std::vector<RegisterDemand> reg_demand;
735 unsigned insert_idx = 0;
736 unsigned pred_idx = block->linear_preds[0];
737 RegisterDemand demand_before = get_demand_before(ctx, block_idx, 0);
738
739 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> live : ctx.next_use_distances_start[block_idx]) {
740 if (!live.first.is_linear())
741 continue;
742 /* still spilled */
743 if (ctx.spills_entry[block_idx].find(live.first) != ctx.spills_entry[block_idx].end())
744 continue;
745
746 /* in register at end of predecessor */
747 if (ctx.spills_exit[pred_idx].find(live.first) == ctx.spills_exit[pred_idx].end()) {
748 std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(live.first);
749 if (it != ctx.renames[pred_idx].end())
750 ctx.renames[block_idx].insert(*it);
751 continue;
752 }
753
754 /* variable is spilled at predecessor and live at current block: create reload instruction */
755 Temp new_name = ctx.program->allocateTmp(live.first.regClass());
756 aco_ptr<Instruction> reload = do_reload(ctx, live.first, new_name, ctx.spills_exit[pred_idx][live.first]);
757 instructions.emplace_back(std::move(reload));
758 reg_demand.push_back(demand_before);
759 ctx.renames[block_idx][live.first] = new_name;
760 }
761
762 if (block->logical_preds.size() == 1) {
763 do {
764 assert(insert_idx < block->instructions.size());
765 instructions.emplace_back(std::move(block->instructions[insert_idx]));
766 reg_demand.push_back(ctx.register_demand[block_idx][insert_idx]);
767 insert_idx++;
768 } while (instructions.back()->opcode != aco_opcode::p_logical_start);
769
770 unsigned pred_idx = block->logical_preds[0];
771 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> live : ctx.next_use_distances_start[block_idx]) {
772 if (live.first.is_linear())
773 continue;
774 /* still spilled */
775 if (ctx.spills_entry[block_idx].find(live.first) != ctx.spills_entry[block_idx].end())
776 continue;
777
778 /* in register at end of predecessor */
779 if (ctx.spills_exit[pred_idx].find(live.first) == ctx.spills_exit[pred_idx].end()) {
780 std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(live.first);
781 if (it != ctx.renames[pred_idx].end())
782 ctx.renames[block_idx].insert(*it);
783 continue;
784 }
785
786 /* variable is spilled at predecessor and live at current block: create reload instruction */
787 Temp new_name = ctx.program->allocateTmp(live.first.regClass());
788 aco_ptr<Instruction> reload = do_reload(ctx, live.first, new_name, ctx.spills_exit[pred_idx][live.first]);
789 instructions.emplace_back(std::move(reload));
790 reg_demand.emplace_back(reg_demand.back());
791 ctx.renames[block_idx][live.first] = new_name;
792 }
793 }
794
795 /* combine new reload instructions with original block */
796 if (!instructions.empty()) {
797 reg_demand.insert(reg_demand.end(), std::next(ctx.register_demand[block->index].begin(), insert_idx),
798 ctx.register_demand[block->index].end());
799 ctx.register_demand[block_idx] = std::move(reg_demand);
800 instructions.insert(instructions.end(),
801 std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(std::next(block->instructions.begin(), insert_idx)),
802 std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
803 block->instructions = std::move(instructions);
804 }
805 return;
806 }
807
808 /* loop header and merge blocks: check if all (linear) predecessors have been processed */
809 for (ASSERTED unsigned pred : block->linear_preds)
810 assert(ctx.processed[pred]);
811
812 /* iterate the phi nodes for which operands to spill at the predecessor */
813 for (aco_ptr<Instruction>& phi : block->instructions) {
814 if (phi->opcode != aco_opcode::p_phi &&
815 phi->opcode != aco_opcode::p_linear_phi)
816 break;
817
818 /* if the phi is not spilled, add to instructions */
819 if (ctx.spills_entry[block_idx].find(phi->definitions[0].getTemp()) == ctx.spills_entry[block_idx].end()) {
820 instructions.emplace_back(std::move(phi));
821 continue;
822 }
823
824 std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
825 uint32_t def_spill_id = ctx.spills_entry[block_idx][phi->definitions[0].getTemp()];
826
827 for (unsigned i = 0; i < phi->operands.size(); i++) {
828 if (phi->operands[i].isUndefined())
829 continue;
830
831 unsigned pred_idx = preds[i];
832 assert(phi->operands[i].isTemp() && phi->operands[i].isKill());
833 Temp var = phi->operands[i].getTemp();
834
835 std::map<Temp, Temp>::iterator rename_it = ctx.renames[pred_idx].find(var);
836 /* prevent the definining instruction from being DCE'd if it could be rematerialized */
837 if (rename_it == ctx.renames[preds[i]].end() && ctx.remat.count(var))
838 ctx.remat_used[ctx.remat[var].instr] = true;
839
840 /* build interferences between the phi def and all spilled variables at the predecessor blocks */
841 for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[pred_idx]) {
842 if (var == pair.first)
843 continue;
844 ctx.add_interference(def_spill_id, pair.second);
845 }
846
847 /* check if variable is already spilled at predecessor */
848 std::map<Temp, uint32_t>::iterator spilled = ctx.spills_exit[pred_idx].find(var);
849 if (spilled != ctx.spills_exit[pred_idx].end()) {
850 if (spilled->second != def_spill_id)
851 ctx.add_affinity(def_spill_id, spilled->second);
852 continue;
853 }
854
855 /* rename if necessary */
856 if (rename_it != ctx.renames[pred_idx].end()) {
857 var = rename_it->second;
858 ctx.renames[pred_idx].erase(rename_it);
859 }
860
861 uint32_t spill_id = ctx.allocate_spill_id(phi->definitions[0].regClass());
862 ctx.add_affinity(def_spill_id, spill_id);
863 aco_ptr<Pseudo_instruction> spill{create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
864 spill->operands[0] = Operand(var);
865 spill->operands[1] = Operand(spill_id);
866 Block& pred = ctx.program->blocks[pred_idx];
867 unsigned idx = pred.instructions.size();
868 do {
869 assert(idx != 0);
870 idx--;
871 } while (phi->opcode == aco_opcode::p_phi && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
872 std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
873 pred.instructions.insert(it, std::move(spill));
874 ctx.spills_exit[pred_idx][phi->operands[i].getTemp()] = spill_id;
875 }
876
877 /* remove phi from instructions */
878 phi.reset();
879 }
880
881 /* iterate all (other) spilled variables for which to spill at the predecessor */
882 // TODO: would be better to have them sorted: first vgprs and first with longest distance
883 for (std::pair<Temp, uint32_t> pair : ctx.spills_entry[block_idx]) {
884 std::vector<unsigned> preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds;
885
886 for (unsigned pred_idx : preds) {
887 /* variable is already spilled at predecessor */
888 std::map<Temp, uint32_t>::iterator spilled = ctx.spills_exit[pred_idx].find(pair.first);
889 if (spilled != ctx.spills_exit[pred_idx].end()) {
890 if (spilled->second != pair.second)
891 ctx.add_affinity(pair.second, spilled->second);
892 continue;
893 }
894
895 /* variable is dead at predecessor, it must be from a phi: this works because of CSSA form */
896 if (ctx.next_use_distances_end[pred_idx].find(pair.first) == ctx.next_use_distances_end[pred_idx].end())
897 continue;
898
899 /* add interferences between spilled variable and predecessors exit spills */
900 for (std::pair<Temp, uint32_t> exit_spill : ctx.spills_exit[pred_idx]) {
901 if (exit_spill.first == pair.first)
902 continue;
903 ctx.add_interference(exit_spill.second, pair.second);
904 }
905
906 /* variable is in register at predecessor and has to be spilled */
907 /* rename if necessary */
908 Temp var = pair.first;
909 std::map<Temp, Temp>::iterator rename_it = ctx.renames[pred_idx].find(var);
910 if (rename_it != ctx.renames[pred_idx].end()) {
911 var = rename_it->second;
912 ctx.renames[pred_idx].erase(rename_it);
913 }
914
915 aco_ptr<Pseudo_instruction> spill{create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
916 spill->operands[0] = Operand(var);
917 spill->operands[1] = Operand(pair.second);
918 Block& pred = ctx.program->blocks[pred_idx];
919 unsigned idx = pred.instructions.size();
920 do {
921 assert(idx != 0);
922 idx--;
923 } while (pair.first.type() == RegType::vgpr && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
924 std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
925 pred.instructions.insert(it, std::move(spill));
926 ctx.spills_exit[pred.index][pair.first] = pair.second;
927 }
928 }
929
930 /* iterate phis for which operands to reload */
931 for (aco_ptr<Instruction>& phi : instructions) {
932 assert(phi->opcode == aco_opcode::p_phi || phi->opcode == aco_opcode::p_linear_phi);
933 assert(ctx.spills_entry[block_idx].find(phi->definitions[0].getTemp()) == ctx.spills_entry[block_idx].end());
934
935 std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block->logical_preds : block->linear_preds;
936 for (unsigned i = 0; i < phi->operands.size(); i++) {
937 if (!phi->operands[i].isTemp())
938 continue;
939 unsigned pred_idx = preds[i];
940
941 /* rename operand */
942 if (ctx.spills_exit[pred_idx].find(phi->operands[i].getTemp()) == ctx.spills_exit[pred_idx].end()) {
943 std::map<Temp, Temp>::iterator it = ctx.renames[pred_idx].find(phi->operands[i].getTemp());
944 if (it != ctx.renames[pred_idx].end())
945 phi->operands[i].setTemp(it->second);
946 /* prevent the definining instruction from being DCE'd if it could be rematerialized */
947 else if (ctx.remat.count(phi->operands[i].getTemp()))
948 ctx.remat_used[ctx.remat[phi->operands[i].getTemp()].instr] = true;
949 continue;
950 }
951
952 Temp tmp = phi->operands[i].getTemp();
953
954 /* reload phi operand at end of predecessor block */
955 Temp new_name = ctx.program->allocateTmp(tmp.regClass());
956 Block& pred = ctx.program->blocks[pred_idx];
957 unsigned idx = pred.instructions.size();
958 do {
959 assert(idx != 0);
960 idx--;
961 } while (phi->opcode == aco_opcode::p_phi && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
962 std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
963
964 aco_ptr<Instruction> reload = do_reload(ctx, tmp, new_name, ctx.spills_exit[pred_idx][tmp]);
965 pred.instructions.insert(it, std::move(reload));
966
967 ctx.spills_exit[pred_idx].erase(tmp);
968 ctx.renames[pred_idx][tmp] = new_name;
969 phi->operands[i].setTemp(new_name);
970 }
971 }
972
973 /* iterate live variables for which to reload */
974 // TODO: reload at current block if variable is spilled on all predecessors
975 for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_start[block_idx]) {
976 /* skip spilled variables */
977 if (ctx.spills_entry[block_idx].find(pair.first) != ctx.spills_entry[block_idx].end())
978 continue;
979 std::vector<unsigned> preds = pair.first.is_linear() ? block->linear_preds : block->logical_preds;
980
981 /* variable is dead at predecessor, it must be from a phi */
982 bool is_dead = false;
983 for (unsigned pred_idx : preds) {
984 if (ctx.next_use_distances_end[pred_idx].find(pair.first) == ctx.next_use_distances_end[pred_idx].end())
985 is_dead = true;
986 }
987 if (is_dead)
988 continue;
989 for (unsigned pred_idx : preds) {
990 /* the variable is not spilled at the predecessor */
991 if (ctx.spills_exit[pred_idx].find(pair.first) == ctx.spills_exit[pred_idx].end())
992 continue;
993
994 /* variable is spilled at predecessor and has to be reloaded */
995 Temp new_name = ctx.program->allocateTmp(pair.first.regClass());
996 Block& pred = ctx.program->blocks[pred_idx];
997 unsigned idx = pred.instructions.size();
998 do {
999 assert(idx != 0);
1000 idx--;
1001 } while (pair.first.type() == RegType::vgpr && pred.instructions[idx]->opcode != aco_opcode::p_logical_end);
1002 std::vector<aco_ptr<Instruction>>::iterator it = std::next(pred.instructions.begin(), idx);
1003
1004 aco_ptr<Instruction> reload = do_reload(ctx, pair.first, new_name, ctx.spills_exit[pred.index][pair.first]);
1005 pred.instructions.insert(it, std::move(reload));
1006
1007 ctx.spills_exit[pred.index].erase(pair.first);
1008 ctx.renames[pred.index][pair.first] = new_name;
1009 }
1010
1011 /* check if we have to create a new phi for this variable */
1012 Temp rename = Temp();
1013 bool is_same = true;
1014 for (unsigned pred_idx : preds) {
1015 if (ctx.renames[pred_idx].find(pair.first) == ctx.renames[pred_idx].end()) {
1016 if (rename == Temp())
1017 rename = pair.first;
1018 else
1019 is_same = rename == pair.first;
1020 } else {
1021 if (rename == Temp())
1022 rename = ctx.renames[pred_idx][pair.first];
1023 else
1024 is_same = rename == ctx.renames[pred_idx][pair.first];
1025 }
1026
1027 if (!is_same)
1028 break;
1029 }
1030
1031 if (!is_same) {
1032 /* the variable was renamed differently in the predecessors: we have to create a phi */
1033 aco_opcode opcode = pair.first.is_linear() ? aco_opcode::p_linear_phi : aco_opcode::p_phi;
1034 aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(opcode, Format::PSEUDO, preds.size(), 1)};
1035 rename = ctx.program->allocateTmp(pair.first.regClass());
1036 for (unsigned i = 0; i < phi->operands.size(); i++) {
1037 Temp tmp;
1038 if (ctx.renames[preds[i]].find(pair.first) != ctx.renames[preds[i]].end()) {
1039 tmp = ctx.renames[preds[i]][pair.first];
1040 } else if (preds[i] >= block_idx) {
1041 tmp = rename;
1042 } else {
1043 tmp = pair.first;
1044 /* prevent the definining instruction from being DCE'd if it could be rematerialized */
1045 if (ctx.remat.count(tmp))
1046 ctx.remat_used[ctx.remat[tmp].instr] = true;
1047 }
1048 phi->operands[i] = Operand(tmp);
1049 }
1050 phi->definitions[0] = Definition(rename);
1051 instructions.emplace_back(std::move(phi));
1052 }
1053
1054 /* the variable was renamed: add new name to renames */
1055 if (!(rename == Temp() || rename == pair.first))
1056 ctx.renames[block_idx][pair.first] = rename;
1057 }
1058
1059 /* combine phis with instructions */
1060 unsigned idx = 0;
1061 while (!block->instructions[idx]) {
1062 idx++;
1063 }
1064
1065 if (!ctx.processed[block_idx]) {
1066 assert(!(block->kind & block_kind_loop_header));
1067 RegisterDemand demand_before = get_demand_before(ctx, block_idx, idx);
1068 ctx.register_demand[block->index].erase(ctx.register_demand[block->index].begin(), ctx.register_demand[block->index].begin() + idx);
1069 ctx.register_demand[block->index].insert(ctx.register_demand[block->index].begin(), instructions.size(), demand_before);
1070 }
1071
1072 std::vector<aco_ptr<Instruction>>::iterator start = std::next(block->instructions.begin(), idx);
1073 instructions.insert(instructions.end(), std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(start),
1074 std::move_iterator<std::vector<aco_ptr<Instruction>>::iterator>(block->instructions.end()));
1075 block->instructions = std::move(instructions);
1076 }
1077
process_block(spill_ctx & ctx,unsigned block_idx,Block * block,std::map<Temp,uint32_t> & current_spills,RegisterDemand spilled_registers)1078 void process_block(spill_ctx& ctx, unsigned block_idx, Block* block,
1079 std::map<Temp, uint32_t> ¤t_spills, RegisterDemand spilled_registers)
1080 {
1081 assert(!ctx.processed[block_idx]);
1082
1083 std::vector<std::map<Temp, uint32_t>> local_next_use_distance;
1084 std::vector<aco_ptr<Instruction>> instructions;
1085 unsigned idx = 0;
1086
1087 /* phis are handled separetely */
1088 while (block->instructions[idx]->opcode == aco_opcode::p_phi ||
1089 block->instructions[idx]->opcode == aco_opcode::p_linear_phi) {
1090 instructions.emplace_back(std::move(block->instructions[idx++]));
1091 }
1092
1093 if (block->register_demand.exceeds(ctx.target_pressure))
1094 local_next_use_distance = local_next_uses(ctx, block);
1095
1096 while (idx < block->instructions.size()) {
1097 aco_ptr<Instruction>& instr = block->instructions[idx];
1098
1099 std::map<Temp, std::pair<Temp, uint32_t>> reloads;
1100 std::map<Temp, uint32_t> spills;
1101 /* rename and reload operands */
1102 for (Operand& op : instr->operands) {
1103 if (!op.isTemp())
1104 continue;
1105 if (current_spills.find(op.getTemp()) == current_spills.end()) {
1106 /* the Operand is in register: check if it was renamed */
1107 if (ctx.renames[block_idx].find(op.getTemp()) != ctx.renames[block_idx].end())
1108 op.setTemp(ctx.renames[block_idx][op.getTemp()]);
1109 /* prevent it's definining instruction from being DCE'd if it could be rematerialized */
1110 else if (ctx.remat.count(op.getTemp()))
1111 ctx.remat_used[ctx.remat[op.getTemp()].instr] = true;
1112 continue;
1113 }
1114 /* the Operand is spilled: add it to reloads */
1115 Temp new_tmp = ctx.program->allocateTmp(op.regClass());
1116 ctx.renames[block_idx][op.getTemp()] = new_tmp;
1117 reloads[new_tmp] = std::make_pair(op.getTemp(), current_spills[op.getTemp()]);
1118 current_spills.erase(op.getTemp());
1119 op.setTemp(new_tmp);
1120 spilled_registers -= new_tmp;
1121 }
1122
1123 /* check if register demand is low enough before and after the current instruction */
1124 if (block->register_demand.exceeds(ctx.target_pressure)) {
1125
1126 RegisterDemand new_demand = ctx.register_demand[block_idx][idx];
1127 new_demand.update(get_demand_before(ctx, block_idx, idx));
1128
1129 assert(!local_next_use_distance.empty());
1130
1131 /* if reg pressure is too high, spill variable with furthest next use */
1132 while (RegisterDemand(new_demand - spilled_registers).exceeds(ctx.target_pressure)) {
1133 unsigned distance = 0;
1134 Temp to_spill;
1135 bool do_rematerialize = false;
1136 if (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) {
1137 for (std::pair<Temp, uint32_t> pair : local_next_use_distance[idx]) {
1138 bool can_rematerialize = ctx.remat.count(pair.first);
1139 if (pair.first.type() == RegType::vgpr &&
1140 ((pair.second > distance && can_rematerialize == do_rematerialize) ||
1141 (can_rematerialize && !do_rematerialize && pair.second > idx)) &&
1142 current_spills.find(pair.first) == current_spills.end() &&
1143 ctx.spills_exit[block_idx].find(pair.first) == ctx.spills_exit[block_idx].end()) {
1144 to_spill = pair.first;
1145 distance = pair.second;
1146 do_rematerialize = can_rematerialize;
1147 }
1148 }
1149 } else {
1150 for (std::pair<Temp, uint32_t> pair : local_next_use_distance[idx]) {
1151 bool can_rematerialize = ctx.remat.count(pair.first);
1152 if (pair.first.type() == RegType::sgpr &&
1153 ((pair.second > distance && can_rematerialize == do_rematerialize) ||
1154 (can_rematerialize && !do_rematerialize && pair.second > idx)) &&
1155 current_spills.find(pair.first) == current_spills.end() &&
1156 ctx.spills_exit[block_idx].find(pair.first) == ctx.spills_exit[block_idx].end()) {
1157 to_spill = pair.first;
1158 distance = pair.second;
1159 do_rematerialize = can_rematerialize;
1160 }
1161 }
1162 }
1163
1164 assert(distance != 0 && distance > idx);
1165 uint32_t spill_id = ctx.allocate_spill_id(to_spill.regClass());
1166
1167 /* add interferences with currently spilled variables */
1168 for (std::pair<Temp, uint32_t> pair : current_spills)
1169 ctx.add_interference(spill_id, pair.second);
1170 for (std::pair<Temp, std::pair<Temp, uint32_t>> pair : reloads)
1171 ctx.add_interference(spill_id, pair.second.second);
1172
1173 current_spills[to_spill] = spill_id;
1174 spilled_registers += to_spill;
1175
1176 /* rename if necessary */
1177 if (ctx.renames[block_idx].find(to_spill) != ctx.renames[block_idx].end()) {
1178 to_spill = ctx.renames[block_idx][to_spill];
1179 }
1180
1181 /* add spill to new instructions */
1182 aco_ptr<Pseudo_instruction> spill{create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 2, 0)};
1183 spill->operands[0] = Operand(to_spill);
1184 spill->operands[1] = Operand(spill_id);
1185 instructions.emplace_back(std::move(spill));
1186 }
1187 }
1188
1189 /* add reloads and instruction to new instructions */
1190 for (std::pair<Temp, std::pair<Temp, uint32_t>> pair : reloads) {
1191 aco_ptr<Instruction> reload = do_reload(ctx, pair.second.first, pair.first, pair.second.second);
1192 instructions.emplace_back(std::move(reload));
1193 }
1194 instructions.emplace_back(std::move(instr));
1195 idx++;
1196 }
1197
1198 block->instructions = std::move(instructions);
1199 ctx.spills_exit[block_idx].insert(current_spills.begin(), current_spills.end());
1200 }
1201
spill_block(spill_ctx & ctx,unsigned block_idx)1202 void spill_block(spill_ctx& ctx, unsigned block_idx)
1203 {
1204 Block* block = &ctx.program->blocks[block_idx];
1205
1206 /* determine set of variables which are spilled at the beginning of the block */
1207 RegisterDemand spilled_registers = init_live_in_vars(ctx, block, block_idx);
1208
1209 /* add interferences for spilled variables */
1210 for (auto it = ctx.spills_entry[block_idx].begin(); it != ctx.spills_entry[block_idx].end(); ++it) {
1211 for (auto it2 = std::next(it); it2 != ctx.spills_entry[block_idx].end(); ++it2)
1212 ctx.add_interference(it->second, it2->second);
1213 }
1214
1215 bool is_loop_header = block->loop_nest_depth && ctx.loop_header.top()->index == block_idx;
1216 if (!is_loop_header) {
1217 /* add spill/reload code on incoming control flow edges */
1218 add_coupling_code(ctx, block, block_idx);
1219 }
1220
1221 std::map<Temp, uint32_t> current_spills = ctx.spills_entry[block_idx];
1222
1223 /* check conditions to process this block */
1224 bool process = RegisterDemand(block->register_demand - spilled_registers).exceeds(ctx.target_pressure) ||
1225 !ctx.renames[block_idx].empty() ||
1226 ctx.remat_used.size();
1227
1228 std::map<Temp, uint32_t>::iterator it = current_spills.begin();
1229 while (!process && it != current_spills.end()) {
1230 if (ctx.next_use_distances_start[block_idx][it->first].first == block_idx)
1231 process = true;
1232 ++it;
1233 }
1234
1235 if (process)
1236 process_block(ctx, block_idx, block, current_spills, spilled_registers);
1237 else
1238 ctx.spills_exit[block_idx].insert(current_spills.begin(), current_spills.end());
1239
1240 ctx.processed[block_idx] = true;
1241
1242 /* check if the next block leaves the current loop */
1243 if (block->loop_nest_depth == 0 || ctx.program->blocks[block_idx + 1].loop_nest_depth >= block->loop_nest_depth)
1244 return;
1245
1246 Block* loop_header = ctx.loop_header.top();
1247
1248 /* preserve original renames at end of loop header block */
1249 std::map<Temp, Temp> renames = std::move(ctx.renames[loop_header->index]);
1250
1251 /* add coupling code to all loop header predecessors */
1252 add_coupling_code(ctx, loop_header, loop_header->index);
1253
1254 /* propagate new renames through loop: i.e. repair the SSA */
1255 renames.swap(ctx.renames[loop_header->index]);
1256 for (std::pair<Temp, Temp> rename : renames) {
1257 for (unsigned idx = loop_header->index; idx <= block_idx; idx++) {
1258 Block& current = ctx.program->blocks[idx];
1259 std::vector<aco_ptr<Instruction>>::iterator instr_it = current.instructions.begin();
1260
1261 /* first rename phis */
1262 while (instr_it != current.instructions.end()) {
1263 aco_ptr<Instruction>& phi = *instr_it;
1264 if (phi->opcode != aco_opcode::p_phi && phi->opcode != aco_opcode::p_linear_phi)
1265 break;
1266 /* no need to rename the loop header phis once again. this happened in add_coupling_code() */
1267 if (idx == loop_header->index) {
1268 instr_it++;
1269 continue;
1270 }
1271
1272 for (Operand& op : phi->operands) {
1273 if (!op.isTemp())
1274 continue;
1275 if (op.getTemp() == rename.first)
1276 op.setTemp(rename.second);
1277 }
1278 instr_it++;
1279 }
1280
1281 std::map<Temp, std::pair<uint32_t, uint32_t>>::iterator it = ctx.next_use_distances_start[idx].find(rename.first);
1282
1283 /* variable is not live at beginning of this block */
1284 if (it == ctx.next_use_distances_start[idx].end())
1285 continue;
1286
1287 /* if the variable is live at the block's exit, add rename */
1288 if (ctx.next_use_distances_end[idx].find(rename.first) != ctx.next_use_distances_end[idx].end())
1289 ctx.renames[idx].insert(rename);
1290
1291 /* rename all uses in this block */
1292 bool renamed = false;
1293 while (!renamed && instr_it != current.instructions.end()) {
1294 aco_ptr<Instruction>& instr = *instr_it;
1295 for (Operand& op : instr->operands) {
1296 if (!op.isTemp())
1297 continue;
1298 if (op.getTemp() == rename.first) {
1299 op.setTemp(rename.second);
1300 /* we can stop with this block as soon as the variable is spilled */
1301 if (instr->opcode == aco_opcode::p_spill)
1302 renamed = true;
1303 }
1304 }
1305 instr_it++;
1306 }
1307 }
1308 }
1309
1310 /* remove loop header info from stack */
1311 ctx.loop_header.pop();
1312 }
1313
load_scratch_resource(spill_ctx & ctx,Temp & scratch_offset,std::vector<aco_ptr<Instruction>> & instructions,unsigned offset,bool is_top_level)1314 Temp load_scratch_resource(spill_ctx& ctx, Temp& scratch_offset,
1315 std::vector<aco_ptr<Instruction>>& instructions,
1316 unsigned offset, bool is_top_level)
1317 {
1318 Builder bld(ctx.program);
1319 if (is_top_level) {
1320 bld.reset(&instructions);
1321 } else {
1322 /* find p_logical_end */
1323 unsigned idx = instructions.size() - 1;
1324 while (instructions[idx]->opcode != aco_opcode::p_logical_end)
1325 idx--;
1326 bld.reset(&instructions, std::next(instructions.begin(), idx));
1327 }
1328
1329 Temp private_segment_buffer = ctx.program->private_segment_buffer;
1330 if (ctx.program->stage != compute_cs)
1331 private_segment_buffer = bld.smem(aco_opcode::s_load_dwordx2, bld.def(s2), private_segment_buffer, Operand(0u));
1332
1333 if (offset)
1334 scratch_offset = bld.sop2(aco_opcode::s_add_u32, bld.def(s1), bld.def(s1, scc), scratch_offset, Operand(offset));
1335
1336 uint32_t rsrc_conf = S_008F0C_ADD_TID_ENABLE(1) |
1337 S_008F0C_INDEX_STRIDE(ctx.program->wave_size == 64 ? 3 : 2);
1338
1339 if (ctx.program->chip_class >= GFX10) {
1340 rsrc_conf |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
1341 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
1342 S_008F0C_RESOURCE_LEVEL(1);
1343 } else if (ctx.program->chip_class <= GFX7) { /* dfmt modifies stride on GFX8/GFX9 when ADD_TID_EN=1 */
1344 rsrc_conf |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
1345 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
1346 }
1347 /* older generations need element size = 4 bytes. element size removed in GFX9 */
1348 if (ctx.program->chip_class <= GFX8)
1349 rsrc_conf |= S_008F0C_ELEMENT_SIZE(1);
1350
1351 return bld.pseudo(aco_opcode::p_create_vector, bld.def(s4),
1352 private_segment_buffer, Operand(-1u),
1353 Operand(rsrc_conf));
1354 }
1355
add_interferences(spill_ctx & ctx,std::vector<bool> & is_assigned,std::vector<uint32_t> & slots,std::vector<bool> & slots_used,unsigned id)1356 void add_interferences(spill_ctx& ctx, std::vector<bool>& is_assigned,
1357 std::vector<uint32_t>& slots, std::vector<bool>& slots_used,
1358 unsigned id)
1359 {
1360 for (unsigned other : ctx.interferences[id].second) {
1361 if (!is_assigned[other])
1362 continue;
1363
1364 RegClass other_rc = ctx.interferences[other].first;
1365 unsigned slot = slots[other];
1366 std::fill(slots_used.begin() + slot, slots_used.begin() + slot + other_rc.size(), true);
1367 }
1368 }
1369
find_available_slot(std::vector<bool> & used,unsigned wave_size,unsigned size,bool is_sgpr,unsigned * num_slots)1370 unsigned find_available_slot(std::vector<bool>& used, unsigned wave_size,
1371 unsigned size, bool is_sgpr, unsigned *num_slots)
1372 {
1373 unsigned wave_size_minus_one = wave_size - 1;
1374 unsigned slot = 0;
1375
1376 while (true) {
1377 bool available = true;
1378 for (unsigned i = 0; i < size; i++) {
1379 if (slot + i < used.size() && used[slot + i]) {
1380 available = false;
1381 break;
1382 }
1383 }
1384 if (!available) {
1385 slot++;
1386 continue;
1387 }
1388
1389 if (is_sgpr && ((slot & wave_size_minus_one) > wave_size - size)) {
1390 slot = align(slot, wave_size);
1391 continue;
1392 }
1393
1394 std::fill(used.begin(), used.end(), false);
1395
1396 if (slot + size > used.size())
1397 used.resize(slot + size);
1398
1399 return slot;
1400 }
1401 }
1402
assign_spill_slots_helper(spill_ctx & ctx,RegType type,std::vector<bool> & is_assigned,std::vector<uint32_t> & slots,unsigned * num_slots)1403 void assign_spill_slots_helper(spill_ctx& ctx, RegType type,
1404 std::vector<bool>& is_assigned,
1405 std::vector<uint32_t>& slots,
1406 unsigned *num_slots)
1407 {
1408 std::vector<bool> slots_used(*num_slots);
1409
1410 /* assign slots for ids with affinities first */
1411 for (std::vector<uint32_t>& vec : ctx.affinities) {
1412 if (ctx.interferences[vec[0]].first.type() != type)
1413 continue;
1414
1415 for (unsigned id : vec) {
1416 if (!ctx.is_reloaded[id])
1417 continue;
1418
1419 add_interferences(ctx, is_assigned, slots, slots_used, id);
1420 }
1421
1422 unsigned slot = find_available_slot(slots_used, ctx.wave_size,
1423 ctx.interferences[vec[0]].first.size(),
1424 type == RegType::sgpr, num_slots);
1425
1426 for (unsigned id : vec) {
1427 assert(!is_assigned[id]);
1428
1429 if (ctx.is_reloaded[id]) {
1430 slots[id] = slot;
1431 is_assigned[id] = true;
1432 }
1433 }
1434 }
1435
1436 /* assign slots for ids without affinities */
1437 for (unsigned id = 0; id < ctx.interferences.size(); id++) {
1438 if (is_assigned[id] || !ctx.is_reloaded[id] || ctx.interferences[id].first.type() != type)
1439 continue;
1440
1441 add_interferences(ctx, is_assigned, slots, slots_used, id);
1442
1443 unsigned slot = find_available_slot(slots_used, ctx.wave_size,
1444 ctx.interferences[id].first.size(),
1445 type == RegType::sgpr, num_slots);
1446
1447 slots[id] = slot;
1448 is_assigned[id] = true;
1449 }
1450
1451 *num_slots = slots_used.size();
1452 }
1453
assign_spill_slots(spill_ctx & ctx,unsigned spills_to_vgpr)1454 void assign_spill_slots(spill_ctx& ctx, unsigned spills_to_vgpr) {
1455 std::vector<uint32_t> slots(ctx.interferences.size());
1456 std::vector<bool> is_assigned(ctx.interferences.size());
1457
1458 /* first, handle affinities: just merge all interferences into both spill ids */
1459 for (std::vector<uint32_t>& vec : ctx.affinities) {
1460 for (unsigned i = 0; i < vec.size(); i++) {
1461 for (unsigned j = i + 1; j < vec.size(); j++) {
1462 assert(vec[i] != vec[j]);
1463 bool reloaded = ctx.is_reloaded[vec[i]] || ctx.is_reloaded[vec[j]];
1464 ctx.is_reloaded[vec[i]] = reloaded;
1465 ctx.is_reloaded[vec[j]] = reloaded;
1466 }
1467 }
1468 }
1469 for (ASSERTED uint32_t i = 0; i < ctx.interferences.size(); i++)
1470 for (ASSERTED uint32_t id : ctx.interferences[i].second)
1471 assert(i != id);
1472
1473 /* for each spill slot, assign as many spill ids as possible */
1474 unsigned sgpr_spill_slots = 0, vgpr_spill_slots = 0;
1475 assign_spill_slots_helper(ctx, RegType::sgpr, is_assigned, slots, &sgpr_spill_slots);
1476 assign_spill_slots_helper(ctx, RegType::vgpr, is_assigned, slots, &vgpr_spill_slots);
1477
1478 for (unsigned id = 0; id < is_assigned.size(); id++)
1479 assert(is_assigned[id] || !ctx.is_reloaded[id]);
1480
1481 for (std::vector<uint32_t>& vec : ctx.affinities) {
1482 for (unsigned i = 0; i < vec.size(); i++) {
1483 for (unsigned j = i + 1; j < vec.size(); j++) {
1484 assert(is_assigned[vec[i]] == is_assigned[vec[j]]);
1485 if (!is_assigned[vec[i]])
1486 continue;
1487 assert(ctx.is_reloaded[vec[i]] == ctx.is_reloaded[vec[j]]);
1488 assert(ctx.interferences[vec[i]].first.type() == ctx.interferences[vec[j]].first.type());
1489 assert(slots[vec[i]] == slots[vec[j]]);
1490 }
1491 }
1492 }
1493
1494 /* hope, we didn't mess up */
1495 std::vector<Temp> vgpr_spill_temps((sgpr_spill_slots + ctx.wave_size - 1) / ctx.wave_size);
1496 assert(vgpr_spill_temps.size() <= spills_to_vgpr);
1497
1498 /* replace pseudo instructions with actual hardware instructions */
1499 Temp scratch_offset = ctx.program->scratch_offset, scratch_rsrc = Temp();
1500 unsigned last_top_level_block_idx = 0;
1501 std::vector<bool> reload_in_loop(vgpr_spill_temps.size());
1502 for (Block& block : ctx.program->blocks) {
1503
1504 /* after loops, we insert a user if there was a reload inside the loop */
1505 if (block.loop_nest_depth == 0) {
1506 int end_vgprs = 0;
1507 for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1508 if (reload_in_loop[i])
1509 end_vgprs++;
1510 }
1511
1512 if (end_vgprs > 0) {
1513 aco_ptr<Instruction> destr{create_instruction<Pseudo_instruction>(aco_opcode::p_end_linear_vgpr, Format::PSEUDO, end_vgprs, 0)};
1514 int k = 0;
1515 for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1516 if (reload_in_loop[i])
1517 destr->operands[k++] = Operand(vgpr_spill_temps[i]);
1518 reload_in_loop[i] = false;
1519 }
1520 /* find insertion point */
1521 std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
1522 while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
1523 ++it;
1524 block.instructions.insert(it, std::move(destr));
1525 }
1526 }
1527
1528 if (block.kind & block_kind_top_level && !block.linear_preds.empty()) {
1529 last_top_level_block_idx = block.index;
1530
1531 /* check if any spilled variables use a created linear vgpr, otherwise destroy them */
1532 for (unsigned i = 0; i < vgpr_spill_temps.size(); i++) {
1533 if (vgpr_spill_temps[i] == Temp())
1534 continue;
1535
1536 bool can_destroy = true;
1537 for (std::pair<Temp, uint32_t> pair : ctx.spills_exit[block.linear_preds[0]]) {
1538
1539 if (ctx.interferences[pair.second].first.type() == RegType::sgpr &&
1540 slots[pair.second] / ctx.wave_size == i) {
1541 can_destroy = false;
1542 break;
1543 }
1544 }
1545 if (can_destroy)
1546 vgpr_spill_temps[i] = Temp();
1547 }
1548 }
1549
1550 std::vector<aco_ptr<Instruction>>::iterator it;
1551 std::vector<aco_ptr<Instruction>> instructions;
1552 instructions.reserve(block.instructions.size());
1553 Builder bld(ctx.program, &instructions);
1554 for (it = block.instructions.begin(); it != block.instructions.end(); ++it) {
1555
1556 if ((*it)->opcode == aco_opcode::p_spill) {
1557 uint32_t spill_id = (*it)->operands[1].constantValue();
1558
1559 if (!ctx.is_reloaded[spill_id]) {
1560 /* never reloaded, so don't spill */
1561 } else if (!is_assigned[spill_id]) {
1562 unreachable("No spill slot assigned for spill id");
1563 } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) {
1564 /* spill vgpr */
1565 ctx.program->config->spilled_vgprs += (*it)->operands[0].size();
1566 uint32_t spill_slot = slots[spill_id];
1567 bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096;
1568 unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
1569
1570 /* check if the scratch resource descriptor already exists */
1571 if (scratch_rsrc == Temp()) {
1572 unsigned offset = add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
1573 scratch_rsrc = load_scratch_resource(ctx, scratch_offset,
1574 last_top_level_block_idx == block.index ?
1575 instructions : ctx.program->blocks[last_top_level_block_idx].instructions,
1576 offset,
1577 last_top_level_block_idx == block.index);
1578 }
1579
1580 unsigned offset = base_offset + spill_slot * 4;
1581 aco_opcode opcode = aco_opcode::buffer_store_dword;
1582 assert((*it)->operands[0].isTemp());
1583 Temp temp = (*it)->operands[0].getTemp();
1584 assert(temp.type() == RegType::vgpr && !temp.is_linear());
1585 if (temp.size() > 1) {
1586 Instruction* split{create_instruction<Pseudo_instruction>(aco_opcode::p_split_vector, Format::PSEUDO, 1, temp.size())};
1587 split->operands[0] = Operand(temp);
1588 for (unsigned i = 0; i < temp.size(); i++)
1589 split->definitions[i] = bld.def(v1);
1590 bld.insert(split);
1591 for (unsigned i = 0; i < temp.size(); i++) {
1592 Instruction *instr = bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset, split->definitions[i].getTemp(), offset + i * 4, false, true);
1593 static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1594 }
1595 } else {
1596 Instruction *instr = bld.mubuf(opcode, scratch_rsrc, Operand(v1), scratch_offset, temp, offset, false, true);
1597 static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1598 }
1599 } else {
1600 ctx.program->config->spilled_sgprs += (*it)->operands[0].size();
1601
1602 uint32_t spill_slot = slots[spill_id];
1603
1604 /* check if the linear vgpr already exists */
1605 if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) {
1606 Temp linear_vgpr = ctx.program->allocateTmp(v1.as_linear());
1607 vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr;
1608 aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
1609 create->definitions[0] = Definition(linear_vgpr);
1610 /* find the right place to insert this definition */
1611 if (last_top_level_block_idx == block.index) {
1612 /* insert right before the current instruction */
1613 instructions.emplace_back(std::move(create));
1614 } else {
1615 assert(last_top_level_block_idx < block.index);
1616 /* insert before the branch at last top level block */
1617 std::vector<aco_ptr<Instruction>>& instructions = ctx.program->blocks[last_top_level_block_idx].instructions;
1618 instructions.insert(std::next(instructions.begin(), instructions.size() - 1), std::move(create));
1619 }
1620 }
1621
1622 /* spill sgpr: just add the vgpr temp to operands */
1623 Pseudo_instruction* spill = create_instruction<Pseudo_instruction>(aco_opcode::p_spill, Format::PSEUDO, 3, 0);
1624 spill->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]);
1625 spill->operands[1] = Operand(spill_slot % ctx.wave_size);
1626 spill->operands[2] = (*it)->operands[0];
1627 instructions.emplace_back(aco_ptr<Instruction>(spill));
1628 }
1629
1630 } else if ((*it)->opcode == aco_opcode::p_reload) {
1631 uint32_t spill_id = (*it)->operands[0].constantValue();
1632 assert(ctx.is_reloaded[spill_id]);
1633
1634 if (!is_assigned[spill_id]) {
1635 unreachable("No spill slot assigned for spill id");
1636 } else if (ctx.interferences[spill_id].first.type() == RegType::vgpr) {
1637 /* reload vgpr */
1638 uint32_t spill_slot = slots[spill_id];
1639 bool add_offset_to_sgpr = ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size + vgpr_spill_slots * 4 > 4096;
1640 unsigned base_offset = add_offset_to_sgpr ? 0 : ctx.program->config->scratch_bytes_per_wave / ctx.program->wave_size;
1641
1642 /* check if the scratch resource descriptor already exists */
1643 if (scratch_rsrc == Temp()) {
1644 unsigned offset = add_offset_to_sgpr ? ctx.program->config->scratch_bytes_per_wave : 0;
1645 scratch_rsrc = load_scratch_resource(ctx, scratch_offset,
1646 last_top_level_block_idx == block.index ?
1647 instructions : ctx.program->blocks[last_top_level_block_idx].instructions,
1648 offset,
1649 last_top_level_block_idx == block.index);
1650 }
1651
1652 unsigned offset = base_offset + spill_slot * 4;
1653 aco_opcode opcode = aco_opcode::buffer_load_dword;
1654 Definition def = (*it)->definitions[0];
1655 if (def.size() > 1) {
1656 Instruction* vec{create_instruction<Pseudo_instruction>(aco_opcode::p_create_vector, Format::PSEUDO, def.size(), 1)};
1657 vec->definitions[0] = def;
1658 for (unsigned i = 0; i < def.size(); i++) {
1659 Temp tmp = bld.tmp(v1);
1660 vec->operands[i] = Operand(tmp);
1661 Instruction *instr = bld.mubuf(opcode, Definition(tmp), scratch_rsrc, Operand(v1), scratch_offset, offset + i * 4, false, true);
1662 static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1663 }
1664 bld.insert(vec);
1665 } else {
1666 Instruction *instr = bld.mubuf(opcode, def, scratch_rsrc, Operand(v1), scratch_offset, offset, false, true);
1667 static_cast<MUBUF_instruction *>(instr)->sync = memory_sync_info(storage_vgpr_spill, semantic_private);
1668 }
1669 } else {
1670 uint32_t spill_slot = slots[spill_id];
1671 reload_in_loop[spill_slot / ctx.wave_size] = block.loop_nest_depth > 0;
1672
1673 /* check if the linear vgpr already exists */
1674 if (vgpr_spill_temps[spill_slot / ctx.wave_size] == Temp()) {
1675 Temp linear_vgpr = ctx.program->allocateTmp(v1.as_linear());
1676 vgpr_spill_temps[spill_slot / ctx.wave_size] = linear_vgpr;
1677 aco_ptr<Pseudo_instruction> create{create_instruction<Pseudo_instruction>(aco_opcode::p_start_linear_vgpr, Format::PSEUDO, 0, 1)};
1678 create->definitions[0] = Definition(linear_vgpr);
1679 /* find the right place to insert this definition */
1680 if (last_top_level_block_idx == block.index) {
1681 /* insert right before the current instruction */
1682 instructions.emplace_back(std::move(create));
1683 } else {
1684 assert(last_top_level_block_idx < block.index);
1685 /* insert before the branch at last top level block */
1686 std::vector<aco_ptr<Instruction>>& instructions = ctx.program->blocks[last_top_level_block_idx].instructions;
1687 instructions.insert(std::next(instructions.begin(), instructions.size() - 1), std::move(create));
1688 }
1689 }
1690
1691 /* reload sgpr: just add the vgpr temp to operands */
1692 Pseudo_instruction* reload = create_instruction<Pseudo_instruction>(aco_opcode::p_reload, Format::PSEUDO, 2, 1);
1693 reload->operands[0] = Operand(vgpr_spill_temps[spill_slot / ctx.wave_size]);
1694 reload->operands[1] = Operand(spill_slot % ctx.wave_size);
1695 reload->definitions[0] = (*it)->definitions[0];
1696 instructions.emplace_back(aco_ptr<Instruction>(reload));
1697 }
1698 } else if (!ctx.remat_used.count(it->get()) || ctx.remat_used[it->get()]) {
1699 instructions.emplace_back(std::move(*it));
1700 }
1701
1702 }
1703 block.instructions = std::move(instructions);
1704 }
1705
1706 /* update required scratch memory */
1707 ctx.program->config->scratch_bytes_per_wave += align(vgpr_spill_slots * 4 * ctx.program->wave_size, 1024);
1708
1709 /* SSA elimination inserts copies for logical phis right before p_logical_end
1710 * So if a linear vgpr is used between that p_logical_end and the branch,
1711 * we need to ensure logical phis don't choose a definition which aliases
1712 * the linear vgpr.
1713 * TODO: Moving the spills and reloads to before p_logical_end might produce
1714 * slightly better code. */
1715 for (Block& block : ctx.program->blocks) {
1716 /* loops exits are already handled */
1717 if (block.logical_preds.size() <= 1)
1718 continue;
1719
1720 bool has_logical_phis = false;
1721 for (aco_ptr<Instruction>& instr : block.instructions) {
1722 if (instr->opcode == aco_opcode::p_phi) {
1723 has_logical_phis = true;
1724 break;
1725 } else if (instr->opcode != aco_opcode::p_linear_phi) {
1726 break;
1727 }
1728 }
1729 if (!has_logical_phis)
1730 continue;
1731
1732 std::set<Temp> vgprs;
1733 for (unsigned pred_idx : block.logical_preds) {
1734 Block& pred = ctx.program->blocks[pred_idx];
1735 for (int i = pred.instructions.size() - 1; i >= 0; i--) {
1736 aco_ptr<Instruction>& pred_instr = pred.instructions[i];
1737 if (pred_instr->opcode == aco_opcode::p_logical_end) {
1738 break;
1739 } else if (pred_instr->opcode == aco_opcode::p_spill ||
1740 pred_instr->opcode == aco_opcode::p_reload) {
1741 vgprs.insert(pred_instr->operands[0].getTemp());
1742 }
1743 }
1744 }
1745 if (!vgprs.size())
1746 continue;
1747
1748 aco_ptr<Instruction> destr{create_instruction<Pseudo_instruction>(aco_opcode::p_end_linear_vgpr, Format::PSEUDO, vgprs.size(), 0)};
1749 int k = 0;
1750 for (Temp tmp : vgprs) {
1751 destr->operands[k++] = Operand(tmp);
1752 }
1753 /* find insertion point */
1754 std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.begin();
1755 while ((*it)->opcode == aco_opcode::p_linear_phi || (*it)->opcode == aco_opcode::p_phi)
1756 ++it;
1757 block.instructions.insert(it, std::move(destr));
1758 }
1759 }
1760
1761 } /* end namespace */
1762
1763
spill(Program * program,live & live_vars)1764 void spill(Program* program, live& live_vars)
1765 {
1766 program->config->spilled_vgprs = 0;
1767 program->config->spilled_sgprs = 0;
1768
1769 /* no spilling when register pressure is low enough */
1770 if (program->num_waves > 0)
1771 return;
1772
1773 /* lower to CSSA before spilling to ensure correctness w.r.t. phis */
1774 lower_to_cssa(program, live_vars);
1775
1776 /* calculate target register demand */
1777 RegisterDemand register_target = program->max_reg_demand;
1778 if (register_target.sgpr > program->sgpr_limit)
1779 register_target.vgpr += (register_target.sgpr - program->sgpr_limit + program->wave_size - 1 + 32) / program->wave_size;
1780 register_target.sgpr = program->sgpr_limit;
1781
1782 if (register_target.vgpr > program->vgpr_limit)
1783 register_target.sgpr = program->sgpr_limit - 5;
1784 int spills_to_vgpr = (program->max_reg_demand.sgpr - register_target.sgpr + program->wave_size - 1 + 32) / program->wave_size;
1785 register_target.vgpr = program->vgpr_limit - spills_to_vgpr;
1786
1787 /* initialize ctx */
1788 spill_ctx ctx(register_target, program, live_vars.register_demand);
1789 compute_global_next_uses(ctx);
1790 get_rematerialize_info(ctx);
1791
1792 /* create spills and reloads */
1793 for (unsigned i = 0; i < program->blocks.size(); i++)
1794 spill_block(ctx, i);
1795
1796 /* assign spill slots and DCE rematerialized code */
1797 assign_spill_slots(ctx, spills_to_vgpr);
1798
1799 /* update live variable information */
1800 live_vars = live_var_analysis(program);
1801
1802 assert(program->num_waves > 0);
1803 }
1804
1805 }
1806
1807