src/glsl/nir/nir_opt_copy_propagate.c

   1 /*
   2  * Copyright © 2014 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  * Authors:
  24  *    Connor Abbott (cwabbott0@gmail.com)
  25  *
  26  */
  27
  28 #include "nir.h"
  29 #include <main/imports.h>
  30
  31 /**
  32  * SSA-based copy propagation
  33  */
  34
  35 static bool is_move(nir_alu_instr *instr)
  36 {
  37    if (instr->op != nir_op_fmov &&
  38        instr->op != nir_op_imov)
  39       return false;
  40
  41    if (instr->dest.saturate)
  42       return false;
  43
  44    /* we handle modifiers in a separate pass */
  45
  46    if (instr->src[0].abs || instr->src[0].negate)
  47       return false;
  48
  49    if (!instr->src[0].src.is_ssa)
  50       return false;
  51
  52    return true;
  53
  54 }
  55
  56 static bool
  57 is_swizzleless_move(nir_alu_instr *instr)
  58 {
  59    if (!is_move(instr))
  60       return false;
  61
  62    for (unsigned i = 0; i < 4; i++) {
  63       if (!((instr->dest.write_mask >> i) & 1))
  64          break;
  65       if (instr->src[0].swizzle[i] != i)
  66          return false;
  67    }
  68
  69    return true;
  70 }
  71
  72 static bool is_vec(nir_alu_instr *instr)
  73 {
  74    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
  75       if (!instr->src[i].src.is_ssa)
  76          return false;
  77
  78    return instr->op == nir_op_vec2 ||
  79           instr->op == nir_op_vec3 ||
  80           instr->op == nir_op_vec4;
  81 }
  82
  83 typedef struct {
  84    nir_ssa_def *def;
  85    bool found;
  86 } search_def_state;
  87
  88 static bool
  89 search_def(nir_src *src, void *_state)
  90 {
  91    search_def_state *state = (search_def_state *) _state;
  92
  93    if (src->is_ssa && src->ssa == state->def)
  94       state->found = true;
  95
  96    return true;
  97 }
  98
  99 static void
 100 rewrite_src_instr(nir_src *src, nir_ssa_def *new_def, nir_instr *parent_instr)
 101 {
 102    nir_ssa_def *old_def = src->ssa;
 103
 104    src->ssa = new_def;
 105
 106    /*
 107     * The instruction could still use the old definition in one of its other
 108     * sources, so only remove the instruction from the uses if there are no
 109     * more uses left.
 110     */
 111
 112    search_def_state search_state;
 113    search_state.def = old_def;
 114    search_state.found = false;
 115    nir_foreach_src(parent_instr, search_def, &search_state);
 116    if (!search_state.found) {
 117       struct set_entry *entry = _mesa_set_search(old_def->uses, parent_instr);
 118       assert(entry);
 119       _mesa_set_remove(old_def->uses, entry);
 120    }
 121
 122    _mesa_set_add(new_def->uses, parent_instr);
 123 }
 124
 125 static void
 126 rewrite_src_if(nir_if *if_stmt, nir_ssa_def *new_def)
 127 {
 128    nir_ssa_def *old_def = if_stmt->condition.ssa;
 129
 130    if_stmt->condition.ssa = new_def;
 131
 132    struct set_entry *entry = _mesa_set_search(old_def->if_uses, if_stmt);
 133    assert(entry);
 134    _mesa_set_remove(old_def->if_uses, entry);
 135
 136    _mesa_set_add(new_def->if_uses, if_stmt);
 137 }
 138
 139 static bool
 140 copy_prop_src(nir_src *src, nir_instr *parent_instr, nir_if *parent_if)
 141 {
 142    if (!src->is_ssa) {
 143       if (src->reg.indirect)
 144          return copy_prop_src(src, parent_instr, parent_if);
 145       return false;
 146    }
 147
 148    nir_instr *src_instr = src->ssa->parent_instr;
 149    if (src_instr->type != nir_instr_type_alu)
 150       return false;
 151
 152    nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
 153    if (!is_swizzleless_move(alu_instr))
 154       return false;
 155
 156    /* Don't let copy propagation land us with a phi that has more
 157     * components in its source than it has in its destination.  That badly
 158     * messes up out-of-ssa.
 159     */
 160    if (parent_instr && parent_instr->type == nir_instr_type_phi) {
 161       nir_phi_instr *phi = nir_instr_as_phi(parent_instr);
 162       assert(phi->dest.is_ssa);
 163       if (phi->dest.ssa.num_components !=
 164           alu_instr->src[0].src.ssa->num_components)
 165          return false;
 166    }
 167
 168    if (parent_instr)
 169       rewrite_src_instr(src, alu_instr->src[0].src.ssa, parent_instr);
 170    else
 171       rewrite_src_if(parent_if, alu_instr->src[0].src.ssa);
 172
 173    return true;
 174 }
 175
 176 static bool
 177 copy_prop_alu_src(nir_alu_instr *parent_alu_instr, unsigned index)
 178 {
 179    nir_alu_src *src = &parent_alu_instr->src[index];
 180    if (!src->src.is_ssa) {
 181       if (src->src.reg.indirect)
 182          return copy_prop_src(src->src.reg.indirect, &parent_alu_instr->instr,
 183                               NULL);
 184       return false;
 185    }
 186
 187    nir_instr *src_instr =  src->src.ssa->parent_instr;
 188    if (src_instr->type != nir_instr_type_alu)
 189       return false;
 190
 191    nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
 192    if (!is_move(alu_instr) && !is_vec(alu_instr))
 193       return false;
 194
 195    nir_ssa_def *def;
 196    unsigned new_swizzle[4] = {0, 0, 0, 0};
 197
 198    if (alu_instr->op == nir_op_fmov ||
 199        alu_instr->op == nir_op_imov) {
 200       for (unsigned i = 0; i < 4; i++)
 201          new_swizzle[i] = alu_instr->src[0].swizzle[src->swizzle[i]];
 202       def = alu_instr->src[0].src.ssa;
 203    } else {
 204       def = NULL;
 205
 206       for (unsigned i = 0; i < 4; i++) {
 207          if (!nir_alu_instr_channel_used(parent_alu_instr, index, i))
 208             continue;
 209
 210          nir_ssa_def *new_def = alu_instr->src[src->swizzle[i]].src.ssa;
 211          if (def == NULL)
 212             def = new_def;
 213          else {
 214             if (def != new_def)
 215                return false;
 216          }
 217          new_swizzle[i] = alu_instr->src[src->swizzle[i]].swizzle[0];
 218       }
 219    }
 220
 221    for (unsigned i = 0; i < 4; i++)
 222       src->swizzle[i] = new_swizzle[i];
 223
 224    rewrite_src_instr(&src->src, def, &parent_alu_instr->instr);
 225
 226    return true;
 227 }
 228
 229 typedef struct {
 230    nir_instr *parent_instr;
 231    bool progress;
 232 } copy_prop_state;
 233
 234 static bool
 235 copy_prop_src_cb(nir_src *src, void *_state)
 236 {
 237    copy_prop_state *state = (copy_prop_state *) _state;
 238    while (copy_prop_src(src, state->parent_instr, NULL))
 239       state->progress = true;
 240
 241    return true;
 242 }
 243
 244 static bool
 245 copy_prop_instr(nir_instr *instr)
 246 {
 247    if (instr->type == nir_instr_type_alu) {
 248       nir_alu_instr *alu_instr = nir_instr_as_alu(instr);
 249       bool progress = false;
 250
 251       for (unsigned i = 0; i < nir_op_infos[alu_instr->op].num_inputs; i++)
 252          while (copy_prop_alu_src(alu_instr, i))
 253             progress = true;
 254
 255       if (!alu_instr->dest.dest.is_ssa && alu_instr->dest.dest.reg.indirect)
 256          while (copy_prop_src(alu_instr->dest.dest.reg.indirect, instr, NULL))
 257             progress = true;
 258
 259       return progress;
 260    }
 261
 262    copy_prop_state state;
 263    state.parent_instr = instr;
 264    state.progress = false;
 265    nir_foreach_src(instr, copy_prop_src_cb, &state);
 266
 267    return state.progress;
 268 }
 269
 270 static bool
 271 copy_prop_if(nir_if *if_stmt)
 272 {
 273    return copy_prop_src(&if_stmt->condition, NULL, if_stmt);
 274 }
 275
 276 static bool
 277 copy_prop_block(nir_block *block, void *_state)
 278 {
 279    bool *progress = (bool *) _state;
 280
 281    nir_foreach_instr(block, instr) {
 282       if (copy_prop_instr(instr))
 283          *progress = true;
 284    }
 285
 286    if (block->cf_node.node.next != NULL && /* check that we aren't the end node */
 287        !nir_cf_node_is_last(&block->cf_node) &&
 288        nir_cf_node_next(&block->cf_node)->type == nir_cf_node_if) {
 289       nir_if *if_stmt = nir_cf_node_as_if(nir_cf_node_next(&block->cf_node));
 290       if (copy_prop_if(if_stmt))
 291          *progress = true;
 292    }
 293
 294    return true;
 295 }
 296
 297 bool
 298 nir_copy_prop_impl(nir_function_impl *impl)
 299 {
 300    bool progress = false;
 301
 302    nir_foreach_block(impl, copy_prop_block, &progress);
 303    return progress;
 304 }
 305
 306 bool
 307 nir_copy_prop(nir_shader *shader)
 308 {
 309    bool progress = false;
 310
 311    nir_foreach_overload(shader, overload) {
 312       if (overload->impl && nir_copy_prop_impl(overload->impl))
 313          progress = true;
 314    }
 315
 316    return progress;
 317 }