src/mesa/drivers/dri/i965/brw_fs_register_coalesce.cpp

   1 /*
   2  * Copyright © 2012 Intel 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 /** @file brw_fs_register_coalesce.cpp
  25  *
  26  * Implements register coalescing: Checks if the two registers involved in a
  27  * raw move don't interfere, in which case they can both be stored in the same
  28  * place and the MOV removed.
  29  *
  30  * To do this, all uses of the source of the MOV in the shader are replaced
  31  * with the destination of the MOV. For example:
  32  *
  33  * add vgrf3:F, vgrf1:F, vgrf2:F
  34  * mov vgrf4:F, vgrf3:F
  35  * mul vgrf5:F, vgrf5:F, vgrf4:F
  36  *
  37  * becomes
  38  *
  39  * add vgrf4:F, vgrf1:F, vgrf2:F
  40  * mul vgrf5:F, vgrf5:F, vgrf4:F
  41  */
  42
  43 #include "brw_fs.h"
  44 #include "brw_fs_live_variables.h"
  45
  46 bool
  47 fs_visitor::register_coalesce()
  48 {
  49    bool progress = false;
  50
  51    calculate_live_intervals();
  52
  53    int src_size = 0;
  54    int channels_remaining = 0;
  55    int reg_from = -1, reg_to = -1;
  56    int reg_to_offset[MAX_SAMPLER_MESSAGE_SIZE];
  57    fs_inst *mov[MAX_SAMPLER_MESSAGE_SIZE];
  58
  59    foreach_list(node, &this->instructions) {
  60       fs_inst *inst = (fs_inst *)node;
  61
  62       if (inst->opcode != BRW_OPCODE_MOV ||
  63           inst->is_partial_write() ||
  64           inst->saturate ||
  65           inst->src[0].file != GRF ||
  66           inst->src[0].negate ||
  67           inst->src[0].abs ||
  68           !inst->src[0].is_contiguous() ||
  69           inst->dst.file != GRF ||
  70           inst->dst.type != inst->src[0].type) {
  71          continue;
  72       }
  73
  74       if (virtual_grf_sizes[inst->src[0].reg] >
  75           virtual_grf_sizes[inst->dst.reg])
  76          continue;
  77
  78       int var_from = live_intervals->var_from_reg(&inst->src[0]);
  79       int var_to = live_intervals->var_from_reg(&inst->dst);
  80
  81       if (live_intervals->vars_interfere(var_from, var_to) &&
  82           !inst->dst.equals(inst->src[0])) {
  83
  84          /* We know that the live ranges of A (var_from) and B (var_to)
  85           * interfere because of the ->vars_interfere() call above. If the end
  86           * of B's live range is after the end of A's range, then we know two
  87           * things:
  88           *  - the start of B's live range must be in A's live range (since we
  89           *    already know the two ranges interfere, this is the only remaining
  90           *    possibility)
  91           *  - the interference isn't of the form we're looking for (where B is
  92           *    entirely inside A)
  93           */
  94          if (live_intervals->end[var_to] > live_intervals->end[var_from])
  95             continue;
  96
  97          bool overwritten = false;
  98          int scan_ip = -1;
  99
 100          foreach_list(n, &this->instructions) {
 101             fs_inst *scan_inst = (fs_inst *)n;
 102             scan_ip++;
 103
 104             if (scan_inst->is_control_flow()) {
 105                overwritten = true;
 106                break;
 107             }
 108
 109             if (scan_ip <= live_intervals->start[var_to])
 110                continue;
 111
 112             if (scan_ip > live_intervals->end[var_to])
 113                break;
 114
 115             if (scan_inst->dst.equals(inst->dst) ||
 116                 scan_inst->dst.equals(inst->src[0])) {
 117                overwritten = true;
 118                break;
 119             }
 120          }
 121
 122          if (overwritten)
 123             continue;
 124       }
 125
 126       if (reg_from != inst->src[0].reg) {
 127          reg_from = inst->src[0].reg;
 128
 129          src_size = virtual_grf_sizes[inst->src[0].reg];
 130          assert(src_size <= MAX_SAMPLER_MESSAGE_SIZE);
 131
 132          channels_remaining = src_size;
 133          memset(mov, 0, sizeof(mov));
 134
 135          reg_to = inst->dst.reg;
 136       }
 137
 138       if (reg_to != inst->dst.reg)
 139          continue;
 140
 141       const int offset = inst->src[0].reg_offset;
 142       reg_to_offset[offset] = inst->dst.reg_offset;
 143       mov[offset] = inst;
 144       channels_remaining--;
 145
 146       if (channels_remaining)
 147          continue;
 148
 149       bool removed = false;
 150       for (int i = 0; i < src_size; i++) {
 151          if (mov[i]) {
 152             removed = true;
 153
 154             mov[i]->opcode = BRW_OPCODE_NOP;
 155             mov[i]->conditional_mod = BRW_CONDITIONAL_NONE;
 156             mov[i]->dst = reg_undef;
 157             mov[i]->src[0] = reg_undef;
 158             mov[i]->src[1] = reg_undef;
 159             mov[i]->src[2] = reg_undef;
 160          }
 161       }
 162
 163       foreach_list(node, &this->instructions) {
 164          fs_inst *scan_inst = (fs_inst *)node;
 165
 166          for (int i = 0; i < src_size; i++) {
 167             if (mov[i]) {
 168                if (scan_inst->dst.file == GRF &&
 169                    scan_inst->dst.reg == reg_from &&
 170                    scan_inst->dst.reg_offset == i) {
 171                   scan_inst->dst.reg = reg_to;
 172                   scan_inst->dst.reg_offset = reg_to_offset[i];
 173                }
 174                for (int j = 0; j < 3; j++) {
 175                   if (scan_inst->src[j].file == GRF &&
 176                       scan_inst->src[j].reg == reg_from &&
 177                       scan_inst->src[j].reg_offset == i) {
 178                      scan_inst->src[j].reg = reg_to;
 179                      scan_inst->src[j].reg_offset = reg_to_offset[i];
 180                   }
 181                }
 182             }
 183          }
 184       }
 185
 186       if (removed) {
 187          live_intervals->start[var_to] = MIN2(live_intervals->start[var_to],
 188                                               live_intervals->start[var_from]);
 189          live_intervals->end[var_to] = MAX2(live_intervals->end[var_to],
 190                                             live_intervals->end[var_from]);
 191          reg_from = -1;
 192       }
 193    }
 194
 195    foreach_list_safe(node, &this->instructions) {
 196       fs_inst *inst = (fs_inst *)node;
 197
 198       if (inst->opcode == BRW_OPCODE_NOP) {
 199          inst->remove();
 200          progress = true;
 201       }
 202    }
 203
 204    if (progress)
 205       invalidate_live_intervals();
 206
 207    return progress;
 208 }