1/*
2 * Copyright © 2021 Valve Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 */
24
25#include "aco_builder.h"
26#include "aco_ir.h"
27
28#include <algorithm>
29#include <array>
30#include <bitset>
31#include <vector>
32
33namespace aco {
34namespace {
35
36constexpr const size_t max_reg_cnt = 512;
37
38struct Idx {
39   bool operator==(const Idx& other) const { return block == other.block && instr == other.instr; }
40   bool operator!=(const Idx& other) const { return !operator==(other); }
41
42   bool found() const { return block != UINT32_MAX; }
43
44   uint32_t block;
45   uint32_t instr;
46};
47
48Idx not_written_in_block{UINT32_MAX, 0};
49Idx clobbered{UINT32_MAX, 1};
50Idx const_or_undef{UINT32_MAX, 2};
51Idx written_by_multiple_instrs{UINT32_MAX, 3};
52
53struct pr_opt_ctx {
54   Program* program;
55   Block* current_block;
56   uint32_t current_instr_idx;
57   std::vector<uint16_t> uses;
58   std::vector<std::array<Idx, max_reg_cnt>> instr_idx_by_regs;
59
60   void reset_block(Block* block)
61   {
62      current_block = block;
63      current_instr_idx = 0;
64
65      if ((block->kind & block_kind_loop_header) || block->linear_preds.empty()) {
66         std::fill(instr_idx_by_regs[block->index].begin(), instr_idx_by_regs[block->index].end(),
67                   not_written_in_block);
68      } else {
69         unsigned first_pred = block->linear_preds[0];
70         for (unsigned i = 0; i < max_reg_cnt; i++) {
71            bool all_same = std::all_of(
72               std::next(block->linear_preds.begin()), block->linear_preds.end(),
73               [&](unsigned pred)
74               { return instr_idx_by_regs[pred][i] == instr_idx_by_regs[first_pred][i]; });
75
76            if (all_same)
77               instr_idx_by_regs[block->index][i] = instr_idx_by_regs[first_pred][i];
78            else
79               instr_idx_by_regs[block->index][i] = not_written_in_block;
80         }
81      }
82   }
83
84   Instruction* get(Idx idx) { return program->blocks[idx.block].instructions[idx.instr].get(); }
85};
86
87void
88save_reg_writes(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
89{
90   for (const Definition& def : instr->definitions) {
91      assert(def.regClass().type() != RegType::sgpr || def.physReg().reg() <= 255);
92      assert(def.regClass().type() != RegType::vgpr || def.physReg().reg() >= 256);
93
94      unsigned dw_size = DIV_ROUND_UP(def.bytes(), 4u);
95      unsigned r = def.physReg().reg();
96      Idx idx{ctx.current_block->index, ctx.current_instr_idx};
97
98      if (def.regClass().is_subdword())
99         idx = clobbered;
100
101      assert((r + dw_size) <= max_reg_cnt);
102      assert(def.size() == dw_size || def.regClass().is_subdword());
103      std::fill(ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r,
104                ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r + dw_size, idx);
105   }
106}
107
108Idx
109last_writer_idx(pr_opt_ctx& ctx, PhysReg physReg, RegClass rc)
110{
111   /* Verify that all of the operand's registers are written by the same instruction. */
112   assert(physReg.reg() < max_reg_cnt);
113   Idx instr_idx = ctx.instr_idx_by_regs[ctx.current_block->index][physReg.reg()];
114   unsigned dw_size = DIV_ROUND_UP(rc.bytes(), 4u);
115   unsigned r = physReg.reg();
116   bool all_same =
117      std::all_of(ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r,
118                  ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r + dw_size,
119                  [instr_idx](Idx i) { return i == instr_idx; });
120
121   return all_same ? instr_idx : written_by_multiple_instrs;
122}
123
124Idx
125last_writer_idx(pr_opt_ctx& ctx, const Operand& op)
126{
127   if (op.isConstant() || op.isUndefined())
128      return const_or_undef;
129
130   assert(op.physReg().reg() < max_reg_cnt);
131   Idx instr_idx = ctx.instr_idx_by_regs[ctx.current_block->index][op.physReg().reg()];
132
133#ifndef NDEBUG
134   /* Debug mode:  */
135   instr_idx = last_writer_idx(ctx, op.physReg(), op.regClass());
136   assert(instr_idx != written_by_multiple_instrs);
137#endif
138
139   return instr_idx;
140}
141
142bool
143is_clobbered_since(pr_opt_ctx& ctx, PhysReg reg, RegClass rc, const Idx& idx)
144{
145   /* If we didn't find an instruction, assume that the register is clobbered. */
146   if (!idx.found())
147      return true;
148
149   /* TODO: We currently can't keep track of subdword registers. */
150   if (rc.is_subdword())
151      return true;
152
153   unsigned begin_reg = reg.reg();
154   unsigned end_reg = begin_reg + rc.size();
155   unsigned current_block_idx = ctx.current_block->index;
156
157   for (unsigned r = begin_reg; r < end_reg; ++r) {
158      Idx& i = ctx.instr_idx_by_regs[current_block_idx][r];
159      if (i == clobbered || i == written_by_multiple_instrs)
160         return true;
161      else if (i == not_written_in_block)
162         continue;
163
164      assert(i.found());
165
166      if (i.block > idx.block || (i.block == idx.block && i.instr > idx.instr))
167         return true;
168   }
169
170   return false;
171}
172
173template <typename T>
174bool
175is_clobbered_since(pr_opt_ctx& ctx, const T& t, const Idx& idx)
176{
177   return is_clobbered_since(ctx, t.physReg(), t.regClass(), idx);
178}
179
180void
181try_apply_branch_vcc(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
182{
183   /* We are looking for the following pattern:
184    *
185    * vcc = ...                      ; last_vcc_wr
186    * sX, scc = s_and_bXX vcc, exec  ; op0_instr
187    * (...vcc and exec must not be clobbered inbetween...)
188    * s_cbranch_XX scc               ; instr
189    *
190    * If possible, the above is optimized into:
191    *
192    * vcc = ...                      ; last_vcc_wr
193    * s_cbranch_XX vcc               ; instr modified to use vcc
194    */
195
196   /* Don't try to optimize this on GFX6-7 because SMEM may corrupt the vccz bit. */
197   if (ctx.program->chip_class < GFX8)
198      return;
199
200   if (instr->format != Format::PSEUDO_BRANCH || instr->operands.size() == 0 ||
201       instr->operands[0].physReg() != scc)
202      return;
203
204   Idx op0_instr_idx = last_writer_idx(ctx, instr->operands[0]);
205   Idx last_vcc_wr_idx = last_writer_idx(ctx, vcc, ctx.program->lane_mask);
206
207   /* We need to make sure:
208    * - the instructions that wrote the operand register and VCC are both found
209    * - the operand register used by the branch, and VCC were both written in the current block
210    * - EXEC hasn't been clobbered since the last VCC write
211    * - VCC hasn't been clobbered since the operand register was written
212    *   (ie. the last VCC writer precedes the op0 writer)
213    */
214   if (!op0_instr_idx.found() || !last_vcc_wr_idx.found() ||
215       op0_instr_idx.block != ctx.current_block->index ||
216       last_vcc_wr_idx.block != ctx.current_block->index ||
217       is_clobbered_since(ctx, exec, ctx.program->lane_mask, last_vcc_wr_idx) ||
218       is_clobbered_since(ctx, vcc, ctx.program->lane_mask, op0_instr_idx))
219      return;
220
221   Instruction* op0_instr = ctx.get(op0_instr_idx);
222   Instruction* last_vcc_wr = ctx.get(last_vcc_wr_idx);
223
224   if ((op0_instr->opcode != aco_opcode::s_and_b64 /* wave64 */ &&
225        op0_instr->opcode != aco_opcode::s_and_b32 /* wave32 */) ||
226       op0_instr->operands[0].physReg() != vcc || op0_instr->operands[1].physReg() != exec ||
227       !last_vcc_wr->isVOPC())
228      return;
229
230   assert(last_vcc_wr->definitions[0].tempId() == op0_instr->operands[0].tempId());
231
232   /* Reduce the uses of the SCC def */
233   ctx.uses[instr->operands[0].tempId()]--;
234   /* Use VCC instead of SCC in the branch */
235   instr->operands[0] = op0_instr->operands[0];
236}
237
238void
239try_optimize_scc_nocompare(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
240{
241   /* We are looking for the following pattern:
242    *
243    * s_bfe_u32 s0, s3, 0x40018  ; outputs SGPR and SCC if the SGPR != 0
244    * s_cmp_eq_i32 s0, 0         ; comparison between the SGPR and 0
245    * s_cbranch_scc0 BB3         ; use the result of the comparison, eg. branch or cselect
246    *
247    * If possible, the above is optimized into:
248    *
249    * s_bfe_u32 s0, s3, 0x40018  ; original instruction
250    * s_cbranch_scc1 BB3         ; modified to use SCC directly rather than the SGPR with comparison
251    *
252    */
253
254   if (!instr->isSALU() && !instr->isBranch())
255      return;
256
257   if (instr->isSOPC() &&
258       (instr->opcode == aco_opcode::s_cmp_eq_u32 || instr->opcode == aco_opcode::s_cmp_eq_i32 ||
259        instr->opcode == aco_opcode::s_cmp_lg_u32 || instr->opcode == aco_opcode::s_cmp_lg_i32 ||
260        instr->opcode == aco_opcode::s_cmp_eq_u64 || instr->opcode == aco_opcode::s_cmp_lg_u64) &&
261       (instr->operands[0].constantEquals(0) || instr->operands[1].constantEquals(0)) &&
262       (instr->operands[0].isTemp() || instr->operands[1].isTemp())) {
263      /* Make sure the constant is always in operand 1 */
264      if (instr->operands[0].isConstant())
265         std::swap(instr->operands[0], instr->operands[1]);
266
267      if (ctx.uses[instr->operands[0].tempId()] > 1)
268         return;
269
270      /* Make sure both SCC and Operand 0 are written by the same instruction. */
271      Idx wr_idx = last_writer_idx(ctx, instr->operands[0]);
272      Idx sccwr_idx = last_writer_idx(ctx, scc, s1);
273      if (!wr_idx.found() || wr_idx != sccwr_idx)
274         return;
275
276      Instruction* wr_instr = ctx.get(wr_idx);
277      if (!wr_instr->isSALU() || wr_instr->definitions.size() < 2 ||
278          wr_instr->definitions[1].physReg() != scc)
279         return;
280
281      /* Look for instructions which set SCC := (D != 0) */
282      switch (wr_instr->opcode) {
283      case aco_opcode::s_bfe_i32:
284      case aco_opcode::s_bfe_i64:
285      case aco_opcode::s_bfe_u32:
286      case aco_opcode::s_bfe_u64:
287      case aco_opcode::s_and_b32:
288      case aco_opcode::s_and_b64:
289      case aco_opcode::s_andn2_b32:
290      case aco_opcode::s_andn2_b64:
291      case aco_opcode::s_or_b32:
292      case aco_opcode::s_or_b64:
293      case aco_opcode::s_orn2_b32:
294      case aco_opcode::s_orn2_b64:
295      case aco_opcode::s_xor_b32:
296      case aco_opcode::s_xor_b64:
297      case aco_opcode::s_not_b32:
298      case aco_opcode::s_not_b64:
299      case aco_opcode::s_nor_b32:
300      case aco_opcode::s_nor_b64:
301      case aco_opcode::s_xnor_b32:
302      case aco_opcode::s_xnor_b64:
303      case aco_opcode::s_nand_b32:
304      case aco_opcode::s_nand_b64:
305      case aco_opcode::s_lshl_b32:
306      case aco_opcode::s_lshl_b64:
307      case aco_opcode::s_lshr_b32:
308      case aco_opcode::s_lshr_b64:
309      case aco_opcode::s_ashr_i32:
310      case aco_opcode::s_ashr_i64:
311      case aco_opcode::s_abs_i32:
312      case aco_opcode::s_absdiff_i32: break;
313      default: return;
314      }
315
316      /* Use the SCC def from wr_instr */
317      ctx.uses[instr->operands[0].tempId()]--;
318      instr->operands[0] = Operand(wr_instr->definitions[1].getTemp(), scc);
319      ctx.uses[instr->operands[0].tempId()]++;
320
321      /* Set the opcode and operand to 32-bit */
322      instr->operands[1] = Operand::zero();
323      instr->opcode =
324         (instr->opcode == aco_opcode::s_cmp_eq_u32 || instr->opcode == aco_opcode::s_cmp_eq_i32 ||
325          instr->opcode == aco_opcode::s_cmp_eq_u64)
326            ? aco_opcode::s_cmp_eq_u32
327            : aco_opcode::s_cmp_lg_u32;
328   } else if ((instr->format == Format::PSEUDO_BRANCH && instr->operands.size() == 1 &&
329               instr->operands[0].physReg() == scc) ||
330              instr->opcode == aco_opcode::s_cselect_b32) {
331
332      /* For cselect, operand 2 is the SCC condition */
333      unsigned scc_op_idx = 0;
334      if (instr->opcode == aco_opcode::s_cselect_b32) {
335         scc_op_idx = 2;
336      }
337
338      Idx wr_idx = last_writer_idx(ctx, instr->operands[scc_op_idx]);
339      if (!wr_idx.found())
340         return;
341
342      Instruction* wr_instr = ctx.get(wr_idx);
343
344      /* Check if we found the pattern above. */
345      if (wr_instr->opcode != aco_opcode::s_cmp_eq_u32 &&
346          wr_instr->opcode != aco_opcode::s_cmp_lg_u32)
347         return;
348      if (wr_instr->operands[0].physReg() != scc)
349         return;
350      if (!wr_instr->operands[1].constantEquals(0))
351         return;
352
353      /* The optimization can be unsafe when there are other users. */
354      if (ctx.uses[instr->operands[scc_op_idx].tempId()] > 1)
355         return;
356
357      if (wr_instr->opcode == aco_opcode::s_cmp_eq_u32) {
358         /* Flip the meaning of the instruction to correctly use the SCC. */
359         if (instr->format == Format::PSEUDO_BRANCH)
360            instr->opcode = instr->opcode == aco_opcode::p_cbranch_z ? aco_opcode::p_cbranch_nz
361                                                                     : aco_opcode::p_cbranch_z;
362         else if (instr->opcode == aco_opcode::s_cselect_b32)
363            std::swap(instr->operands[0], instr->operands[1]);
364         else
365            unreachable(
366               "scc_nocompare optimization is only implemented for p_cbranch and s_cselect");
367      }
368
369      /* Use the SCC def from the original instruction, not the comparison */
370      ctx.uses[instr->operands[scc_op_idx].tempId()]--;
371      instr->operands[scc_op_idx] = wr_instr->operands[0];
372   }
373}
374
375void
376try_combine_dpp(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
377{
378   /* We are looking for the following pattern:
379    *
380    * v_mov_dpp vA, vB, ...      ; move instruction with DPP
381    * v_xxx vC, vA, ...          ; current instr that uses the result from the move
382    *
383    * If possible, the above is optimized into:
384    *
385    * v_xxx_dpp vC, vB, ...      ; current instr modified to use DPP directly
386    *
387    */
388
389   if (!instr->isVALU() || instr->isDPP() || !can_use_DPP(instr, false))
390      return;
391
392   for (unsigned i = 0; i < MIN2(2, instr->operands.size()); i++) {
393      Idx op_instr_idx = last_writer_idx(ctx, instr->operands[i]);
394      if (!op_instr_idx.found())
395         continue;
396
397      Instruction* mov = ctx.get(op_instr_idx);
398      if (mov->opcode != aco_opcode::v_mov_b32 || !mov->isDPP())
399         continue;
400
401      /* If we aren't going to remove the v_mov_b32, we have to ensure that it doesn't overwrite
402       * it's own operand before we use it.
403       */
404      if (mov->definitions[0].physReg() == mov->operands[0].physReg() &&
405          (!mov->definitions[0].tempId() || ctx.uses[mov->definitions[0].tempId()] > 1))
406         continue;
407
408      /* Don't propagate DPP if the source register is overwritten since the move. */
409      if (is_clobbered_since(ctx, mov->operands[0], op_instr_idx))
410         continue;
411
412      if (i && !can_swap_operands(instr, &instr->opcode))
413         continue;
414
415      /* anything else doesn't make sense in SSA */
416      assert(mov->dpp().row_mask == 0xf && mov->dpp().bank_mask == 0xf);
417
418      if (--ctx.uses[mov->definitions[0].tempId()])
419         ctx.uses[mov->operands[0].tempId()]++;
420
421      convert_to_DPP(instr);
422
423      DPP_instruction* dpp = &instr->dpp();
424      if (i) {
425         std::swap(dpp->operands[0], dpp->operands[1]);
426         std::swap(dpp->neg[0], dpp->neg[1]);
427         std::swap(dpp->abs[0], dpp->abs[1]);
428      }
429      dpp->operands[0] = mov->operands[0];
430      dpp->dpp_ctrl = mov->dpp().dpp_ctrl;
431      dpp->bound_ctrl = true;
432      dpp->neg[0] ^= mov->dpp().neg[0] && !dpp->abs[0];
433      dpp->abs[0] |= mov->dpp().abs[0];
434      return;
435   }
436}
437
438void
439process_instruction(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
440{
441   try_apply_branch_vcc(ctx, instr);
442
443   try_optimize_scc_nocompare(ctx, instr);
444
445   try_combine_dpp(ctx, instr);
446
447   if (instr)
448      save_reg_writes(ctx, instr);
449
450   ctx.current_instr_idx++;
451}
452
453} // namespace
454
455void
456optimize_postRA(Program* program)
457{
458   pr_opt_ctx ctx;
459   ctx.program = program;
460   ctx.uses = dead_code_analysis(program);
461   ctx.instr_idx_by_regs.resize(program->blocks.size());
462
463   /* Forward pass
464    * Goes through each instruction exactly once, and can transform
465    * instructions or adjust the use counts of temps.
466    */
467   for (auto& block : program->blocks) {
468      ctx.reset_block(&block);
469
470      for (aco_ptr<Instruction>& instr : block.instructions)
471         process_instruction(ctx, instr);
472   }
473
474   /* Cleanup pass
475    * Gets rid of instructions which are manually deleted or
476    * no longer have any uses.
477    */
478   for (auto& block : program->blocks) {
479      auto new_end = std::remove_if(block.instructions.begin(), block.instructions.end(),
480                                    [&ctx](const aco_ptr<Instruction>& instr)
481                                    { return !instr || is_dead(ctx.uses, instr.get()); });
482      block.instructions.resize(new_end - block.instructions.begin());
483   }
484}
485
486} // namespace aco
487