src/amd/compiler/aco_ssa_elimination.cpp

   1 /*
   2  * Copyright © 2018 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
  26 #include "aco_ir.h"
  27
  28 #include <map>
  29
  30 namespace aco {
  31 namespace {
  32
  33 /* map: block-id -> pair (dest, src) to store phi information */
  34 typedef std::map<uint32_t, std::vector<std::pair<Definition, Operand>>> phi_info;
  35
  36 struct ssa_elimination_ctx {
  37    phi_info logical_phi_info;
  38    phi_info linear_phi_info;
  39    std::vector<bool> empty_blocks;
  40    Program* program;
  41
  42    ssa_elimination_ctx(Program* program) : empty_blocks(program->blocks.size(), true), program(program) {}
  43 };
  44
  45 void collect_phi_info(ssa_elimination_ctx& ctx)
  46 {
  47    for (Block& block : ctx.program->blocks) {
  48       for (aco_ptr<Instruction>& phi : block.instructions) {
  49          if (phi->opcode != aco_opcode::p_phi && phi->opcode != aco_opcode::p_linear_phi)
  50             break;
  51
  52          for (unsigned i = 0; i < phi->operands.size(); i++) {
  53             if (phi->operands[i].isUndefined())
  54                continue;
  55             if (phi->operands[i].isTemp() && phi->operands[i].physReg() == phi->definitions[0].physReg())
  56                continue;
  57
  58             std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ? block.logical_preds : block.linear_preds;
  59             phi_info& info = phi->opcode == aco_opcode::p_phi ? ctx.logical_phi_info : ctx.linear_phi_info;
  60             const auto result = info.emplace(preds[i], std::vector<std::pair<Definition, Operand>>());
  61             assert(phi->definitions[0].size() == phi->operands[i].size());
  62             result.first->second.emplace_back(phi->definitions[0], phi->operands[i]);
  63             ctx.empty_blocks[preds[i]] = false;
  64          }
  65       }
  66    }
  67 }
  68
  69 void insert_parallelcopies(ssa_elimination_ctx& ctx)
  70 {
  71    /* insert the parallelcopies from logical phis before p_logical_end */
  72    for (auto&& entry : ctx.logical_phi_info) {
  73       Block& block = ctx.program->blocks[entry.first];
  74       unsigned idx = block.instructions.size() - 1;
  75       while (block.instructions[idx]->opcode != aco_opcode::p_logical_end) {
  76          assert(idx > 0);
  77          idx--;
  78       }
  79
  80       std::vector<aco_ptr<Instruction>>::iterator it = std::next(block.instructions.begin(), idx);
  81       aco_ptr<Pseudo_instruction> pc{create_instruction<Pseudo_instruction>(aco_opcode::p_parallelcopy, Format::PSEUDO, entry.second.size(), entry.second.size())};
  82       unsigned i = 0;
  83       for (std::pair<Definition, Operand>& pair : entry.second)
  84       {
  85          pc->definitions[i] = pair.first;
  86          pc->operands[i] = pair.second;
  87          i++;
  88       }
  89       /* this shouldn't be needed since we're only copying vgprs */
  90       pc->tmp_in_scc = false;
  91       block.instructions.insert(it, std::move(pc));
  92    }
  93
  94    /* insert parallelcopies for the linear phis at the end of blocks just before the branch */
  95    for (auto&& entry : ctx.linear_phi_info) {
  96       Block& block = ctx.program->blocks[entry.first];
  97       std::vector<aco_ptr<Instruction>>::iterator it = block.instructions.end();
  98       --it;
  99       assert((*it)->format == Format::PSEUDO_BRANCH);
 100       aco_ptr<Pseudo_instruction> pc{create_instruction<Pseudo_instruction>(aco_opcode::p_parallelcopy, Format::PSEUDO, entry.second.size(), entry.second.size())};
 101       unsigned i = 0;
 102       for (std::pair<Definition, Operand>& pair : entry.second)
 103       {
 104          pc->definitions[i] = pair.first;
 105          pc->operands[i] = pair.second;
 106          i++;
 107       }
 108       pc->tmp_in_scc = block.scc_live_out;
 109       pc->scratch_sgpr = block.scratch_sgpr;
 110       block.instructions.insert(it, std::move(pc));
 111    }
 112 }
 113
 114
 115 void try_remove_merge_block(ssa_elimination_ctx& ctx, Block* block)
 116 {
 117    /* check if the successor is another merge block which restores exec */
 118    // TODO: divergent loops also restore exec
 119    if (block->linear_succs.size() != 1 ||
 120        !(ctx.program->blocks[block->linear_succs[0]].kind & block_kind_merge))
 121       return;
 122
 123    /* check if this block is empty and the exec mask is not needed */
 124    for (aco_ptr<Instruction>& instr : block->instructions) {
 125       if (instr->opcode == aco_opcode::p_parallelcopy) {
 126          if (instr->definitions[0].physReg() == exec)
 127             continue;
 128          else
 129             return;
 130       }
 131
 132       if (instr->opcode != aco_opcode::p_linear_phi &&
 133           instr->opcode != aco_opcode::p_phi &&
 134           instr->opcode != aco_opcode::p_logical_start &&
 135           instr->opcode != aco_opcode::p_logical_end &&
 136           instr->opcode != aco_opcode::p_branch)
 137          return;
 138    }
 139
 140    /* keep the branch instruction and remove the rest */
 141    aco_ptr<Instruction> branch = std::move(block->instructions.back());
 142    block->instructions.clear();
 143    block->instructions.emplace_back(std::move(branch));
 144 }
 145
 146 void try_remove_invert_block(ssa_elimination_ctx& ctx, Block* block)
 147 {
 148    assert(block->linear_succs.size() == 2);
 149    if (block->linear_succs[0] != block->linear_succs[1])
 150       return;
 151
 152    /* check if we can remove this block */
 153    for (aco_ptr<Instruction>& instr : block->instructions) {
 154       if (instr->opcode != aco_opcode::p_linear_phi &&
 155           instr->opcode != aco_opcode::p_phi &&
 156           (instr->opcode != aco_opcode::s_andn2_b64 || ctx.program->wave_size != 64) &&
 157           (instr->opcode != aco_opcode::s_andn2_b32 || ctx.program->wave_size != 32) &&
 158           instr->opcode != aco_opcode::p_branch)
 159          return;
 160    }
 161
 162    unsigned succ_idx = block->linear_succs[0];
 163    assert(block->linear_preds.size() == 2);
 164    for (unsigned i = 0; i < 2; i++) {
 165       Block *pred = &ctx.program->blocks[block->linear_preds[i]];
 166       pred->linear_succs[0] = succ_idx;
 167       ctx.program->blocks[succ_idx].linear_preds[i] = pred->index;
 168
 169       Pseudo_branch_instruction *branch = static_cast<Pseudo_branch_instruction*>(pred->instructions.back().get());
 170       assert(branch->format == Format::PSEUDO_BRANCH);
 171       branch->target[0] = succ_idx;
 172       branch->target[1] = succ_idx;
 173    }
 174
 175    block->instructions.clear();
 176    block->linear_preds.clear();
 177    block->linear_succs.clear();
 178 }
 179
 180 void try_remove_simple_block(ssa_elimination_ctx& ctx, Block* block)
 181 {
 182    for (aco_ptr<Instruction>& instr : block->instructions) {
 183       if (instr->opcode != aco_opcode::p_logical_start &&
 184           instr->opcode != aco_opcode::p_logical_end &&
 185           instr->opcode != aco_opcode::p_branch)
 186          return;
 187    }
 188
 189    Block& pred = ctx.program->blocks[block->linear_preds[0]];
 190    Block& succ = ctx.program->blocks[block->linear_succs[0]];
 191    Pseudo_branch_instruction* branch = static_cast<Pseudo_branch_instruction*>(pred.instructions.back().get());
 192    if (branch->opcode == aco_opcode::p_branch) {
 193       branch->target[0] = succ.index;
 194       branch->target[1] = succ.index;
 195    } else if (branch->target[0] == block->index) {
 196       branch->target[0] = succ.index;
 197    } else if (branch->target[0] == succ.index) {
 198       assert(branch->target[1] == block->index);
 199       branch->target[1] = succ.index;
 200       branch->opcode = aco_opcode::p_branch;
 201    } else if (branch->target[1] == block->index) {
 202       /* check if there is a fall-through path from block to succ */
 203       bool falls_through = true;
 204       for (unsigned j = block->index + 1; falls_through && j < succ.index; j++) {
 205          assert(ctx.program->blocks[j].index == j);
 206          if (!ctx.program->blocks[j].instructions.empty())
 207             falls_through = false;
 208       }
 209       if (falls_through) {
 210          branch->target[1] = succ.index;
 211       } else {
 212          /* check if there is a fall-through path for the alternative target */
 213          for (unsigned j = block->index + 1; j < branch->target[0]; j++) {
 214             if (!ctx.program->blocks[j].instructions.empty())
 215                return;
 216          }
 217
 218          /* This is a (uniform) break or continue block. The branch condition has to be inverted. */
 219          if (branch->opcode == aco_opcode::p_cbranch_z)
 220             branch->opcode = aco_opcode::p_cbranch_nz;
 221          else if (branch->opcode == aco_opcode::p_cbranch_nz)
 222             branch->opcode = aco_opcode::p_cbranch_z;
 223          else
 224             assert(false);
 225          /* also invert the linear successors */
 226          pred.linear_succs[0] = pred.linear_succs[1];
 227          pred.linear_succs[1] = succ.index;
 228          branch->target[1] = branch->target[0];
 229          branch->target[0] = succ.index;
 230       }
 231    } else {
 232       assert(false);
 233    }
 234
 235    if (branch->target[0] == branch->target[1])
 236       branch->opcode = aco_opcode::p_branch;
 237
 238    for (unsigned i = 0; i < pred.linear_succs.size(); i++)
 239       if (pred.linear_succs[i] == block->index)
 240          pred.linear_succs[i] = succ.index;
 241
 242    for (unsigned i = 0; i < succ.linear_preds.size(); i++)
 243       if (succ.linear_preds[i] == block->index)
 244          succ.linear_preds[i] = pred.index;
 245
 246    block->instructions.clear();
 247    block->linear_preds.clear();
 248    block->linear_succs.clear();
 249 }
 250
 251 void jump_threading(ssa_elimination_ctx& ctx)
 252 {
 253    for (int i = ctx.program->blocks.size() - 1; i >= 0; i--) {
 254       Block* block = &ctx.program->blocks[i];
 255
 256       if (!ctx.empty_blocks[i])
 257          continue;
 258
 259       if (block->kind & block_kind_invert) {
 260          try_remove_invert_block(ctx, block);
 261          continue;
 262       }
 263
 264       if (block->linear_succs.size() > 1)
 265          continue;
 266
 267       if (block->kind & block_kind_merge ||
 268           block->kind & block_kind_loop_exit)
 269          try_remove_merge_block(ctx, block);
 270
 271       if (block->linear_preds.size() == 1)
 272          try_remove_simple_block(ctx, block);
 273    }
 274 }
 275
 276 } /* end namespace */
 277
 278
 279 void ssa_elimination(Program* program)
 280 {
 281    ssa_elimination_ctx ctx(program);
 282
 283    /* Collect information about every phi-instruction */
 284    collect_phi_info(ctx);
 285
 286    /* eliminate empty blocks */
 287    jump_threading(ctx);
 288
 289    /* insert parallelcopies from SSA elimination */
 290    insert_parallelcopies(ctx);
 291
 292 }
 293 }