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.f,
  85                                           instr->def.num_components
  86                                              * sizeof(instr->value.f[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    hash = HASH(hash, instr->const_offset);
 156    unsigned component = instr->component;
 157    hash = HASH(hash, component);
 158    hash = HASH(hash, instr->texture_index);
 159    hash = HASH(hash, instr->texture_array_size);
 160    hash = HASH(hash, instr->sampler_index);
 161
 162    assert(!instr->texture && !instr->sampler);
 163
 164    return hash;
 165 }
 166
 167 /* Computes a hash of an instruction for use in a hash table. Note that this
 168  * will only work for instructions where instr_can_rewrite() returns true, and
 169  * it should return identical hashes for two instructions that are the same
 170  * according nir_instrs_equal().
 171  */
 172
 173 static uint32_t
 174 hash_instr(const void *data)
 175 {
 176    const nir_instr *instr = data;
 177    uint32_t hash = _mesa_fnv32_1a_offset_bias;
 178
 179    switch (instr->type) {
 180    case nir_instr_type_alu:
 181       hash = hash_alu(hash, nir_instr_as_alu(instr));
 182       break;
 183    case nir_instr_type_load_const:
 184       hash = hash_load_const(hash, nir_instr_as_load_const(instr));
 185       break;
 186    case nir_instr_type_phi:
 187       hash = hash_phi(hash, nir_instr_as_phi(instr));
 188       break;
 189    case nir_instr_type_intrinsic:
 190       hash = hash_intrinsic(hash, nir_instr_as_intrinsic(instr));
 191       break;
 192    case nir_instr_type_tex:
 193       hash = hash_tex(hash, nir_instr_as_tex(instr));
 194       break;
 195    default:
 196       unreachable("Invalid instruction type");
 197    }
 198
 199    return hash;
 200 }
 201
 202 bool
 203 nir_srcs_equal(nir_src src1, nir_src src2)
 204 {
 205    if (src1.is_ssa) {
 206       if (src2.is_ssa) {
 207          return src1.ssa == src2.ssa;
 208       } else {
 209          return false;
 210       }
 211    } else {
 212       if (src2.is_ssa) {
 213          return false;
 214       } else {
 215          if ((src1.reg.indirect == NULL) != (src2.reg.indirect == NULL))
 216             return false;
 217
 218          if (src1.reg.indirect) {
 219             if (!nir_srcs_equal(*src1.reg.indirect, *src2.reg.indirect))
 220                return false;
 221          }
 222
 223          return src1.reg.reg == src2.reg.reg &&
 224                 src1.reg.base_offset == src2.reg.base_offset;
 225       }
 226    }
 227 }
 228
 229 static bool
 230 nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2,
 231                    unsigned src1, unsigned src2)
 232 {
 233    if (alu1->src[src1].abs != alu2->src[src2].abs ||
 234        alu1->src[src1].negate != alu2->src[src2].negate)
 235       return false;
 236
 237    for (unsigned i = 0; i < nir_ssa_alu_instr_src_components(alu1, src1); i++) {
 238       if (alu1->src[src1].swizzle[i] != alu2->src[src2].swizzle[i])
 239          return false;
 240    }
 241
 242    return nir_srcs_equal(alu1->src[src1].src, alu2->src[src2].src);
 243 }
 244
 245 /* Returns "true" if two instructions are equal. Note that this will only
 246  * work for the subset of instructions defined by instr_can_rewrite(). Also,
 247  * it should only return "true" for instructions that hash_instr() will return
 248  * the same hash for (ignoring collisions, of course).
 249  */
 250
 251 static bool
 252 nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2)
 253 {
 254    if (instr1->type != instr2->type)
 255       return false;
 256
 257    switch (instr1->type) {
 258    case nir_instr_type_alu: {
 259       nir_alu_instr *alu1 = nir_instr_as_alu(instr1);
 260       nir_alu_instr *alu2 = nir_instr_as_alu(instr2);
 261
 262       if (alu1->op != alu2->op)
 263          return false;
 264
 265       /* TODO: We can probably acutally do something more inteligent such
 266        * as allowing different numbers and taking a maximum or something
 267        * here */
 268       if (alu1->dest.dest.ssa.num_components != alu2->dest.dest.ssa.num_components)
 269          return false;
 270
 271       if (nir_op_infos[alu1->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) {
 272          assert(nir_op_infos[alu1->op].num_inputs == 2);
 273          return (nir_alu_srcs_equal(alu1, alu2, 0, 0) &&
 274                  nir_alu_srcs_equal(alu1, alu2, 1, 1)) ||
 275                 (nir_alu_srcs_equal(alu1, alu2, 0, 1) &&
 276                  nir_alu_srcs_equal(alu1, alu2, 1, 0));
 277       } else {
 278          for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) {
 279             if (!nir_alu_srcs_equal(alu1, alu2, i, i))
 280                return false;
 281          }
 282       }
 283       return true;
 284    }
 285    case nir_instr_type_tex: {
 286       nir_tex_instr *tex1 = nir_instr_as_tex(instr1);
 287       nir_tex_instr *tex2 = nir_instr_as_tex(instr2);
 288
 289       if (tex1->op != tex2->op)
 290          return false;
 291
 292       if (tex1->num_srcs != tex2->num_srcs)
 293          return false;
 294       for (unsigned i = 0; i < tex1->num_srcs; i++) {
 295          if (tex1->src[i].src_type != tex2->src[i].src_type ||
 296              !nir_srcs_equal(tex1->src[i].src, tex2->src[i].src)) {
 297             return false;
 298          }
 299       }
 300
 301       if (tex1->coord_components != tex2->coord_components ||
 302           tex1->sampler_dim != tex2->sampler_dim ||
 303           tex1->is_array != tex2->is_array ||
 304           tex1->is_shadow != tex2->is_shadow ||
 305           tex1->is_new_style_shadow != tex2->is_new_style_shadow ||
 306           memcmp(tex1->const_offset, tex2->const_offset,
 307                  sizeof(tex1->const_offset)) != 0 ||
 308           tex1->component != tex2->component ||
 309          tex1->texture_index != tex2->texture_index ||
 310          tex1->texture_array_size != tex2->texture_array_size ||
 311          tex1->sampler_index != tex2->sampler_index) {
 312          return false;
 313       }
 314
 315       /* Don't support un-lowered sampler derefs currently. */
 316       assert(!tex1->texture && !tex1->sampler &&
 317              !tex2->texture && !tex2->sampler);
 318
 319       return true;
 320    }
 321    case nir_instr_type_load_const: {
 322       nir_load_const_instr *load1 = nir_instr_as_load_const(instr1);
 323       nir_load_const_instr *load2 = nir_instr_as_load_const(instr2);
 324
 325       if (load1->def.num_components != load2->def.num_components)
 326          return false;
 327
 328       return memcmp(load1->value.f, load2->value.f,
 329                     load1->def.num_components * sizeof(*load2->value.f)) == 0;
 330    }
 331    case nir_instr_type_phi: {
 332       nir_phi_instr *phi1 = nir_instr_as_phi(instr1);
 333       nir_phi_instr *phi2 = nir_instr_as_phi(instr2);
 334
 335       if (phi1->instr.block != phi2->instr.block)
 336          return false;
 337
 338       nir_foreach_phi_src(phi1, src1) {
 339          nir_foreach_phi_src(phi2, src2) {
 340             if (src1->pred == src2->pred) {
 341                if (!nir_srcs_equal(src1->src, src2->src))
 342                   return false;
 343
 344                break;
 345             }
 346          }
 347       }
 348
 349       return true;
 350    }
 351    case nir_instr_type_intrinsic: {
 352       nir_intrinsic_instr *intrinsic1 = nir_instr_as_intrinsic(instr1);
 353       nir_intrinsic_instr *intrinsic2 = nir_instr_as_intrinsic(instr2);
 354       const nir_intrinsic_info *info =
 355          &nir_intrinsic_infos[intrinsic1->intrinsic];
 356
 357       if (intrinsic1->intrinsic != intrinsic2->intrinsic ||
 358           intrinsic1->num_components != intrinsic2->num_components)
 359          return false;
 360
 361       if (info->has_dest && intrinsic1->dest.ssa.num_components !=
 362                             intrinsic2->dest.ssa.num_components)
 363          return false;
 364
 365       for (unsigned i = 0; i < info->num_srcs; i++) {
 366          if (!nir_srcs_equal(intrinsic1->src[i], intrinsic2->src[i]))
 367             return false;
 368       }
 369
 370       assert(info->num_variables == 0);
 371
 372       for (unsigned i = 0; i < info->num_indices; i++) {
 373          if (intrinsic1->const_index[i] != intrinsic2->const_index[i])
 374             return false;
 375       }
 376
 377       return true;
 378    }
 379    case nir_instr_type_call:
 380    case nir_instr_type_jump:
 381    case nir_instr_type_ssa_undef:
 382    case nir_instr_type_parallel_copy:
 383    default:
 384       unreachable("Invalid instruction type");
 385    }
 386
 387    return false;
 388 }
 389
 390 static bool
 391 src_is_ssa(nir_src *src, void *data)
 392 {
 393    (void) data;
 394    return src->is_ssa;
 395 }
 396
 397 static bool
 398 dest_is_ssa(nir_dest *dest, void *data)
 399 {
 400    (void) data;
 401    return dest->is_ssa;
 402 }
 403
 404 /* This function determines if uses of an instruction can safely be rewritten
 405  * to use another identical instruction instead. Note that this function must
 406  * be kept in sync with hash_instr() and nir_instrs_equal() -- only
 407  * instructions that pass this test will be handed on to those functions, and
 408  * conversely they must handle everything that this function returns true for.
 409  */
 410
 411 static bool
 412 instr_can_rewrite(nir_instr *instr)
 413 {
 414    /* We only handle SSA. */
 415    if (!nir_foreach_dest(instr, dest_is_ssa, NULL) ||
 416        !nir_foreach_src(instr, src_is_ssa, NULL))
 417       return false;
 418
 419    switch (instr->type) {
 420    case nir_instr_type_alu:
 421    case nir_instr_type_load_const:
 422    case nir_instr_type_phi:
 423       return true;
 424    case nir_instr_type_tex: {
 425       nir_tex_instr *tex = nir_instr_as_tex(instr);
 426
 427       /* Don't support un-lowered sampler derefs currently. */
 428       if (tex->texture || tex->sampler)
 429          return false;
 430
 431       return true;
 432    }
 433    case nir_instr_type_intrinsic: {
 434       const nir_intrinsic_info *info =
 435          &nir_intrinsic_infos[nir_instr_as_intrinsic(instr)->intrinsic];
 436       return (info->flags & NIR_INTRINSIC_CAN_ELIMINATE) &&
 437              (info->flags & NIR_INTRINSIC_CAN_REORDER) &&
 438              info->num_variables == 0; /* not implemented yet */
 439    }
 440    case nir_instr_type_call:
 441    case nir_instr_type_jump:
 442    case nir_instr_type_ssa_undef:
 443       return false;
 444    case nir_instr_type_parallel_copy:
 445    default:
 446       unreachable("Invalid instruction type");
 447    }
 448
 449    return false;
 450 }
 451
 452 static nir_ssa_def *
 453 nir_instr_get_dest_ssa_def(nir_instr *instr)
 454 {
 455    switch (instr->type) {
 456    case nir_instr_type_alu:
 457       assert(nir_instr_as_alu(instr)->dest.dest.is_ssa);
 458       return &nir_instr_as_alu(instr)->dest.dest.ssa;
 459    case nir_instr_type_load_const:
 460       return &nir_instr_as_load_const(instr)->def;
 461    case nir_instr_type_phi:
 462       assert(nir_instr_as_phi(instr)->dest.is_ssa);
 463       return &nir_instr_as_phi(instr)->dest.ssa;
 464    case nir_instr_type_intrinsic:
 465       assert(nir_instr_as_intrinsic(instr)->dest.is_ssa);
 466       return &nir_instr_as_intrinsic(instr)->dest.ssa;
 467    case nir_instr_type_tex:
 468       assert(nir_instr_as_tex(instr)->dest.is_ssa);
 469       return &nir_instr_as_tex(instr)->dest.ssa;
 470    default:
 471       unreachable("We never ask for any of these");
 472    }
 473 }
 474
 475 static bool
 476 cmp_func(const void *data1, const void *data2)
 477 {
 478    return nir_instrs_equal(data1, data2);
 479 }
 480
 481 struct set *
 482 nir_instr_set_create(void *mem_ctx)
 483 {
 484    return _mesa_set_create(mem_ctx, hash_instr, cmp_func);
 485 }
 486
 487 void
 488 nir_instr_set_destroy(struct set *instr_set)
 489 {
 490    _mesa_set_destroy(instr_set, NULL);
 491 }
 492
 493 bool
 494 nir_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr)
 495 {
 496    if (!instr_can_rewrite(instr))
 497       return false;
 498
 499    struct set_entry *entry = _mesa_set_search(instr_set, instr);
 500    if (entry) {
 501       nir_ssa_def *def = nir_instr_get_dest_ssa_def(instr);
 502       nir_ssa_def *new_def =
 503          nir_instr_get_dest_ssa_def((nir_instr *) entry->key);
 504       nir_ssa_def_rewrite_uses(def, nir_src_for_ssa(new_def));
 505       return true;
 506    }
 507
 508    _mesa_set_add(instr_set, instr);
 509    return false;
 510 }
 511
 512 void
 513 nir_instr_set_remove(struct set *instr_set, nir_instr *instr)
 514 {
 515    if (!instr_can_rewrite(instr))
 516       return;
 517
 518    struct set_entry *entry = _mesa_set_search(instr_set, instr);
 519    if (entry)
 520       _mesa_set_remove(instr_set, entry);
 521 }
 522