xref: /aosp_15_r20/external/mesa3d/src/amd/compiler/aco_insert_exec_mask.cpp (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2019 Valve Corporation
3  *
4  * SPDX-License-Identifier: MIT
5  */
6 
7 #include "aco_builder.h"
8 #include "aco_ir.h"
9 
10 #include "util/u_math.h"
11 
12 #include <set>
13 #include <vector>
14 
15 namespace aco {
16 
17 namespace {
18 
19 enum WQMState : uint8_t {
20    Unspecified = 0,
21    Exact = 1 << 0,
22    WQM = 1 << 1, /* with control flow applied */
23 };
24 
25 enum mask_type : uint8_t {
26    mask_type_global = 1 << 0,
27    mask_type_exact = 1 << 1,
28    mask_type_wqm = 1 << 2,
29    mask_type_loop = 1 << 3, /* active lanes of a loop */
30 };
31 
32 struct loop_info {
33    Block* loop_header;
34    uint16_t num_exec_masks;
35    bool has_divergent_break;
36    bool has_divergent_continue;
37    bool has_discard; /* has a discard or demote */
loop_infoaco::__anon9ac258bc0111::loop_info38    loop_info(Block* b, uint16_t num, bool breaks, bool cont, bool discard)
39        : loop_header(b), num_exec_masks(num), has_divergent_break(breaks),
40          has_divergent_continue(cont), has_discard(discard)
41    {}
42 };
43 
44 struct block_info {
45    std::vector<std::pair<Operand, uint8_t>>
46       exec; /* Vector of exec masks. Either a temporary or const -1. */
47 };
48 
49 struct exec_ctx {
50    Program* program;
51    std::vector<block_info> info;
52    std::vector<loop_info> loop;
53    bool handle_wqm = false;
exec_ctxaco::__anon9ac258bc0111::exec_ctx54    exec_ctx(Program* program_) : program(program_), info(program->blocks.size()) {}
55 };
56 
57 bool
needs_exact(aco_ptr<Instruction> & instr)58 needs_exact(aco_ptr<Instruction>& instr)
59 {
60    if (instr->isMUBUF()) {
61       return instr->mubuf().disable_wqm;
62    } else if (instr->isMTBUF()) {
63       return instr->mtbuf().disable_wqm;
64    } else if (instr->isMIMG()) {
65       return instr->mimg().disable_wqm;
66    } else if (instr->isFlatLike()) {
67       return instr->flatlike().disable_wqm;
68    } else {
69       /* Require Exact for p_jump_to_epilog because if p_exit_early_if is
70        * emitted inside the same block, the main FS will always jump to the PS
71        * epilog without considering the exec mask.
72        */
73       return instr->isEXP() || instr->opcode == aco_opcode::p_jump_to_epilog ||
74              instr->opcode == aco_opcode::p_dual_src_export_gfx11;
75    }
76 }
77 
78 WQMState
get_instr_needs(aco_ptr<Instruction> & instr)79 get_instr_needs(aco_ptr<Instruction>& instr)
80 {
81    if (needs_exact(instr))
82       return Exact;
83 
84    bool pred_by_exec = needs_exec_mask(instr.get()) || instr->opcode == aco_opcode::p_logical_end ||
85                        instr->isBranch();
86 
87    return pred_by_exec ? WQM : Unspecified;
88 }
89 
90 Operand
get_exec_op(Operand t)91 get_exec_op(Operand t)
92 {
93    if (t.isUndefined())
94       return Operand(exec, t.regClass());
95    else
96       return t;
97 }
98 
99 void
transition_to_WQM(exec_ctx & ctx,Builder bld,unsigned idx)100 transition_to_WQM(exec_ctx& ctx, Builder bld, unsigned idx)
101 {
102    if (ctx.info[idx].exec.back().second & mask_type_wqm)
103       return;
104    if (ctx.info[idx].exec.back().second & mask_type_global) {
105       Operand exec_mask = ctx.info[idx].exec.back().first;
106       if (exec_mask.isUndefined())
107          ctx.info[idx].exec.back().first = bld.copy(bld.def(bld.lm), Operand(exec, bld.lm));
108 
109       exec_mask = bld.sop1(Builder::s_wqm, Definition(exec, bld.lm), bld.def(s1, scc),
110                            get_exec_op(exec_mask));
111       ctx.info[idx].exec.emplace_back(exec_mask, mask_type_global | mask_type_wqm);
112       return;
113    }
114    /* otherwise, the WQM mask should be one below the current mask */
115    ctx.info[idx].exec.pop_back();
116    assert(ctx.info[idx].exec.back().second & mask_type_wqm);
117    assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
118    assert(ctx.info[idx].exec.back().first.isTemp());
119    ctx.info[idx].exec.back().first =
120       bld.copy(Definition(exec, bld.lm), ctx.info[idx].exec.back().first);
121 }
122 
123 void
transition_to_Exact(exec_ctx & ctx,Builder bld,unsigned idx)124 transition_to_Exact(exec_ctx& ctx, Builder bld, unsigned idx)
125 {
126    if (ctx.info[idx].exec.back().second & mask_type_exact)
127       return;
128    /* We can't remove the loop exec mask, because that can cause exec.size() to
129     * be less than num_exec_masks. The loop exec mask also needs to be kept
130     * around for various uses. */
131    if ((ctx.info[idx].exec.back().second & mask_type_global) &&
132        !(ctx.info[idx].exec.back().second & mask_type_loop)) {
133       ctx.info[idx].exec.pop_back();
134       assert(ctx.info[idx].exec.back().second & mask_type_exact);
135       assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
136       assert(ctx.info[idx].exec.back().first.isTemp());
137       ctx.info[idx].exec.back().first =
138          bld.copy(Definition(exec, bld.lm), ctx.info[idx].exec.back().first);
139       return;
140    }
141    /* otherwise, we create an exact mask and push to the stack */
142    Operand wqm = ctx.info[idx].exec.back().first;
143    if (wqm.isUndefined()) {
144       wqm = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
145                      Definition(exec, bld.lm), ctx.info[idx].exec[0].first, Operand(exec, bld.lm));
146    } else {
147       bld.sop2(Builder::s_and, Definition(exec, bld.lm), bld.def(s1, scc),
148                ctx.info[idx].exec[0].first, wqm);
149    }
150    ctx.info[idx].exec.back().first = Operand(wqm);
151    ctx.info[idx].exec.emplace_back(Operand(bld.lm), mask_type_exact);
152 }
153 
154 unsigned
add_coupling_code(exec_ctx & ctx,Block * block,std::vector<aco_ptr<Instruction>> & instructions)155 add_coupling_code(exec_ctx& ctx, Block* block, std::vector<aco_ptr<Instruction>>& instructions)
156 {
157    unsigned idx = block->index;
158    Builder bld(ctx.program, &instructions);
159    Block::edge_vec& preds = block->linear_preds;
160    bool restore_exec = false;
161 
162    /* start block */
163    if (preds.empty()) {
164       aco_ptr<Instruction>& startpgm = block->instructions[0];
165       assert(startpgm->opcode == aco_opcode::p_startpgm);
166       bld.insert(std::move(startpgm));
167 
168       unsigned count = 1;
169       while (block->instructions[count]->opcode == aco_opcode::p_init_scratch ||
170              block->instructions[count]->opcode == aco_opcode::s_setprio) {
171          bld.insert(std::move(block->instructions[count]));
172          count++;
173       }
174 
175       Operand start_exec(bld.lm);
176 
177       /* exec seems to need to be manually initialized with combined shaders */
178       if (ctx.program->stage.num_sw_stages() > 1 ||
179           ctx.program->stage.hw == AC_HW_NEXT_GEN_GEOMETRY_SHADER ||
180           (ctx.program->stage.sw == SWStage::VS &&
181            (ctx.program->stage.hw == AC_HW_HULL_SHADER ||
182             ctx.program->stage.hw == AC_HW_LEGACY_GEOMETRY_SHADER)) ||
183           (ctx.program->stage.sw == SWStage::TES &&
184            ctx.program->stage.hw == AC_HW_LEGACY_GEOMETRY_SHADER)) {
185          start_exec = Operand::c32_or_c64(-1u, bld.lm == s2);
186          bld.copy(Definition(exec, bld.lm), start_exec);
187       }
188 
189       /* EXEC is automatically initialized by the HW for compute shaders.
190        * We know for sure exec is initially -1 when the shader always has full subgroups.
191        */
192       if (ctx.program->stage == compute_cs && ctx.program->info.cs.uses_full_subgroups)
193          start_exec = Operand::c32_or_c64(-1u, bld.lm == s2);
194 
195       if (ctx.handle_wqm) {
196          ctx.info[idx].exec.emplace_back(start_exec, mask_type_global | mask_type_exact);
197          /* Initialize WQM already */
198          transition_to_WQM(ctx, bld, idx);
199       } else {
200          uint8_t mask = mask_type_global;
201          if (ctx.program->needs_wqm) {
202             bld.sop1(Builder::s_wqm, Definition(exec, bld.lm), bld.def(s1, scc),
203                      Operand(exec, bld.lm));
204             mask |= mask_type_wqm;
205          } else {
206             mask |= mask_type_exact;
207          }
208          ctx.info[idx].exec.emplace_back(start_exec, mask);
209       }
210 
211       return count;
212    }
213 
214    /* loop entry block */
215    if (block->kind & block_kind_loop_header) {
216       assert(preds[0] == idx - 1);
217       ctx.info[idx].exec = ctx.info[idx - 1].exec;
218       loop_info& info = ctx.loop.back();
219       assert(ctx.info[idx].exec.size() == info.num_exec_masks);
220 
221       /* create ssa names for outer exec masks */
222       if (info.has_discard && preds.size() > 1) {
223          aco_ptr<Instruction> phi;
224          for (int i = 0; i < info.num_exec_masks - 1; i++) {
225             phi.reset(
226                create_instruction(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1));
227             phi->definitions[0] = bld.def(bld.lm);
228             phi->operands[0] = get_exec_op(ctx.info[preds[0]].exec[i].first);
229             ctx.info[idx].exec[i].first = bld.insert(std::move(phi));
230          }
231       }
232 
233       ctx.info[idx].exec.back().second |= mask_type_loop;
234 
235       if (info.has_divergent_continue) {
236          /* create ssa name for loop active mask */
237          aco_ptr<Instruction> phi{
238             create_instruction(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
239          phi->definitions[0] = bld.def(bld.lm);
240          phi->operands[0] = get_exec_op(ctx.info[preds[0]].exec.back().first);
241          ctx.info[idx].exec.back().first = bld.insert(std::move(phi));
242 
243          restore_exec = true;
244          uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
245          ctx.info[idx].exec.emplace_back(ctx.info[idx].exec.back().first, mask_type);
246       }
247 
248    } else if (block->kind & block_kind_loop_exit) {
249       Block* header = ctx.loop.back().loop_header;
250       loop_info& info = ctx.loop.back();
251 
252       for (ASSERTED unsigned pred : preds)
253          assert(ctx.info[pred].exec.size() >= info.num_exec_masks);
254 
255       /* fill the loop header phis */
256       Block::edge_vec& header_preds = header->linear_preds;
257       int instr_idx = 0;
258       if (info.has_discard && header_preds.size() > 1) {
259          while (instr_idx < info.num_exec_masks - 1) {
260             aco_ptr<Instruction>& phi = header->instructions[instr_idx];
261             assert(phi->opcode == aco_opcode::p_linear_phi);
262             for (unsigned i = 1; i < phi->operands.size(); i++)
263                phi->operands[i] = get_exec_op(ctx.info[header_preds[i]].exec[instr_idx].first);
264             instr_idx++;
265          }
266       }
267 
268       if (info.has_divergent_continue) {
269          aco_ptr<Instruction>& phi = header->instructions[instr_idx++];
270          assert(phi->opcode == aco_opcode::p_linear_phi);
271          for (unsigned i = 1; i < phi->operands.size(); i++)
272             phi->operands[i] =
273                get_exec_op(ctx.info[header_preds[i]].exec[info.num_exec_masks - 1].first);
274          restore_exec = true;
275       }
276 
277       if (info.has_divergent_break) {
278          restore_exec = true;
279          /* Drop the loop active mask. */
280          info.num_exec_masks--;
281       }
282       assert(!(block->kind & block_kind_top_level) || info.num_exec_masks <= 2);
283 
284       /* create the loop exit phis if not trivial */
285       for (unsigned exec_idx = 0; exec_idx < info.num_exec_masks; exec_idx++) {
286          Operand same = ctx.info[preds[0]].exec[exec_idx].first;
287          uint8_t type = ctx.info[header_preds[0]].exec[exec_idx].second;
288          bool trivial = true;
289 
290          for (unsigned i = 1; i < preds.size() && trivial; i++) {
291             if (ctx.info[preds[i]].exec[exec_idx].first != same)
292                trivial = false;
293          }
294 
295          if (trivial) {
296             ctx.info[idx].exec.emplace_back(same, type);
297          } else {
298             /* create phi for loop footer */
299             aco_ptr<Instruction> phi{
300                create_instruction(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
301             phi->definitions[0] = bld.def(bld.lm);
302             for (unsigned i = 0; i < phi->operands.size(); i++)
303                phi->operands[i] = get_exec_op(ctx.info[preds[i]].exec[exec_idx].first);
304             ctx.info[idx].exec.emplace_back(bld.insert(std::move(phi)), type);
305          }
306       }
307 
308       assert(ctx.info[idx].exec.size() == info.num_exec_masks);
309       ctx.loop.pop_back();
310 
311    } else if (preds.size() == 1) {
312       ctx.info[idx].exec = ctx.info[preds[0]].exec;
313    } else {
314       assert(preds.size() == 2);
315       /* if one of the predecessors ends in exact mask, we pop it from stack */
316       unsigned num_exec_masks =
317          std::min(ctx.info[preds[0]].exec.size(), ctx.info[preds[1]].exec.size());
318 
319       if (block->kind & block_kind_merge) {
320          restore_exec = true;
321          num_exec_masks--;
322       }
323       if (block->kind & block_kind_top_level)
324          num_exec_masks = std::min(num_exec_masks, 2u);
325 
326       /* create phis for diverged exec masks */
327       for (unsigned i = 0; i < num_exec_masks; i++) {
328          /* skip trivial phis */
329          if (ctx.info[preds[0]].exec[i].first == ctx.info[preds[1]].exec[i].first) {
330             Operand t = ctx.info[preds[0]].exec[i].first;
331             /* discard/demote can change the state of the current exec mask */
332             assert(!t.isTemp() ||
333                    ctx.info[preds[0]].exec[i].second == ctx.info[preds[1]].exec[i].second);
334             uint8_t mask = ctx.info[preds[0]].exec[i].second & ctx.info[preds[1]].exec[i].second;
335             ctx.info[idx].exec.emplace_back(t, mask);
336             continue;
337          }
338 
339          Temp phi = bld.pseudo(aco_opcode::p_linear_phi, bld.def(bld.lm),
340                                get_exec_op(ctx.info[preds[0]].exec[i].first),
341                                get_exec_op(ctx.info[preds[1]].exec[i].first));
342          uint8_t mask_type = ctx.info[preds[0]].exec[i].second & ctx.info[preds[1]].exec[i].second;
343          ctx.info[idx].exec.emplace_back(phi, mask_type);
344       }
345    }
346 
347    unsigned i = 0;
348    while (block->instructions[i]->opcode == aco_opcode::p_phi ||
349           block->instructions[i]->opcode == aco_opcode::p_linear_phi) {
350       bld.insert(std::move(block->instructions[i]));
351       i++;
352    }
353 
354    if (ctx.handle_wqm) {
355       /* End WQM handling if not needed anymore */
356       if (block->kind & block_kind_top_level && ctx.info[idx].exec.size() == 2) {
357          if (block->instructions[i]->opcode == aco_opcode::p_end_wqm) {
358             ctx.info[idx].exec.back().second |= mask_type_global;
359             transition_to_Exact(ctx, bld, idx);
360             ctx.handle_wqm = false;
361             restore_exec = false;
362             i++;
363          }
364       }
365    }
366 
367    /* restore exec mask after divergent control flow */
368    if (restore_exec) {
369       Operand restore = get_exec_op(ctx.info[idx].exec.back().first);
370       assert(restore.size() == bld.lm.size());
371       bld.copy(Definition(exec, bld.lm), restore);
372       if (!restore.isConstant())
373          ctx.info[idx].exec.back().first = Operand(bld.lm);
374    }
375 
376    return i;
377 }
378 
379 /* Avoid live-range splits in Exact mode:
380  * Because the data register of atomic VMEM instructions
381  * is shared between src and dst, it might be necessary
382  * to create live-range splits during RA.
383  * Make the live-range splits explicit in WQM mode.
384  */
385 void
handle_atomic_data(exec_ctx & ctx,Builder & bld,unsigned block_idx,aco_ptr<Instruction> & instr)386 handle_atomic_data(exec_ctx& ctx, Builder& bld, unsigned block_idx, aco_ptr<Instruction>& instr)
387 {
388    /* check if this is an atomic VMEM instruction */
389    int idx = -1;
390    if (!instr->isVMEM() || instr->definitions.empty())
391       return;
392    else if (instr->isMIMG())
393       idx = instr->operands[2].isTemp() ? 2 : -1;
394    else if (instr->operands.size() == 4)
395       idx = 3;
396 
397    if (idx != -1) {
398       /* insert explicit copy of atomic data in WQM-mode */
399       transition_to_WQM(ctx, bld, block_idx);
400       Temp data = instr->operands[idx].getTemp();
401       data = bld.copy(bld.def(data.regClass()), data);
402       instr->operands[idx].setTemp(data);
403    }
404 }
405 
406 void
process_instructions(exec_ctx & ctx,Block * block,std::vector<aco_ptr<Instruction>> & instructions,unsigned idx)407 process_instructions(exec_ctx& ctx, Block* block, std::vector<aco_ptr<Instruction>>& instructions,
408                      unsigned idx)
409 {
410    block_info& info = ctx.info[block->index];
411    WQMState state;
412    if (info.exec.back().second & mask_type_wqm) {
413       state = WQM;
414    } else {
415       assert(!ctx.handle_wqm || info.exec.back().second & mask_type_exact);
416       state = Exact;
417    }
418 
419    Builder bld(ctx.program, &instructions);
420 
421    for (; idx < block->instructions.size(); idx++) {
422       aco_ptr<Instruction> instr = std::move(block->instructions[idx]);
423 
424       WQMState needs = ctx.handle_wqm ? get_instr_needs(instr) : Unspecified;
425 
426       if (needs == WQM && state != WQM) {
427          transition_to_WQM(ctx, bld, block->index);
428          state = WQM;
429       } else if (needs == Exact) {
430          if (ctx.handle_wqm)
431             handle_atomic_data(ctx, bld, block->index, instr);
432          transition_to_Exact(ctx, bld, block->index);
433          state = Exact;
434       }
435 
436       if (instr->opcode == aco_opcode::p_discard_if) {
437          Operand current_exec = Operand(exec, bld.lm);
438 
439          if (block->instructions[idx + 1]->opcode == aco_opcode::p_end_wqm) {
440             /* Transition to Exact without extra instruction. */
441             info.exec.resize(1);
442             assert(info.exec[0].second == (mask_type_exact | mask_type_global));
443             current_exec = get_exec_op(info.exec[0].first);
444             info.exec[0].first = Operand(bld.lm);
445             state = Exact;
446          } else if (info.exec.size() >= 2 && ctx.handle_wqm) {
447             /* Preserve the WQM mask */
448             info.exec[1].second &= ~mask_type_global;
449          }
450 
451          Temp cond, exit_cond;
452          if (instr->operands[0].isConstant()) {
453             assert(instr->operands[0].constantValue() == -1u);
454             /* save condition and set exec to zero */
455             exit_cond = bld.tmp(s1);
456             cond =
457                bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.scc(Definition(exit_cond)),
458                         Definition(exec, bld.lm), Operand::zero(), Operand(exec, bld.lm));
459          } else {
460             cond = instr->operands[0].getTemp();
461             /* discard from current exec */
462             exit_cond = bld.sop2(Builder::s_andn2, Definition(exec, bld.lm), bld.def(s1, scc),
463                                  current_exec, cond)
464                            .def(1)
465                            .getTemp();
466          }
467 
468          /* discard from inner to outer exec mask on stack */
469          int num = info.exec.size() - 2;
470          for (int i = num; i >= 0; i--) {
471             Instruction* andn2 = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.def(s1, scc),
472                                           info.exec[i].first, cond);
473             info.exec[i].first = Operand(andn2->definitions[0].getTemp());
474             exit_cond = andn2->definitions[1].getTemp();
475          }
476 
477          instr->opcode = aco_opcode::p_exit_early_if;
478          instr->operands[0] = bld.scc(exit_cond);
479          assert(!ctx.handle_wqm || (info.exec[0].second & mask_type_wqm) == 0);
480 
481       } else if (instr->opcode == aco_opcode::p_is_helper) {
482          Definition dst = instr->definitions[0];
483          assert(dst.size() == bld.lm.size());
484          if (state == Exact) {
485             instr.reset(create_instruction(bld.w64or32(Builder::s_mov), Format::SOP1, 1, 1));
486             instr->operands[0] = Operand::zero();
487             instr->definitions[0] = dst;
488          } else {
489             std::pair<Operand, uint8_t>& exact_mask = info.exec[0];
490             assert(exact_mask.second & mask_type_exact);
491 
492             instr.reset(create_instruction(bld.w64or32(Builder::s_andn2), Format::SOP2, 2, 2));
493             instr->operands[0] = Operand(exec, bld.lm); /* current exec */
494             instr->operands[1] = Operand(exact_mask.first);
495             instr->definitions[0] = dst;
496             instr->definitions[1] = bld.def(s1, scc);
497          }
498       } else if (instr->opcode == aco_opcode::p_demote_to_helper) {
499          assert((info.exec[0].second & mask_type_exact) &&
500                 (info.exec[0].second & mask_type_global));
501 
502          const bool nested_cf = !(info.exec.back().second & mask_type_global);
503          if (ctx.handle_wqm && state == Exact && nested_cf) {
504             /* Transition back to WQM without extra instruction. */
505             info.exec.pop_back();
506             state = WQM;
507          } else if (block->instructions[idx + 1]->opcode == aco_opcode::p_end_wqm) {
508             /* Transition to Exact without extra instruction. */
509             info.exec.resize(1);
510             state = Exact;
511          } else if (nested_cf) {
512             /* Save curent exec temporarily. */
513             info.exec.back().first = bld.copy(bld.def(bld.lm), Operand(exec, bld.lm));
514          }
515 
516          /* Remove invocations from global exact mask. */
517          Definition def = state == Exact ? Definition(exec, bld.lm) : bld.def(bld.lm);
518          Operand src = instr->operands[0].isConstant() ? Operand(exec, bld.lm) : instr->operands[0];
519 
520          Definition exit_cond =
521             bld.sop2(Builder::s_andn2, def, bld.def(s1, scc), get_exec_op(info.exec[0].first), src)
522                .def(1);
523          info.exec[0].first = Operand(def.getTemp());
524 
525          /* Update global WQM mask and store in exec. */
526          if (state == WQM) {
527             assert(info.exec.size() > 1);
528             exit_cond =
529                bld.sop1(Builder::s_wqm, Definition(exec, bld.lm), bld.def(s1, scc), def.getTemp())
530                   .def(1);
531          }
532 
533          /* End shader if global mask is zero. */
534          instr->opcode = aco_opcode::p_exit_early_if;
535          instr->operands[0] = bld.scc(exit_cond.getTemp());
536          bld.insert(std::move(instr));
537 
538          /* Update all other exec masks. */
539          if (nested_cf) {
540             const unsigned global_idx = state == WQM ? 1 : 0;
541             for (unsigned i = global_idx + 1; i < info.exec.size() - 1; i++) {
542                info.exec[i].first =
543                   bld.sop2(Builder::s_and, bld.def(bld.lm), bld.def(s1, scc),
544                            get_exec_op(info.exec[i].first), Operand(exec, bld.lm));
545             }
546             /* Update current exec and save WQM mask. */
547             info.exec[global_idx].first =
548                bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
549                         Definition(exec, bld.lm), info.exec.back().first, Operand(exec, bld.lm));
550             info.exec.back().first = Operand(bld.lm);
551          }
552          continue;
553 
554       } else if (instr->opcode == aco_opcode::p_elect) {
555          bool all_lanes_enabled = info.exec.back().first.constantEquals(-1u);
556          Definition dst = instr->definitions[0];
557 
558          if (all_lanes_enabled) {
559             bld.copy(Definition(dst), Operand::c32_or_c64(1u, dst.size() == 2));
560          } else {
561             Temp first_lane_idx = bld.sop1(Builder::s_ff1_i32, bld.def(s1), Operand(exec, bld.lm));
562             bld.sop2(Builder::s_lshl, Definition(dst), bld.def(s1, scc),
563                      Operand::c32_or_c64(1u, dst.size() == 2), Operand(first_lane_idx));
564          }
565          continue;
566       } else if (instr->opcode == aco_opcode::p_end_wqm) {
567          assert(block->kind & block_kind_top_level);
568          assert(info.exec.size() <= 2);
569          /* This instruction indicates the end of WQM mode. */
570          info.exec.back().second |= mask_type_global;
571          transition_to_Exact(ctx, bld, block->index);
572          state = Exact;
573          ctx.handle_wqm = false;
574          continue;
575       }
576 
577       bld.insert(std::move(instr));
578    }
579 }
580 
581 void
add_branch_code(exec_ctx & ctx,Block * block)582 add_branch_code(exec_ctx& ctx, Block* block)
583 {
584    unsigned idx = block->index;
585    Builder bld(ctx.program, block);
586 
587    if (block->linear_succs.empty())
588       return;
589 
590    if (block->kind & block_kind_loop_preheader) {
591       /* collect information about the succeeding loop */
592       bool has_divergent_break = false;
593       bool has_divergent_continue = false;
594       bool has_discard = false;
595       unsigned loop_nest_depth = ctx.program->blocks[idx + 1].loop_nest_depth;
596 
597       for (unsigned i = idx + 1; ctx.program->blocks[i].loop_nest_depth >= loop_nest_depth; i++) {
598          Block& loop_block = ctx.program->blocks[i];
599 
600          if (loop_block.kind & block_kind_uses_discard)
601             has_discard = true;
602          if (loop_block.loop_nest_depth != loop_nest_depth)
603             continue;
604 
605          if (loop_block.kind & block_kind_uniform)
606             continue;
607          else if (loop_block.kind & block_kind_break)
608             has_divergent_break = true;
609          else if (loop_block.kind & block_kind_continue)
610             has_divergent_continue = true;
611       }
612 
613       if (has_divergent_break) {
614          /* save restore exec mask */
615          uint8_t mask = ctx.info[idx].exec.back().second;
616          if (ctx.info[idx].exec.back().first.constantEquals(-1u)) {
617             ctx.info[idx].exec.emplace_back(Operand(exec, bld.lm), mask);
618          } else {
619             bld.reset(bld.instructions, std::prev(bld.instructions->end()));
620             Operand restore = bld.copy(bld.def(bld.lm), Operand(exec, bld.lm));
621             ctx.info[idx].exec.emplace(std::prev(ctx.info[idx].exec.end()), restore, mask);
622             bld.reset(bld.instructions);
623          }
624          ctx.info[idx].exec.back().second &= (mask_type_wqm | mask_type_exact);
625       }
626       unsigned num_exec_masks = ctx.info[idx].exec.size();
627 
628       ctx.loop.emplace_back(&ctx.program->blocks[block->linear_succs[0]], num_exec_masks,
629                             has_divergent_break, has_divergent_continue, has_discard);
630    }
631 
632    /* For normal breaks, this is the exec mask. For discard+break, it's the
633     * old exec mask before it was zero'd.
634     */
635    Operand break_cond = Operand(exec, bld.lm);
636 
637    if (block->kind & block_kind_continue_or_break) {
638       assert(ctx.program->blocks[ctx.program->blocks[block->linear_succs[1]].linear_succs[0]].kind &
639              block_kind_loop_header);
640       assert(ctx.program->blocks[ctx.program->blocks[block->linear_succs[0]].linear_succs[0]].kind &
641              block_kind_loop_exit);
642       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
643       block->instructions.pop_back();
644 
645       bool need_parallelcopy = false;
646       while (!(ctx.info[idx].exec.back().second & mask_type_loop)) {
647          ctx.info[idx].exec.pop_back();
648          need_parallelcopy = true;
649       }
650 
651       if (need_parallelcopy)
652          ctx.info[idx].exec.back().first =
653             bld.copy(Definition(exec, bld.lm), ctx.info[idx].exec.back().first);
654       bld.branch(aco_opcode::p_cbranch_nz, bld.def(s2), Operand(exec, bld.lm),
655                  block->linear_succs[1], block->linear_succs[0]);
656       return;
657    }
658 
659    if (block->kind & block_kind_uniform) {
660       Pseudo_branch_instruction& branch = block->instructions.back()->branch();
661       if (branch.opcode == aco_opcode::p_branch) {
662          branch.target[0] = block->linear_succs[0];
663       } else {
664          branch.target[0] = block->linear_succs[1];
665          branch.target[1] = block->linear_succs[0];
666       }
667       return;
668    }
669 
670    if (block->kind & block_kind_branch) {
671       // orig = s_and_saveexec_b64
672       assert(block->linear_succs.size() == 2);
673       assert(block->instructions.back()->opcode == aco_opcode::p_cbranch_z);
674       Temp cond = block->instructions.back()->operands[0].getTemp();
675       aco_ptr<Instruction> branch = std::move(block->instructions.back());
676       block->instructions.pop_back();
677 
678       uint8_t mask_type = ctx.info[idx].exec.back().second & (mask_type_wqm | mask_type_exact);
679       if (ctx.info[idx].exec.back().first.constantEquals(-1u)) {
680          bld.copy(Definition(exec, bld.lm), cond);
681       } else {
682          Temp old_exec = bld.sop1(Builder::s_and_saveexec, bld.def(bld.lm), bld.def(s1, scc),
683                                   Definition(exec, bld.lm), cond, Operand(exec, bld.lm));
684 
685          ctx.info[idx].exec.back().first = Operand(old_exec);
686       }
687 
688       /* add next current exec to the stack */
689       ctx.info[idx].exec.emplace_back(Operand(bld.lm), mask_type);
690 
691       Builder::Result r = bld.branch(aco_opcode::p_cbranch_z, bld.def(s2), Operand(exec, bld.lm),
692                                      block->linear_succs[1], block->linear_succs[0]);
693       r->branch().rarely_taken = branch->branch().rarely_taken;
694       r->branch().never_taken = branch->branch().never_taken;
695       return;
696    }
697 
698    if (block->kind & block_kind_invert) {
699       // exec = s_andn2_b64 (original_exec, exec)
700       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
701       aco_ptr<Instruction> branch = std::move(block->instructions.back());
702       block->instructions.pop_back();
703       assert(ctx.info[idx].exec.size() >= 2);
704       Operand orig_exec = ctx.info[idx].exec[ctx.info[idx].exec.size() - 2].first;
705       bld.sop2(Builder::s_andn2, Definition(exec, bld.lm), bld.def(s1, scc), orig_exec,
706                Operand(exec, bld.lm));
707 
708       Builder::Result r = bld.branch(aco_opcode::p_cbranch_z, bld.def(s2), Operand(exec, bld.lm),
709                                      block->linear_succs[1], block->linear_succs[0]);
710       r->branch().rarely_taken = branch->branch().rarely_taken;
711       r->branch().never_taken = branch->branch().never_taken;
712       return;
713    }
714 
715    if (block->kind & block_kind_break) {
716       // loop_mask = s_andn2_b64 (loop_mask, exec)
717       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
718       block->instructions.pop_back();
719 
720       Temp cond = Temp();
721       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
722          cond = bld.tmp(s1);
723          Operand exec_mask = ctx.info[idx].exec[exec_idx].first;
724          exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
725                               exec_mask, break_cond);
726          ctx.info[idx].exec[exec_idx].first = exec_mask;
727          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
728             break;
729       }
730 
731       /* check if the successor is the merge block, otherwise set exec to 0 */
732       // TODO: this could be done better by directly branching to the merge block
733       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
734       Block& succ = ctx.program->blocks[succ_idx];
735       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
736          bld.copy(Definition(exec, bld.lm), Operand::zero(bld.lm.bytes()));
737       }
738 
739       bld.branch(aco_opcode::p_cbranch_nz, bld.def(s2), bld.scc(cond), block->linear_succs[1],
740                  block->linear_succs[0]);
741       return;
742    }
743 
744    if (block->kind & block_kind_continue) {
745       assert(block->instructions.back()->opcode == aco_opcode::p_branch);
746       block->instructions.pop_back();
747 
748       Temp cond = Temp();
749       for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
750          if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
751             break;
752          cond = bld.tmp(s1);
753          Operand exec_mask = ctx.info[idx].exec[exec_idx].first;
754          exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
755                               exec_mask, Operand(exec, bld.lm));
756          ctx.info[idx].exec[exec_idx].first = exec_mask;
757       }
758       assert(cond != Temp());
759 
760       /* check if the successor is the merge block, otherwise set exec to 0 */
761       // TODO: this could be done better by directly branching to the merge block
762       unsigned succ_idx = ctx.program->blocks[block->linear_succs[1]].linear_succs[0];
763       Block& succ = ctx.program->blocks[succ_idx];
764       if (!(succ.kind & block_kind_invert || succ.kind & block_kind_merge)) {
765          bld.copy(Definition(exec, bld.lm), Operand::zero(bld.lm.bytes()));
766       }
767 
768       bld.branch(aco_opcode::p_cbranch_nz, bld.def(s2), bld.scc(cond), block->linear_succs[1],
769                  block->linear_succs[0]);
770       return;
771    }
772 }
773 
774 void
process_block(exec_ctx & ctx,Block * block)775 process_block(exec_ctx& ctx, Block* block)
776 {
777    std::vector<aco_ptr<Instruction>> instructions;
778    instructions.reserve(block->instructions.size());
779 
780    unsigned idx = add_coupling_code(ctx, block, instructions);
781 
782    assert(!block->linear_succs.empty() || ctx.info[block->index].exec.size() <= 2);
783 
784    process_instructions(ctx, block, instructions, idx);
785 
786    block->instructions = std::move(instructions);
787 
788    add_branch_code(ctx, block);
789 }
790 
791 } /* end namespace */
792 
793 void
insert_exec_mask(Program * program)794 insert_exec_mask(Program* program)
795 {
796    exec_ctx ctx(program);
797 
798    if (program->needs_wqm && program->needs_exact)
799       ctx.handle_wqm = true;
800 
801    for (Block& block : program->blocks)
802       process_block(ctx, &block);
803 }
804 
805 } // namespace aco
806