src/amd/compiler/aco_insert_NOPs.cpp

   1 /*
   2  * Copyright © 2019 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 <algorithm>
  26
  27 #include "aco_ir.h"
  28 #include <stack>
  29
  30 namespace aco {
  31 namespace {
  32
  33 struct NOP_ctx_gfx8_9 {
  34    enum chip_class chip_class;
  35    unsigned vcc_physical;
  36
  37    /* just initialize these with something less than max NOPs */
  38    int VALU_wrexec = -10;
  39    int VALU_wrvcc = -10;
  40    int VALU_wrsgpr = -10;
  41
  42    NOP_ctx_gfx8_9(Program* program) : chip_class(program->chip_class) {
  43       vcc_physical = program->config->num_sgprs - 2;
  44    }
  45 };
  46
  47 struct NOP_ctx_gfx10 {
  48    bool has_VOPC = false;
  49    bool has_nonVALU_exec_read = false;
  50    bool has_VMEM = false;
  51    bool has_branch_after_VMEM = false;
  52    bool has_DS = false;
  53    bool has_branch_after_DS = false;
  54    std::bitset<128> sgprs_read_by_VMEM;
  55    std::bitset<128> sgprs_read_by_SMEM;
  56
  57    void join(const NOP_ctx_gfx10 &other) {
  58       has_VOPC |= other.has_VOPC;
  59       has_nonVALU_exec_read |= other.has_nonVALU_exec_read;
  60       has_VMEM |= other.has_VMEM;
  61       has_branch_after_VMEM |= other.has_branch_after_VMEM;
  62       has_DS |= other.has_DS;
  63       has_branch_after_DS |= other.has_branch_after_DS;
  64       sgprs_read_by_VMEM |= other.sgprs_read_by_VMEM;
  65       sgprs_read_by_SMEM |= other.sgprs_read_by_SMEM;
  66    }
  67
  68    bool operator==(const NOP_ctx_gfx10 &other)
  69    {
  70       return
  71          has_VOPC == other.has_VOPC &&
  72          has_nonVALU_exec_read == other.has_nonVALU_exec_read &&
  73          has_VMEM == other.has_VMEM &&
  74          has_branch_after_VMEM == other.has_branch_after_VMEM &&
  75          has_DS == other.has_DS &&
  76          has_branch_after_DS == other.has_branch_after_DS &&
  77          sgprs_read_by_VMEM == other.sgprs_read_by_VMEM &&
  78          sgprs_read_by_SMEM == other.sgprs_read_by_SMEM;
  79    }
  80 };
  81
  82 template <std::size_t N>
  83 bool check_written_regs(const aco_ptr<Instruction> &instr, const std::bitset<N> &check_regs)
  84 {
  85    return std::any_of(instr->definitions.begin(), instr->definitions.end(), [&check_regs](const Definition &def) -> bool {
  86       bool writes_any = false;
  87       for (unsigned i = 0; i < def.size(); i++) {
  88          unsigned def_reg = def.physReg() + i;
  89          writes_any |= def_reg < check_regs.size() && check_regs[def_reg];
  90       }
  91       return writes_any;
  92    });
  93 }
  94
  95 template <std::size_t N>
  96 void mark_read_regs(const aco_ptr<Instruction> &instr, std::bitset<N> &reg_reads)
  97 {
  98    for (const Operand &op : instr->operands) {
  99       for (unsigned i = 0; i < op.size(); i++) {
 100          unsigned reg = op.physReg() + i;
 101          if (reg < reg_reads.size())
 102             reg_reads.set(reg);
 103       }
 104    }
 105 }
 106
 107 bool VALU_writes_sgpr(aco_ptr<Instruction>& instr)
 108 {
 109    if ((uint32_t) instr->format & (uint32_t) Format::VOPC)
 110       return true;
 111    if (instr->isVOP3() && instr->definitions.size() == 2)
 112       return true;
 113    if (instr->opcode == aco_opcode::v_readfirstlane_b32 || instr->opcode == aco_opcode::v_readlane_b32)
 114       return true;
 115    return false;
 116 }
 117
 118 bool instr_writes_exec(const aco_ptr<Instruction>& instr)
 119 {
 120    return std::any_of(instr->definitions.begin(), instr->definitions.end(), [](const Definition &def) -> bool {
 121       return def.physReg() == exec_lo || def.physReg() == exec_hi;
 122    });
 123 }
 124
 125 bool instr_writes_sgpr(const aco_ptr<Instruction>& instr)
 126 {
 127    return std::any_of(instr->definitions.begin(), instr->definitions.end(), [](const Definition &def) -> bool {
 128       return def.getTemp().type() == RegType::sgpr;
 129    });
 130 }
 131
 132 inline bool instr_is_branch(const aco_ptr<Instruction>& instr)
 133 {
 134    return instr->opcode == aco_opcode::s_branch ||
 135           instr->opcode == aco_opcode::s_cbranch_scc0 ||
 136           instr->opcode == aco_opcode::s_cbranch_scc1 ||
 137           instr->opcode == aco_opcode::s_cbranch_vccz ||
 138           instr->opcode == aco_opcode::s_cbranch_vccnz ||
 139           instr->opcode == aco_opcode::s_cbranch_execz ||
 140           instr->opcode == aco_opcode::s_cbranch_execnz ||
 141           instr->opcode == aco_opcode::s_cbranch_cdbgsys ||
 142           instr->opcode == aco_opcode::s_cbranch_cdbguser ||
 143           instr->opcode == aco_opcode::s_cbranch_cdbgsys_or_user ||
 144           instr->opcode == aco_opcode::s_cbranch_cdbgsys_and_user ||
 145           instr->opcode == aco_opcode::s_subvector_loop_begin ||
 146           instr->opcode == aco_opcode::s_subvector_loop_end ||
 147           instr->opcode == aco_opcode::s_setpc_b64 ||
 148           instr->opcode == aco_opcode::s_swappc_b64 ||
 149           instr->opcode == aco_opcode::s_getpc_b64 ||
 150           instr->opcode == aco_opcode::s_call_b64;
 151 }
 152
 153 bool regs_intersect(PhysReg a_reg, unsigned a_size, PhysReg b_reg, unsigned b_size)
 154 {
 155    return a_reg > b_reg ?
 156           (a_reg - b_reg < b_size) :
 157           (b_reg - a_reg < a_size);
 158 }
 159
 160 unsigned handle_SMEM_clause(aco_ptr<Instruction>& instr, int new_idx,
 161                             std::vector<aco_ptr<Instruction>>& new_instructions)
 162 {
 163    //TODO: s_dcache_inv needs to be in it's own group on GFX10 (and previous versions?)
 164    const bool is_store = instr->definitions.empty();
 165    for (int pred_idx = new_idx - 1; pred_idx >= 0; pred_idx--) {
 166       aco_ptr<Instruction>& pred = new_instructions[pred_idx];
 167       if (pred->format != Format::SMEM)
 168          break;
 169
 170       /* Don't allow clauses with store instructions since the clause's
 171        * instructions may use the same address. */
 172       if (is_store || pred->definitions.empty())
 173          return 1;
 174
 175       Definition& instr_def = instr->definitions[0];
 176       Definition& pred_def = pred->definitions[0];
 177
 178       /* ISA reference doesn't say anything about this, but best to be safe */
 179       if (regs_intersect(instr_def.physReg(), instr_def.size(), pred_def.physReg(), pred_def.size()))
 180          return 1;
 181
 182       for (const Operand& op : pred->operands) {
 183          if (op.isConstant() || !op.isFixed())
 184             continue;
 185          if (regs_intersect(instr_def.physReg(), instr_def.size(), op.physReg(), op.size()))
 186             return 1;
 187       }
 188       for (const Operand& op : instr->operands) {
 189          if (op.isConstant() || !op.isFixed())
 190             continue;
 191          if (regs_intersect(pred_def.physReg(), pred_def.size(), op.physReg(), op.size()))
 192             return 1;
 193       }
 194    }
 195
 196    return 0;
 197 }
 198
 199 int handle_instruction_gfx8_9(NOP_ctx_gfx8_9& ctx, aco_ptr<Instruction>& instr,
 200                               std::vector<aco_ptr<Instruction>>& old_instructions,
 201                               std::vector<aco_ptr<Instruction>>& new_instructions)
 202 {
 203    int new_idx = new_instructions.size();
 204
 205    // TODO: setreg / getreg / m0 writes
 206    // TODO: try to schedule the NOP-causing instruction up to reduce the number of stall cycles
 207
 208    /* break off from prevous SMEM clause if needed */
 209    if (instr->format == Format::SMEM && ctx.chip_class >= GFX8) {
 210       return handle_SMEM_clause(instr, new_idx, new_instructions);
 211    } else if (instr->isVALU() || instr->format == Format::VINTRP) {
 212       int NOPs = 0;
 213
 214       if (instr->isDPP()) {
 215          /* VALU does not forward EXEC to DPP. */
 216          if (ctx.VALU_wrexec + 5 >= new_idx)
 217             NOPs = 5 + ctx.VALU_wrexec - new_idx + 1;
 218
 219          /* VALU DPP reads VGPR written by VALU */
 220          for (int pred_idx = new_idx - 1; pred_idx >= 0 && pred_idx >= new_idx - 2; pred_idx--) {
 221             aco_ptr<Instruction>& pred = new_instructions[pred_idx];
 222             if ((pred->isVALU() || pred->format == Format::VINTRP) &&
 223                 !pred->definitions.empty() &&
 224                 pred->definitions[0].physReg() == instr->operands[0].physReg()) {
 225                NOPs = std::max(NOPs, 2 + pred_idx - new_idx + 1);
 226                break;
 227             }
 228          }
 229       }
 230
 231       /* SALU writes M0 */
 232       if (instr->format == Format::VINTRP && new_idx > 0 && ctx.chip_class >= GFX9) {
 233          aco_ptr<Instruction>& pred = new_instructions.back();
 234          if (pred->isSALU() &&
 235              !pred->definitions.empty() &&
 236              pred->definitions[0].physReg() == m0)
 237             NOPs = std::max(NOPs, 1);
 238       }
 239
 240       for (const Operand& op : instr->operands) {
 241          /* VALU which uses VCCZ */
 242          if (op.physReg() == PhysReg{251} &&
 243              ctx.VALU_wrvcc + 5 >= new_idx)
 244             NOPs = std::max(NOPs, 5 + ctx.VALU_wrvcc - new_idx + 1);
 245
 246          /* VALU which uses EXECZ */
 247          if (op.physReg() == PhysReg{252} &&
 248              ctx.VALU_wrexec + 5 >= new_idx)
 249             NOPs = std::max(NOPs, 5 + ctx.VALU_wrexec - new_idx + 1);
 250
 251          /* VALU which reads VCC as a constant */
 252          if (ctx.VALU_wrvcc + 1 >= new_idx) {
 253             for (unsigned k = 0; k < op.size(); k++) {
 254                unsigned reg = op.physReg() + k;
 255                if (reg == ctx.vcc_physical || reg == ctx.vcc_physical + 1)
 256                   NOPs = std::max(NOPs, 1);
 257             }
 258          }
 259       }
 260
 261       switch (instr->opcode) {
 262          case aco_opcode::v_readlane_b32:
 263          case aco_opcode::v_writelane_b32: {
 264             if (ctx.VALU_wrsgpr + 4 < new_idx)
 265                break;
 266             PhysReg reg = instr->operands[1].physReg();
 267             for (int pred_idx = new_idx - 1; pred_idx >= 0 && pred_idx >= new_idx - 4; pred_idx--) {
 268                aco_ptr<Instruction>& pred = new_instructions[pred_idx];
 269                if (!pred->isVALU() || !VALU_writes_sgpr(pred))
 270                   continue;
 271                for (const Definition& def : pred->definitions) {
 272                   if (def.physReg() == reg)
 273                      NOPs = std::max(NOPs, 4 + pred_idx - new_idx + 1);
 274                }
 275             }
 276             break;
 277          }
 278          case aco_opcode::v_div_fmas_f32:
 279          case aco_opcode::v_div_fmas_f64: {
 280             if (ctx.VALU_wrvcc + 4 >= new_idx)
 281                NOPs = std::max(NOPs, 4 + ctx.VALU_wrvcc - new_idx + 1);
 282             break;
 283          }
 284          default:
 285             break;
 286       }
 287
 288       /* Write VGPRs holding writedata > 64 bit from MIMG/MUBUF instructions */
 289       // FIXME: handle case if the last instruction of a block without branch is such store
 290       if (new_idx > 0) {
 291          aco_ptr<Instruction>& pred = new_instructions.back();
 292          /* >64-bit MUBUF/MTBUF store with a constant in SOFFSET */
 293          bool consider_buf = (pred->format == Format::MUBUF || pred->format == Format::MTBUF) &&
 294                              pred->operands.size() == 4 &&
 295                              pred->operands[3].size() > 2 &&
 296                              pred->operands[2].physReg() >= 128;
 297          /* MIMG store with a 128-bit T# with more than two bits set in dmask (making it a >64-bit store) */
 298          bool consider_mimg = pred->format == Format::MIMG &&
 299                               pred->operands.size() == 4 &&
 300                               pred->operands[3].size() > 2 &&
 301                               pred->operands[1].size() != 8;
 302          /* FLAT/GLOBAL/SCRATCH store with >64-bit data */
 303          bool consider_flat = (pred->isFlatOrGlobal() || pred->format == Format::SCRATCH) &&
 304                               pred->operands.size() == 3 &&
 305                               pred->operands[2].size() > 2;
 306          if (consider_buf || consider_mimg || consider_flat) {
 307             PhysReg wrdata = pred->operands[3].physReg();
 308             unsigned size = pred->operands[3].size();
 309             assert(wrdata >= 256);
 310             for (const Definition& def : instr->definitions) {
 311                if (regs_intersect(def.physReg(), def.size(), wrdata, size))
 312                   NOPs = std::max(NOPs, 1);
 313             }
 314          }
 315       }
 316
 317       if (VALU_writes_sgpr(instr)) {
 318          for (const Definition& def : instr->definitions) {
 319             if (def.physReg() == vcc)
 320                ctx.VALU_wrvcc = NOPs ? new_idx : new_idx + 1;
 321             else if (def.physReg() == exec)
 322                ctx.VALU_wrexec = NOPs ? new_idx : new_idx + 1;
 323             else if (def.physReg() <= 102)
 324                ctx.VALU_wrsgpr = NOPs ? new_idx : new_idx + 1;
 325          }
 326       }
 327       return NOPs;
 328    } else if (instr->isVMEM() && ctx.VALU_wrsgpr + 5 >= new_idx) {
 329       /* If the VALU writes the SGPR that is used by a VMEM, the user must add five wait states. */
 330       for (int pred_idx = new_idx - 1; pred_idx >= 0 && pred_idx >= new_idx - 5; pred_idx--) {
 331          aco_ptr<Instruction>& pred = new_instructions[pred_idx];
 332          if (!(pred->isVALU() && VALU_writes_sgpr(pred)))
 333             continue;
 334
 335          for (const Definition& def : pred->definitions) {
 336             if (def.physReg() > 102)
 337                continue;
 338
 339             if (instr->operands.size() > 1 &&
 340                 regs_intersect(instr->operands[1].physReg(), instr->operands[1].size(),
 341                                def.physReg(), def.size())) {
 342                   return 5 + pred_idx - new_idx + 1;
 343             }
 344
 345             if (instr->operands.size() > 2 &&
 346                 regs_intersect(instr->operands[2].physReg(), instr->operands[2].size(),
 347                                def.physReg(), def.size())) {
 348                   return 5 + pred_idx - new_idx + 1;
 349             }
 350          }
 351       }
 352    }
 353
 354    return 0;
 355 }
 356
 357 void handle_block_gfx8_9(NOP_ctx_gfx8_9& ctx, Block& block)
 358 {
 359    std::vector<aco_ptr<Instruction>> instructions;
 360    instructions.reserve(block.instructions.size());
 361    for (unsigned i = 0; i < block.instructions.size(); i++) {
 362       aco_ptr<Instruction>& instr = block.instructions[i];
 363       unsigned NOPs = handle_instruction_gfx8_9(ctx, instr, block.instructions, instructions);
 364       if (NOPs) {
 365          // TODO: try to move the instruction down
 366          /* create NOP */
 367          aco_ptr<SOPP_instruction> nop{create_instruction<SOPP_instruction>(aco_opcode::s_nop, Format::SOPP, 0, 0)};
 368          nop->imm = NOPs - 1;
 369          nop->block = -1;
 370          instructions.emplace_back(std::move(nop));
 371       }
 372
 373       instructions.emplace_back(std::move(instr));
 374    }
 375
 376    ctx.VALU_wrvcc -= instructions.size();
 377    ctx.VALU_wrexec -= instructions.size();
 378    ctx.VALU_wrsgpr -= instructions.size();
 379    block.instructions = std::move(instructions);
 380 }
 381
 382 void insert_NOPs_gfx8_9(Program* program)
 383 {
 384    NOP_ctx_gfx8_9 ctx(program);
 385
 386    for (Block& block : program->blocks) {
 387       if (block.instructions.empty())
 388          continue;
 389
 390       handle_block_gfx8_9(ctx, block);
 391    }
 392 }
 393
 394 void handle_instruction_gfx10(Program *program, NOP_ctx_gfx10 &ctx, aco_ptr<Instruction>& instr,
 395                               std::vector<aco_ptr<Instruction>>& old_instructions,
 396                               std::vector<aco_ptr<Instruction>>& new_instructions)
 397 {
 398    /* VMEMtoScalarWriteHazard
 399     * Handle EXEC/M0/SGPR write following a VMEM instruction without a VALU or "waitcnt vmcnt(0)" in-between.
 400     */
 401    if (instr->isVMEM() || instr->format == Format::FLAT || instr->format == Format::GLOBAL ||
 402        instr->format == Format::SCRATCH || instr->format == Format::DS) {
 403       /* Remember all SGPRs that are read by the VMEM instruction */
 404       mark_read_regs(instr, ctx.sgprs_read_by_VMEM);
 405       ctx.sgprs_read_by_VMEM.set(exec);
 406       if (program->wave_size == 64)
 407          ctx.sgprs_read_by_VMEM.set(exec_hi);
 408    } else if (instr->isSALU() || instr->format == Format::SMEM) {
 409       /* Check if SALU writes an SGPR that was previously read by the VALU */
 410       if (check_written_regs(instr, ctx.sgprs_read_by_VMEM)) {
 411          ctx.sgprs_read_by_VMEM.reset();
 412
 413          /* Insert v_nop to mitigate the problem */
 414          aco_ptr<VOP1_instruction> nop{create_instruction<VOP1_instruction>(aco_opcode::v_nop, Format::VOP1, 0, 0)};
 415          new_instructions.emplace_back(std::move(nop));
 416       }
 417    } else if (instr->opcode == aco_opcode::s_waitcnt) {
 418       /* Hazard is mitigated by "s_waitcnt vmcnt(0)" */
 419       uint16_t imm = static_cast<SOPP_instruction*>(instr.get())->imm;
 420       unsigned vmcnt = (imm & 0xF) | ((imm & (0x3 << 14)) >> 10);
 421       if (vmcnt == 0)
 422          ctx.sgprs_read_by_VMEM.reset();
 423    } else if (instr->isVALU()) {
 424       /* Hazard is mitigated by any VALU instruction */
 425       ctx.sgprs_read_by_VMEM.reset();
 426    }
 427
 428    /* VcmpxPermlaneHazard
 429     * Handle any permlane following a VOPC instruction, insert v_mov between them.
 430     */
 431    if (instr->format == Format::VOPC) {
 432       ctx.has_VOPC = true;
 433    } else if (ctx.has_VOPC &&
 434               (instr->opcode == aco_opcode::v_permlane16_b32 ||
 435                instr->opcode == aco_opcode::v_permlanex16_b32)) {
 436       ctx.has_VOPC = false;
 437
 438       /* v_nop would be discarded by SQ, so use v_mov with the first operand of the permlane */
 439       aco_ptr<VOP1_instruction> v_mov{create_instruction<VOP1_instruction>(aco_opcode::v_mov_b32, Format::VOP1, 1, 1)};
 440       v_mov->definitions[0] = Definition(instr->operands[0].physReg(), v1);
 441       v_mov->operands[0] = Operand(instr->operands[0].physReg(), v1);
 442       new_instructions.emplace_back(std::move(v_mov));
 443    } else if (instr->isVALU() && instr->opcode != aco_opcode::v_nop) {
 444       ctx.has_VOPC = false;
 445    }
 446
 447    /* VcmpxExecWARHazard
 448     * Handle any VALU instruction writing the exec mask after it was read by a non-VALU instruction.
 449     */
 450    if (!instr->isVALU() && instr->reads_exec()) {
 451       ctx.has_nonVALU_exec_read = true;
 452    } else if (instr->isVALU()) {
 453       if (instr_writes_exec(instr)) {
 454          ctx.has_nonVALU_exec_read = false;
 455
 456          /* Insert s_waitcnt_depctr instruction with magic imm to mitigate the problem */
 457          aco_ptr<SOPP_instruction> depctr{create_instruction<SOPP_instruction>(aco_opcode::s_waitcnt_depctr, Format::SOPP, 0, 1)};
 458          depctr->imm = 0xfffe;
 459          depctr->definitions[0] = Definition(sgpr_null, s1);
 460          new_instructions.emplace_back(std::move(depctr));
 461       } else if (instr_writes_sgpr(instr)) {
 462          /* Any VALU instruction that writes an SGPR mitigates the problem */
 463          ctx.has_nonVALU_exec_read = false;
 464       }
 465    } else if (instr->opcode == aco_opcode::s_waitcnt_depctr) {
 466       /* s_waitcnt_depctr can mitigate the problem if it has a magic imm */
 467       const SOPP_instruction *sopp = static_cast<const SOPP_instruction *>(instr.get());
 468       if ((sopp->imm & 0xfffe) == 0xfffe)
 469          ctx.has_nonVALU_exec_read = false;
 470    }
 471
 472    /* SMEMtoVectorWriteHazard
 473     * Handle any VALU instruction writing an SGPR after an SMEM reads it.
 474     */
 475    if (instr->format == Format::SMEM) {
 476       /* Remember all SGPRs that are read by the SMEM instruction */
 477       mark_read_regs(instr, ctx.sgprs_read_by_SMEM);
 478    } else if (VALU_writes_sgpr(instr)) {
 479       /* Check if VALU writes an SGPR that was previously read by SMEM */
 480       if (check_written_regs(instr, ctx.sgprs_read_by_SMEM)) {
 481          ctx.sgprs_read_by_SMEM.reset();
 482
 483          /* Insert s_mov to mitigate the problem */
 484          aco_ptr<SOP1_instruction> s_mov{create_instruction<SOP1_instruction>(aco_opcode::s_mov_b32, Format::SOP1, 1, 1)};
 485          s_mov->definitions[0] = Definition(sgpr_null, s1);
 486          s_mov->operands[0] = Operand(0u);
 487          new_instructions.emplace_back(std::move(s_mov));
 488       }
 489    } else if (instr->isSALU()) {
 490       if (instr->format != Format::SOPP) {
 491          /* SALU can mitigate the hazard */
 492          ctx.sgprs_read_by_SMEM.reset();
 493       } else {
 494          /* Reducing lgkmcnt count to 0 always mitigates the hazard. */
 495          const SOPP_instruction *sopp = static_cast<const SOPP_instruction *>(instr.get());
 496          if (sopp->opcode == aco_opcode::s_waitcnt_lgkmcnt) {
 497             if (sopp->imm == 0 && sopp->definitions[0].physReg() == sgpr_null)
 498                ctx.sgprs_read_by_SMEM.reset();
 499          } else if (sopp->opcode == aco_opcode::s_waitcnt) {
 500             unsigned lgkm = (sopp->imm >> 8) & 0x3f;
 501             if (lgkm == 0)
 502                ctx.sgprs_read_by_SMEM.reset();
 503          }
 504       }
 505    }
 506
 507    /* LdsBranchVmemWARHazard
 508     * Handle VMEM/GLOBAL/SCRATCH->branch->DS and DS->branch->VMEM/GLOBAL/SCRATCH patterns.
 509     */
 510    if (instr->isVMEM() || instr->format == Format::GLOBAL || instr->format == Format::SCRATCH) {
 511       ctx.has_VMEM = true;
 512       ctx.has_branch_after_VMEM = false;
 513       /* Mitigation for DS is needed only if there was already a branch after */
 514       ctx.has_DS = ctx.has_branch_after_DS;
 515    } else if (instr->format == Format::DS) {
 516       ctx.has_DS = true;
 517       ctx.has_branch_after_DS = false;
 518       /* Mitigation for VMEM is needed only if there was already a branch after */
 519       ctx.has_VMEM = ctx.has_branch_after_VMEM;
 520    } else if (instr_is_branch(instr)) {
 521       ctx.has_branch_after_VMEM = ctx.has_VMEM;
 522       ctx.has_branch_after_DS = ctx.has_DS;
 523    } else if (instr->opcode == aco_opcode::s_waitcnt_vscnt) {
 524       /* Only s_waitcnt_vscnt can mitigate the hazard */
 525       const SOPK_instruction *sopk = static_cast<const SOPK_instruction *>(instr.get());
 526       if (sopk->definitions[0].physReg() == sgpr_null && sopk->imm == 0)
 527          ctx.has_VMEM = ctx.has_branch_after_VMEM = ctx.has_DS = ctx.has_branch_after_DS = false;
 528    }
 529    if ((ctx.has_VMEM && ctx.has_branch_after_DS) || (ctx.has_DS && ctx.has_branch_after_VMEM)) {
 530       ctx.has_VMEM = ctx.has_branch_after_VMEM = ctx.has_DS = ctx.has_branch_after_DS = false;
 531
 532       /* Insert s_waitcnt_vscnt to mitigate the problem */
 533       aco_ptr<SOPK_instruction> wait{create_instruction<SOPK_instruction>(aco_opcode::s_waitcnt_vscnt, Format::SOPK, 0, 1)};
 534       wait->definitions[0] = Definition(sgpr_null, s1);
 535       wait->imm = 0;
 536       new_instructions.emplace_back(std::move(wait));
 537    }
 538 }
 539
 540 void handle_block_gfx10(Program *program, NOP_ctx_gfx10& ctx, Block& block)
 541 {
 542    if (block.instructions.empty())
 543       return;
 544
 545    std::vector<aco_ptr<Instruction>> instructions;
 546    instructions.reserve(block.instructions.size());
 547
 548    for (aco_ptr<Instruction>& instr : block.instructions) {
 549       handle_instruction_gfx10(program, ctx, instr, block.instructions, instructions);
 550       instructions.emplace_back(std::move(instr));
 551    }
 552
 553    block.instructions = std::move(instructions);
 554 }
 555
 556 void mitigate_hazards_gfx10(Program *program)
 557 {
 558    NOP_ctx_gfx10 all_ctx[program->blocks.size()];
 559    std::stack<unsigned> loop_header_indices;
 560
 561    for (unsigned i = 0; i < program->blocks.size(); i++) {
 562       Block& block = program->blocks[i];
 563       NOP_ctx_gfx10 &ctx = all_ctx[i];
 564
 565       if (block.kind & block_kind_loop_header) {
 566          loop_header_indices.push(i);
 567       } else if (block.kind & block_kind_loop_exit) {
 568          /* Go through the whole loop again */
 569          for (unsigned idx = loop_header_indices.top(); idx < i; idx++) {
 570             NOP_ctx_gfx10 loop_block_ctx;
 571             for (unsigned b : program->blocks[idx].linear_preds)
 572                loop_block_ctx.join(all_ctx[b]);
 573
 574             handle_block_gfx10(program, loop_block_ctx, program->blocks[idx]);
 575
 576             /* We only need to continue if the loop header context changed */
 577             if (idx == loop_header_indices.top() && loop_block_ctx == all_ctx[idx])
 578                break;
 579
 580             all_ctx[idx] = loop_block_ctx;
 581          }
 582
 583          loop_header_indices.pop();
 584       }
 585
 586       for (unsigned b : block.linear_preds)
 587          ctx.join(all_ctx[b]);
 588
 589       handle_block_gfx10(program, ctx, block);
 590    }
 591 }
 592
 593 } /* end namespace */
 594
 595 void insert_NOPs(Program* program)
 596 {
 597    if (program->chip_class >= GFX10)
 598       mitigate_hazards_gfx10(program);
 599    else
 600       insert_NOPs_gfx8_9(program);
 601 }
 602
 603 }