src/compiler/nir/nir_instr_set.c

   1 /*
   2  * Copyright © 2014 Connor Abbott
   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 #include "nir_instr_set.h"
  25 #include "nir_vla.h"
  26
  27 #define HASH(hash, data) _mesa_fnv32_1a_accumulate((hash), (data))
  28
  29 static uint32_t
  30 hash_src(uint32_t hash, const nir_src *src)
  31 {
  32    assert(src->is_ssa);
  33    hash = HASH(hash, src->ssa);
  34    return hash;
  35 }
  36
  37 static uint32_t
  38 hash_alu_src(uint32_t hash, const nir_alu_src *src, unsigned num_components)
  39 {
  40    hash = HASH(hash, src->abs);
  41    hash = HASH(hash, src->negate);
  42
  43    for (unsigned i = 0; i < num_components; i++)
  44       hash = HASH(hash, src->swizzle[i]);
  45
  46    hash = hash_src(hash, &src->src);
  47    return hash;
  48 }
  49
  50 static uint32_t
  51 hash_alu(uint32_t hash, const nir_alu_instr *instr)
  52 {
  53    hash = HASH(hash, instr->op);
  54    hash = HASH(hash, instr->dest.dest.ssa.num_components);
  55
  56    if (nir_op_infos[instr->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) {
  57       assert(nir_op_infos[instr->op].num_inputs == 2);
  58       uint32_t hash0 = hash_alu_src(hash, &instr->src[0],
  59                                     nir_ssa_alu_instr_src_components(instr, 0));
  60       uint32_t hash1 = hash_alu_src(hash, &instr->src[1],
  61                                     nir_ssa_alu_instr_src_components(instr, 1));
  62       /* For commutative operations, we need some commutative way of
  63        * combining the hashes.  One option would be to XOR them but that
  64        * means that anything with two identical sources will hash to 0 and
  65        * that's common enough we probably don't want the guaranteed
  66        * collision.  Either addition or multiplication will also work.
  67        */
  68       hash = hash0 * hash1;
  69    } else {
  70       for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
  71          hash = hash_alu_src(hash, &instr->src[i],
  72                              nir_ssa_alu_instr_src_components(instr, i));
  73       }
  74    }
  75
  76    return hash;
  77 }
  78
  79 static uint32_t
  80 hash_load_const(uint32_t hash, const nir_load_const_instr *instr)
  81 {
  82    hash = HASH(hash, instr->def.num_components);
  83
  84    hash = _mesa_fnv32_1a_accumulate_block(hash, instr->value.f32,
  85                                           instr->def.num_components
  86                                              * sizeof(instr->value.f32[0]));
  87
  88    return hash;
  89 }
  90
  91 static int
  92 cmp_phi_src(const void *data1, const void *data2)
  93 {
  94    nir_phi_src *src1 = *(nir_phi_src **)data1;
  95    nir_phi_src *src2 = *(nir_phi_src **)data2;
  96    return src1->pred - src2->pred;
  97 }
  98
  99 static uint32_t
 100 hash_phi(uint32_t hash, const nir_phi_instr *instr)
 101 {
 102    hash = HASH(hash, instr->instr.block);
 103
 104    /* sort sources by predecessor, since the order shouldn't matter */
 105    unsigned num_preds = instr->instr.block->predecessors->entries;
 106    NIR_VLA(nir_phi_src *, srcs, num_preds);
 107    unsigned i = 0;
 108    nir_foreach_phi_src(instr, src) {
 109       srcs[i++] = src;
 110    }
 111
 112    qsort(srcs, num_preds, sizeof(nir_phi_src *), cmp_phi_src);
 113
 114    for (i = 0; i < num_preds; i++) {
 115       hash = hash_src(hash, &srcs[i]->src);
 116       hash = HASH(hash, srcs[i]->pred);
 117    }
 118
 119    return hash;
 120 }
 121
 122 static uint32_t
 123 hash_intrinsic(uint32_t hash, const nir_intrinsic_instr *instr)
 124 {
 125    const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
 126    hash = HASH(hash, instr->intrinsic);
 127
 128    if (info->has_dest)
 129       hash = HASH(hash, instr->dest.ssa.num_components);
 130
 131    assert(info->num_variables == 0);
 132
 133    hash = _mesa_fnv32_1a_accumulate_block(hash, instr->const_index,
 134                                           info->num_indices
 135                                              * sizeof(instr->const_index[0]));
 136    return hash;
 137 }
 138
 139 static uint32_t
 140 hash_tex(uint32_t hash, const nir_tex_instr *instr)
 141 {
 142    hash = HASH(hash, instr->op);
 143    hash = HASH(hash, instr->num_srcs);
 144
 145    for (unsigned i = 0; i < instr->num_srcs; i++) {
 146       hash = HASH(hash, instr->src[i].src_type);
 147       hash = hash_src(hash, &instr->src[i].src);
 148    }
 149
 150    hash = HASH(hash, instr->coord_components);
 151    hash = HASH(hash, instr->sampler_dim);
 152    hash = HASH(hash, instr->is_array);
 153    hash = HASH(hash, instr->is_shadow);
 154    hash = HASH(hash, instr->is_new_style_shadow);
 155    unsigned component = instr->component;
 156    hash = HASH(hash, component);
 157    hash = HASH(hash, instr->texture_index);
 158    hash = HASH(hash, instr->texture_array_size);
 159    hash = HASH(hash, instr->sampler_index);
 160
 161    assert(!instr->texture && !instr->sampler);
 162
 163    return hash;
 164 }
 165
 166 /* Computes a hash of an instruction for use in a hash table. Note that this
 167  * will only work for instructions where instr_can_rewrite() returns true, and
 168  * it should return identical hashes for two instructions that are the same
 169  * according nir_instrs_equal().
 170  */
 171
 172 static uint32_t
 173 hash_instr(const void *data)
 174 {
 175    const nir_instr *instr = data;
 176    uint32_t hash = _mesa_fnv32_1a_offset_bias;
 177
 178    switch (instr->type) {
 179    case nir_instr_type_alu:
 180       hash = hash_alu(hash, nir_instr_as_alu(instr));
 181       break;
 182    case nir_instr_type_load_const:
 183       hash = hash_load_const(hash, nir_instr_as_load_const(instr));
 184       break;
 185    case nir_instr_type_phi:
 186       hash = hash_phi(hash, nir_instr_as_phi(instr));
 187       break;
 188    case nir_instr_type_intrinsic:
 189       hash = hash_intrinsic(hash, nir_instr_as_intrinsic(instr));
 190       break;
 191    case nir_instr_type_tex:
 192       hash = hash_tex(hash, nir_instr_as_tex(instr));
 193       break;
 194    default:
 195       unreachable("Invalid instruction type");
 196    }
 197
 198    return hash;
 199 }
 200
 201 bool
 202 nir_srcs_equal(nir_src src1, nir_src src2)
 203 {
 204    if (src1.is_ssa) {
 205       if (src2.is_ssa) {
 206          return src1.ssa == src2.ssa;
 207       } else {
 208          return false;
 209       }
 210    } else {
 211       if (src2.is_ssa) {
 212          return false;
 213       } else {
 214          if ((src1.reg.indirect == NULL) != (src2.reg.indirect == NULL))
 215             return false;
 216
 217          if (src1.reg.indirect) {
 218             if (!nir_srcs_equal(*src1.reg.indirect, *src2.reg.indirect))
 219                return false;
 220          }
 221
 222          return src1.reg.reg == src2.reg.reg &&
 223                 src1.reg.base_offset == src2.reg.base_offset;
 224       }
 225    }
 226 }
 227
 228 static bool
 229 nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2,
 230                    unsigned src1, unsigned src2)
 231 {
 232    if (alu1->src[src1].abs != alu2->src[src2].abs ||
 233        alu1->src[src1].negate != alu2->src[src2].negate)
 234       return false;
 235
 236    for (unsigned i = 0; i < nir_ssa_alu_instr_src_components(alu1, src1); i++) {
 237       if (alu1->src[src1].swizzle[i] != alu2->src[src2].swizzle[i])
 238          return false;
 239    }
 240
 241    return nir_srcs_equal(alu1->src[src1].src, alu2->src[src2].src);
 242 }
 243
 244 /* Returns "true" if two instructions are equal. Note that this will only
 245  * work for the subset of instructions defined by instr_can_rewrite(). Also,
 246  * it should only return "true" for instructions that hash_instr() will return
 247  * the same hash for (ignoring collisions, of course).
 248  */
 249
 250 static bool
 251 nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2)
 252 {
 253    if (instr1->type != instr2->type)
 254       return false;
 255
 256    switch (instr1->type) {
 257    case nir_instr_type_alu: {
 258       nir_alu_instr *alu1 = nir_instr_as_alu(instr1);
 259       nir_alu_instr *alu2 = nir_instr_as_alu(instr2);
 260
 261       if (alu1->op != alu2->op)
 262          return false;
 263
 264       /* TODO: We can probably acutally do something more inteligent such
 265        * as allowing different numbers and taking a maximum or something
 266        * here */
 267       if (alu1->dest.dest.ssa.num_components != alu2->dest.dest.ssa.num_components)
 268          return false;
 269
 270       if (nir_op_infos[alu1->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) {
 271          assert(nir_op_infos[alu1->op].num_inputs == 2);
 272          return (nir_alu_srcs_equal(alu1, alu2, 0, 0) &&
 273                  nir_alu_srcs_equal(alu1, alu2, 1, 1)) ||
 274                 (nir_alu_srcs_equal(alu1, alu2, 0, 1) &&
 275                  nir_alu_srcs_equal(alu1, alu2, 1, 0));
 276       } else {
 277          for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) {
 278             if (!nir_alu_srcs_equal(alu1, alu2, i, i))
 279                return false;
 280          }
 281       }
 282       return true;
 283    }
 284    case nir_instr_type_tex: {
 285       nir_tex_instr *tex1 = nir_instr_as_tex(instr1);
 286       nir_tex_instr *tex2 = nir_instr_as_tex(instr2);
 287
 288       if (tex1->op != tex2->op)
 289          return false;
 290
 291       if (tex1->num_srcs != tex2->num_srcs)
 292          return false;
 293       for (unsigned i = 0; i < tex1->num_srcs; i++) {
 294          if (tex1->src[i].src_type != tex2->src[i].src_type ||
 295              !nir_srcs_equal(tex1->src[i].src, tex2->src[i].src)) {
 296             return false;
 297          }
 298       }
 299
 300       if (tex1->coord_components != tex2->coord_components ||
 301           tex1->sampler_dim != tex2->sampler_dim ||
 302           tex1->is_array != tex2->is_array ||
 303           tex1->is_shadow != tex2->is_shadow ||
 304           tex1->is_new_style_shadow != tex2->is_new_style_shadow ||
 305           tex1->component != tex2->component ||
 306          tex1->texture_index != tex2->texture_index ||
 307          tex1->texture_array_size != tex2->texture_array_size ||
 308          tex1->sampler_index != tex2->sampler_index) {
 309          return false;
 310       }
 311
 312       /* Don't support un-lowered sampler derefs currently. */
 313       assert(!tex1->texture && !tex1->sampler &&
 314              !tex2->texture && !tex2->sampler);
 315
 316       return true;
 317    }
 318    case nir_instr_type_load_const: {
 319       nir_load_const_instr *load1 = nir_instr_as_load_const(instr1);
 320       nir_load_const_instr *load2 = nir_instr_as_load_const(instr2);
 321
 322       if (load1->def.num_components != load2->def.num_components)
 323          return false;
 324
 325       return memcmp(load1->value.f32, load2->value.f32,
 326                     load1->def.num_components * sizeof(*load2->value.f32)) == 0;
 327    }
 328    case nir_instr_type_phi: {
 329       nir_phi_instr *phi1 = nir_instr_as_phi(instr1);
 330       nir_phi_instr *phi2 = nir_instr_as_phi(instr2);
 331
 332       if (phi1->instr.block != phi2->instr.block)
 333          return false;
 334
 335       nir_foreach_phi_src(phi1, src1) {
 336          nir_foreach_phi_src(phi2, src2) {
 337             if (src1->pred == src2->pred) {
 338                if (!nir_srcs_equal(src1->src, src2->src))
 339                   return false;
 340
 341                break;
 342             }
 343          }
 344       }
 345
 346       return true;
 347    }
 348    case nir_instr_type_intrinsic: {
 349       nir_intrinsic_instr *intrinsic1 = nir_instr_as_intrinsic(instr1);
 350       nir_intrinsic_instr *intrinsic2 = nir_instr_as_intrinsic(instr2);
 351       const nir_intrinsic_info *info =
 352          &nir_intrinsic_infos[intrinsic1->intrinsic];
 353
 354       if (intrinsic1->intrinsic != intrinsic2->intrinsic ||
 355           intrinsic1->num_components != intrinsic2->num_components)
 356          return false;
 357
 358       if (info->has_dest && intrinsic1->dest.ssa.num_components !=
 359                             intrinsic2->dest.ssa.num_components)
 360          return false;
 361
 362       for (unsigned i = 0; i < info->num_srcs; i++) {
 363          if (!nir_srcs_equal(intrinsic1->src[i], intrinsic2->src[i]))
 364             return false;
 365       }
 366
 367       assert(info->num_variables == 0);
 368
 369       for (unsigned i = 0; i < info->num_indices; i++) {
 370          if (intrinsic1->const_index[i] != intrinsic2->const_index[i])
 371             return false;
 372       }
 373
 374       return true;
 375    }
 376    case nir_instr_type_call:
 377    case nir_instr_type_jump:
 378    case nir_instr_type_ssa_undef:
 379    case nir_instr_type_parallel_copy:
 380    default:
 381       unreachable("Invalid instruction type");
 382    }
 383
 384    return false;
 385 }
 386
 387 static bool
 388 src_is_ssa(nir_src *src, void *data)
 389 {
 390    (void) data;
 391    return src->is_ssa;
 392 }
 393
 394 static bool
 395 dest_is_ssa(nir_dest *dest, void *data)
 396 {
 397    (void) data;
 398    return dest->is_ssa;
 399 }
 400
 401 /* This function determines if uses of an instruction can safely be rewritten
 402  * to use another identical instruction instead. Note that this function must
 403  * be kept in sync with hash_instr() and nir_instrs_equal() -- only
 404  * instructions that pass this test will be handed on to those functions, and
 405  * conversely they must handle everything that this function returns true for.
 406  */
 407
 408 static bool
 409 instr_can_rewrite(nir_instr *instr)
 410 {
 411    /* We only handle SSA. */
 412    if (!nir_foreach_dest(instr, dest_is_ssa, NULL) ||
 413        !nir_foreach_src(instr, src_is_ssa, NULL))
 414       return false;
 415
 416    switch (instr->type) {
 417    case nir_instr_type_alu:
 418    case nir_instr_type_load_const:
 419    case nir_instr_type_phi:
 420       return true;
 421    case nir_instr_type_tex: {
 422       nir_tex_instr *tex = nir_instr_as_tex(instr);
 423
 424       /* Don't support un-lowered sampler derefs currently. */
 425       if (tex->texture || tex->sampler)
 426          return false;
 427
 428       return true;
 429    }
 430    case nir_instr_type_intrinsic: {
 431       const nir_intrinsic_info *info =
 432          &nir_intrinsic_infos[nir_instr_as_intrinsic(instr)->intrinsic];
 433       return (info->flags & NIR_INTRINSIC_CAN_ELIMINATE) &&
 434              (info->flags & NIR_INTRINSIC_CAN_REORDER) &&
 435              info->num_variables == 0; /* not implemented yet */
 436    }
 437    case nir_instr_type_call:
 438    case nir_instr_type_jump:
 439    case nir_instr_type_ssa_undef:
 440       return false;
 441    case nir_instr_type_parallel_copy:
 442    default:
 443       unreachable("Invalid instruction type");
 444    }
 445
 446    return false;
 447 }
 448
 449 static nir_ssa_def *
 450 nir_instr_get_dest_ssa_def(nir_instr *instr)
 451 {
 452    switch (instr->type) {
 453    case nir_instr_type_alu:
 454       assert(nir_instr_as_alu(instr)->dest.dest.is_ssa);
 455       return &nir_instr_as_alu(instr)->dest.dest.ssa;
 456    case nir_instr_type_load_const:
 457       return &nir_instr_as_load_const(instr)->def;
 458    case nir_instr_type_phi:
 459       assert(nir_instr_as_phi(instr)->dest.is_ssa);
 460       return &nir_instr_as_phi(instr)->dest.ssa;
 461    case nir_instr_type_intrinsic:
 462       assert(nir_instr_as_intrinsic(instr)->dest.is_ssa);
 463       return &nir_instr_as_intrinsic(instr)->dest.ssa;
 464    case nir_instr_type_tex:
 465       assert(nir_instr_as_tex(instr)->dest.is_ssa);
 466       return &nir_instr_as_tex(instr)->dest.ssa;
 467    default:
 468       unreachable("We never ask for any of these");
 469    }
 470 }
 471
 472 static bool
 473 cmp_func(const void *data1, const void *data2)
 474 {
 475    return nir_instrs_equal(data1, data2);
 476 }
 477
 478 struct set *
 479 nir_instr_set_create(void *mem_ctx)
 480 {
 481    return _mesa_set_create(mem_ctx, hash_instr, cmp_func);
 482 }
 483
 484 void
 485 nir_instr_set_destroy(struct set *instr_set)
 486 {
 487    _mesa_set_destroy(instr_set, NULL);
 488 }
 489
 490 bool
 491 nir_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr)
 492 {
 493    if (!instr_can_rewrite(instr))
 494       return false;
 495
 496    struct set_entry *entry = _mesa_set_search(instr_set, instr);
 497    if (entry) {
 498       nir_ssa_def *def = nir_instr_get_dest_ssa_def(instr);
 499       nir_ssa_def *new_def =
 500          nir_instr_get_dest_ssa_def((nir_instr *) entry->key);
 501       nir_ssa_def_rewrite_uses(def, nir_src_for_ssa(new_def));
 502       return true;
 503    }
 504
 505    _mesa_set_add(instr_set, instr);
 506    return false;
 507 }
 508
 509 void
 510 nir_instr_set_remove(struct set *instr_set, nir_instr *instr)
 511 {
 512    if (!instr_can_rewrite(instr))
 513       return;
 514
 515    struct set_entry *entry = _mesa_set_search(instr_set, instr);
 516    if (entry)
 517       _mesa_set_remove(instr_set, entry);
 518 }
 519