src/compiler/nir/nir_serialize.c

   1 /*
   2  * Copyright © 2017 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_serialize.h"
  25 #include "nir_control_flow.h"
  26 #include "util/u_dynarray.h"
  27
  28 typedef struct {
  29    size_t blob_offset;
  30    nir_ssa_def *src;
  31    nir_block *block;
  32 } write_phi_fixup;
  33
  34 typedef struct {
  35    const nir_shader *nir;
  36
  37    struct blob *blob;
  38
  39    /* maps pointer to index */
  40    struct hash_table *remap_table;
  41
  42    /* the next index to assign to a NIR in-memory object */
  43    uintptr_t next_idx;
  44
  45    /* Array of write_phi_fixup structs representing phi sources that need to
  46     * be resolved in the second pass.
  47     */
  48    struct util_dynarray phi_fixups;
  49 } write_ctx;
  50
  51 typedef struct {
  52    nir_shader *nir;
  53
  54    struct blob_reader *blob;
  55
  56    /* the next index to assign to a NIR in-memory object */
  57    uintptr_t next_idx;
  58
  59    /* The length of the index -> object table */
  60    uintptr_t idx_table_len;
  61
  62    /* map from index to deserialized pointer */
  63    void **idx_table;
  64
  65    /* List of phi sources. */
  66    struct list_head phi_srcs;
  67
  68 } read_ctx;
  69
  70 static void
  71 write_add_object(write_ctx *ctx, const void *obj)
  72 {
  73    uintptr_t index = ctx->next_idx++;
  74    _mesa_hash_table_insert(ctx->remap_table, obj, (void *) index);
  75 }
  76
  77 static uintptr_t
  78 write_lookup_object(write_ctx *ctx, const void *obj)
  79 {
  80    struct hash_entry *entry = _mesa_hash_table_search(ctx->remap_table, obj);
  81    assert(entry);
  82    return (uintptr_t) entry->data;
  83 }
  84
  85 static void
  86 write_object(write_ctx *ctx, const void *obj)
  87 {
  88    blob_write_intptr(ctx->blob, write_lookup_object(ctx, obj));
  89 }
  90
  91 static void
  92 read_add_object(read_ctx *ctx, void *obj)
  93 {
  94    assert(ctx->next_idx < ctx->idx_table_len);
  95    ctx->idx_table[ctx->next_idx++] = obj;
  96 }
  97
  98 static void *
  99 read_lookup_object(read_ctx *ctx, uintptr_t idx)
 100 {
 101    assert(idx < ctx->idx_table_len);
 102    return ctx->idx_table[idx];
 103 }
 104
 105 static void *
 106 read_object(read_ctx *ctx)
 107 {
 108    return read_lookup_object(ctx, blob_read_intptr(ctx->blob));
 109 }
 110
 111 static void
 112 write_constant(write_ctx *ctx, const nir_constant *c)
 113 {
 114    blob_write_bytes(ctx->blob, c->values, sizeof(c->values));
 115    blob_write_uint32(ctx->blob, c->num_elements);
 116    for (unsigned i = 0; i < c->num_elements; i++)
 117       write_constant(ctx, c->elements[i]);
 118 }
 119
 120 static nir_constant *
 121 read_constant(read_ctx *ctx, nir_variable *nvar)
 122 {
 123    nir_constant *c = ralloc(nvar, nir_constant);
 124
 125    blob_copy_bytes(ctx->blob, (uint8_t *)c->values, sizeof(c->values));
 126    c->num_elements = blob_read_uint32(ctx->blob);
 127    c->elements = ralloc_array(nvar, nir_constant *, c->num_elements);
 128    for (unsigned i = 0; i < c->num_elements; i++)
 129       c->elements[i] = read_constant(ctx, nvar);
 130
 131    return c;
 132 }
 133
 134 static void
 135 write_variable(write_ctx *ctx, const nir_variable *var)
 136 {
 137    write_add_object(ctx, var);
 138    encode_type_to_blob(ctx->blob, var->type);
 139    blob_write_uint32(ctx->blob, !!(var->name));
 140    if (var->name)
 141       blob_write_string(ctx->blob, var->name);
 142    blob_write_bytes(ctx->blob, (uint8_t *) &var->data, sizeof(var->data));
 143    blob_write_uint32(ctx->blob, var->num_state_slots);
 144    for (unsigned i = 0; i < var->num_state_slots; i++) {
 145       for (unsigned j = 0; j < STATE_LENGTH; j++)
 146          blob_write_uint32(ctx->blob, var->state_slots[i].tokens[j]);
 147       blob_write_uint32(ctx->blob, var->state_slots[i].swizzle);
 148    }
 149    blob_write_uint32(ctx->blob, !!(var->constant_initializer));
 150    if (var->constant_initializer)
 151       write_constant(ctx, var->constant_initializer);
 152    blob_write_uint32(ctx->blob, !!(var->interface_type));
 153    if (var->interface_type)
 154       encode_type_to_blob(ctx->blob, var->interface_type);
 155    blob_write_uint32(ctx->blob, var->num_members);
 156    if (var->num_members > 0) {
 157       blob_write_bytes(ctx->blob, (uint8_t *) var->members,
 158                        var->num_members * sizeof(*var->members));
 159    }
 160 }
 161
 162 static nir_variable *
 163 read_variable(read_ctx *ctx)
 164 {
 165    nir_variable *var = rzalloc(ctx->nir, nir_variable);
 166    read_add_object(ctx, var);
 167
 168    var->type = decode_type_from_blob(ctx->blob);
 169    bool has_name = blob_read_uint32(ctx->blob);
 170    if (has_name) {
 171       const char *name = blob_read_string(ctx->blob);
 172       var->name = ralloc_strdup(var, name);
 173    } else {
 174       var->name = NULL;
 175    }
 176    blob_copy_bytes(ctx->blob, (uint8_t *) &var->data, sizeof(var->data));
 177    var->num_state_slots = blob_read_uint32(ctx->blob);
 178    if (var->num_state_slots != 0) {
 179       var->state_slots = ralloc_array(var, nir_state_slot,
 180                                       var->num_state_slots);
 181       for (unsigned i = 0; i < var->num_state_slots; i++) {
 182          for (unsigned j = 0; j < STATE_LENGTH; j++)
 183             var->state_slots[i].tokens[j] = blob_read_uint32(ctx->blob);
 184          var->state_slots[i].swizzle = blob_read_uint32(ctx->blob);
 185       }
 186    }
 187    bool has_const_initializer = blob_read_uint32(ctx->blob);
 188    if (has_const_initializer)
 189       var->constant_initializer = read_constant(ctx, var);
 190    else
 191       var->constant_initializer = NULL;
 192    bool has_interface_type = blob_read_uint32(ctx->blob);
 193    if (has_interface_type)
 194       var->interface_type = decode_type_from_blob(ctx->blob);
 195    else
 196       var->interface_type = NULL;
 197    var->num_members = blob_read_uint32(ctx->blob);
 198    if (var->num_members > 0) {
 199       var->members = ralloc_array(var, struct nir_variable_data,
 200                                   var->num_members);
 201       blob_copy_bytes(ctx->blob, (uint8_t *) var->members,
 202                       var->num_members * sizeof(*var->members));
 203    }
 204
 205    return var;
 206 }
 207
 208 static void
 209 write_var_list(write_ctx *ctx, const struct exec_list *src)
 210 {
 211    blob_write_uint32(ctx->blob, exec_list_length(src));
 212    foreach_list_typed(nir_variable, var, node, src) {
 213       write_variable(ctx, var);
 214    }
 215 }
 216
 217 static void
 218 read_var_list(read_ctx *ctx, struct exec_list *dst)
 219 {
 220    exec_list_make_empty(dst);
 221    unsigned num_vars = blob_read_uint32(ctx->blob);
 222    for (unsigned i = 0; i < num_vars; i++) {
 223       nir_variable *var = read_variable(ctx);
 224       exec_list_push_tail(dst, &var->node);
 225    }
 226 }
 227
 228 static void
 229 write_register(write_ctx *ctx, const nir_register *reg)
 230 {
 231    write_add_object(ctx, reg);
 232    blob_write_uint32(ctx->blob, reg->num_components);
 233    blob_write_uint32(ctx->blob, reg->bit_size);
 234    blob_write_uint32(ctx->blob, reg->num_array_elems);
 235    blob_write_uint32(ctx->blob, reg->index);
 236    blob_write_uint32(ctx->blob, !!(reg->name));
 237    if (reg->name)
 238       blob_write_string(ctx->blob, reg->name);
 239 }
 240
 241 static nir_register *
 242 read_register(read_ctx *ctx)
 243 {
 244    nir_register *reg = ralloc(ctx->nir, nir_register);
 245    read_add_object(ctx, reg);
 246    reg->num_components = blob_read_uint32(ctx->blob);
 247    reg->bit_size = blob_read_uint32(ctx->blob);
 248    reg->num_array_elems = blob_read_uint32(ctx->blob);
 249    reg->index = blob_read_uint32(ctx->blob);
 250    bool has_name = blob_read_uint32(ctx->blob);
 251    if (has_name) {
 252       const char *name = blob_read_string(ctx->blob);
 253       reg->name = ralloc_strdup(reg, name);
 254    } else {
 255       reg->name = NULL;
 256    }
 257
 258    list_inithead(&reg->uses);
 259    list_inithead(&reg->defs);
 260    list_inithead(&reg->if_uses);
 261
 262    return reg;
 263 }
 264
 265 static void
 266 write_reg_list(write_ctx *ctx, const struct exec_list *src)
 267 {
 268    blob_write_uint32(ctx->blob, exec_list_length(src));
 269    foreach_list_typed(nir_register, reg, node, src)
 270       write_register(ctx, reg);
 271 }
 272
 273 static void
 274 read_reg_list(read_ctx *ctx, struct exec_list *dst)
 275 {
 276    exec_list_make_empty(dst);
 277    unsigned num_regs = blob_read_uint32(ctx->blob);
 278    for (unsigned i = 0; i < num_regs; i++) {
 279       nir_register *reg = read_register(ctx);
 280       exec_list_push_tail(dst, &reg->node);
 281    }
 282 }
 283
 284 static void
 285 write_src(write_ctx *ctx, const nir_src *src)
 286 {
 287    /* Since sources are very frequent, we try to save some space when storing
 288     * them. In particular, we store whether the source is a register and
 289     * whether the register has an indirect index in the low two bits. We can
 290     * assume that the high two bits of the index are zero, since otherwise our
 291     * address space would've been exhausted allocating the remap table!
 292     */
 293    if (src->is_ssa) {
 294       uintptr_t idx = write_lookup_object(ctx, src->ssa) << 2;
 295       idx |= 1;
 296       blob_write_intptr(ctx->blob, idx);
 297    } else {
 298       uintptr_t idx = write_lookup_object(ctx, src->reg.reg) << 2;
 299       if (src->reg.indirect)
 300          idx |= 2;
 301       blob_write_intptr(ctx->blob, idx);
 302       blob_write_uint32(ctx->blob, src->reg.base_offset);
 303       if (src->reg.indirect) {
 304          write_src(ctx, src->reg.indirect);
 305       }
 306    }
 307 }
 308
 309 static void
 310 read_src(read_ctx *ctx, nir_src *src, void *mem_ctx)
 311 {
 312    uintptr_t val = blob_read_intptr(ctx->blob);
 313    uintptr_t idx = val >> 2;
 314    src->is_ssa = val & 0x1;
 315    if (src->is_ssa) {
 316       src->ssa = read_lookup_object(ctx, idx);
 317    } else {
 318       bool is_indirect = val & 0x2;
 319       src->reg.reg = read_lookup_object(ctx, idx);
 320       src->reg.base_offset = blob_read_uint32(ctx->blob);
 321       if (is_indirect) {
 322          src->reg.indirect = ralloc(mem_ctx, nir_src);
 323          read_src(ctx, src->reg.indirect, mem_ctx);
 324       } else {
 325          src->reg.indirect = NULL;
 326       }
 327    }
 328 }
 329
 330 static void
 331 write_dest(write_ctx *ctx, const nir_dest *dst)
 332 {
 333    uint32_t val = dst->is_ssa;
 334    if (dst->is_ssa) {
 335       val |= !!(dst->ssa.name) << 1;
 336       val |= dst->ssa.num_components << 2;
 337       val |= dst->ssa.bit_size << 5;
 338    } else {
 339       val |= !!(dst->reg.indirect) << 1;
 340    }
 341    blob_write_uint32(ctx->blob, val);
 342    if (dst->is_ssa) {
 343       write_add_object(ctx, &dst->ssa);
 344       if (dst->ssa.name)
 345          blob_write_string(ctx->blob, dst->ssa.name);
 346    } else {
 347       blob_write_intptr(ctx->blob, write_lookup_object(ctx, dst->reg.reg));
 348       blob_write_uint32(ctx->blob, dst->reg.base_offset);
 349       if (dst->reg.indirect)
 350          write_src(ctx, dst->reg.indirect);
 351    }
 352 }
 353
 354 static void
 355 read_dest(read_ctx *ctx, nir_dest *dst, nir_instr *instr)
 356 {
 357    uint32_t val = blob_read_uint32(ctx->blob);
 358    bool is_ssa = val & 0x1;
 359    if (is_ssa) {
 360       bool has_name = val & 0x2;
 361       unsigned num_components = (val >> 2) & 0x7;
 362       unsigned bit_size = val >> 5;
 363       char *name = has_name ? blob_read_string(ctx->blob) : NULL;
 364       nir_ssa_dest_init(instr, dst, num_components, bit_size, name);
 365       read_add_object(ctx, &dst->ssa);
 366    } else {
 367       bool is_indirect = val & 0x2;
 368       dst->reg.reg = read_object(ctx);
 369       dst->reg.base_offset = blob_read_uint32(ctx->blob);
 370       if (is_indirect) {
 371          dst->reg.indirect = ralloc(instr, nir_src);
 372          read_src(ctx, dst->reg.indirect, instr);
 373       }
 374    }
 375 }
 376
 377 static void
 378 write_alu(write_ctx *ctx, const nir_alu_instr *alu)
 379 {
 380    blob_write_uint32(ctx->blob, alu->op);
 381    uint32_t flags = alu->exact;
 382    flags |= alu->dest.saturate << 1;
 383    flags |= alu->dest.write_mask << 2;
 384    blob_write_uint32(ctx->blob, flags);
 385
 386    write_dest(ctx, &alu->dest.dest);
 387
 388    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 389       write_src(ctx, &alu->src[i].src);
 390       flags = alu->src[i].negate;
 391       flags |= alu->src[i].abs << 1;
 392       for (unsigned j = 0; j < 4; j++)
 393          flags |= alu->src[i].swizzle[j] << (2 + 2 * j);
 394       blob_write_uint32(ctx->blob, flags);
 395    }
 396 }
 397
 398 static nir_alu_instr *
 399 read_alu(read_ctx *ctx)
 400 {
 401    nir_op op = blob_read_uint32(ctx->blob);
 402    nir_alu_instr *alu = nir_alu_instr_create(ctx->nir, op);
 403
 404    uint32_t flags = blob_read_uint32(ctx->blob);
 405    alu->exact = flags & 1;
 406    alu->dest.saturate = flags & 2;
 407    alu->dest.write_mask = flags >> 2;
 408
 409    read_dest(ctx, &alu->dest.dest, &alu->instr);
 410
 411    for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++) {
 412       read_src(ctx, &alu->src[i].src, &alu->instr);
 413       flags = blob_read_uint32(ctx->blob);
 414       alu->src[i].negate = flags & 1;
 415       alu->src[i].abs = flags & 2;
 416       for (unsigned j = 0; j < 4; j++)
 417          alu->src[i].swizzle[j] = (flags >> (2 * j + 2)) & 3;
 418    }
 419
 420    return alu;
 421 }
 422
 423 static void
 424 write_deref(write_ctx *ctx, const nir_deref_instr *deref)
 425 {
 426    blob_write_uint32(ctx->blob, deref->deref_type);
 427
 428    blob_write_uint32(ctx->blob, deref->mode);
 429    encode_type_to_blob(ctx->blob, deref->type);
 430
 431    write_dest(ctx, &deref->dest);
 432
 433    if (deref->deref_type == nir_deref_type_var) {
 434       write_object(ctx, deref->var);
 435       return;
 436    }
 437
 438    write_src(ctx, &deref->parent);
 439
 440    switch (deref->deref_type) {
 441    case nir_deref_type_struct:
 442       blob_write_uint32(ctx->blob, deref->strct.index);
 443       break;
 444
 445    case nir_deref_type_array:
 446    case nir_deref_type_ptr_as_array:
 447       write_src(ctx, &deref->arr.index);
 448       break;
 449
 450    case nir_deref_type_cast:
 451       blob_write_uint32(ctx->blob, deref->cast.ptr_stride);
 452       break;
 453
 454    case nir_deref_type_array_wildcard:
 455       /* Nothing to do */
 456       break;
 457
 458    default:
 459       unreachable("Invalid deref type");
 460    }
 461 }
 462
 463 static nir_deref_instr *
 464 read_deref(read_ctx *ctx)
 465 {
 466    nir_deref_type deref_type = blob_read_uint32(ctx->blob);
 467    nir_deref_instr *deref = nir_deref_instr_create(ctx->nir, deref_type);
 468
 469    deref->mode = blob_read_uint32(ctx->blob);
 470    deref->type = decode_type_from_blob(ctx->blob);
 471
 472    read_dest(ctx, &deref->dest, &deref->instr);
 473
 474    if (deref_type == nir_deref_type_var) {
 475       deref->var = read_object(ctx);
 476       return deref;
 477    }
 478
 479    read_src(ctx, &deref->parent, &deref->instr);
 480
 481    switch (deref->deref_type) {
 482    case nir_deref_type_struct:
 483       deref->strct.index = blob_read_uint32(ctx->blob);
 484       break;
 485
 486    case nir_deref_type_array:
 487    case nir_deref_type_ptr_as_array:
 488       read_src(ctx, &deref->arr.index, &deref->instr);
 489       break;
 490
 491    case nir_deref_type_cast:
 492       deref->cast.ptr_stride = blob_read_uint32(ctx->blob);
 493       break;
 494
 495    case nir_deref_type_array_wildcard:
 496       /* Nothing to do */
 497       break;
 498
 499    default:
 500       unreachable("Invalid deref type");
 501    }
 502
 503    return deref;
 504 }
 505
 506 static void
 507 write_intrinsic(write_ctx *ctx, const nir_intrinsic_instr *intrin)
 508 {
 509    blob_write_uint32(ctx->blob, intrin->intrinsic);
 510
 511    unsigned num_srcs = nir_intrinsic_infos[intrin->intrinsic].num_srcs;
 512    unsigned num_indices = nir_intrinsic_infos[intrin->intrinsic].num_indices;
 513
 514    blob_write_uint32(ctx->blob, intrin->num_components);
 515
 516    if (nir_intrinsic_infos[intrin->intrinsic].has_dest)
 517       write_dest(ctx, &intrin->dest);
 518
 519    for (unsigned i = 0; i < num_srcs; i++)
 520       write_src(ctx, &intrin->src[i]);
 521
 522    for (unsigned i = 0; i < num_indices; i++)
 523       blob_write_uint32(ctx->blob, intrin->const_index[i]);
 524 }
 525
 526 static nir_intrinsic_instr *
 527 read_intrinsic(read_ctx *ctx)
 528 {
 529    nir_intrinsic_op op = blob_read_uint32(ctx->blob);
 530
 531    nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(ctx->nir, op);
 532
 533    unsigned num_srcs = nir_intrinsic_infos[op].num_srcs;
 534    unsigned num_indices = nir_intrinsic_infos[op].num_indices;
 535
 536    intrin->num_components = blob_read_uint32(ctx->blob);
 537
 538    if (nir_intrinsic_infos[op].has_dest)
 539       read_dest(ctx, &intrin->dest, &intrin->instr);
 540
 541    for (unsigned i = 0; i < num_srcs; i++)
 542       read_src(ctx, &intrin->src[i], &intrin->instr);
 543
 544    for (unsigned i = 0; i < num_indices; i++)
 545       intrin->const_index[i] = blob_read_uint32(ctx->blob);
 546
 547    return intrin;
 548 }
 549
 550 static void
 551 write_load_const(write_ctx *ctx, const nir_load_const_instr *lc)
 552 {
 553    uint32_t val = lc->def.num_components;
 554    val |= lc->def.bit_size << 3;
 555    blob_write_uint32(ctx->blob, val);
 556    blob_write_bytes(ctx->blob, lc->value, sizeof(*lc->value) * lc->def.num_components);
 557    write_add_object(ctx, &lc->def);
 558 }
 559
 560 static nir_load_const_instr *
 561 read_load_const(read_ctx *ctx)
 562 {
 563    uint32_t val = blob_read_uint32(ctx->blob);
 564
 565    nir_load_const_instr *lc =
 566       nir_load_const_instr_create(ctx->nir, val & 0x7, val >> 3);
 567
 568    blob_copy_bytes(ctx->blob, lc->value, sizeof(*lc->value) * lc->def.num_components);
 569    read_add_object(ctx, &lc->def);
 570    return lc;
 571 }
 572
 573 static void
 574 write_ssa_undef(write_ctx *ctx, const nir_ssa_undef_instr *undef)
 575 {
 576    uint32_t val = undef->def.num_components;
 577    val |= undef->def.bit_size << 3;
 578    blob_write_uint32(ctx->blob, val);
 579    write_add_object(ctx, &undef->def);
 580 }
 581
 582 static nir_ssa_undef_instr *
 583 read_ssa_undef(read_ctx *ctx)
 584 {
 585    uint32_t val = blob_read_uint32(ctx->blob);
 586
 587    nir_ssa_undef_instr *undef =
 588       nir_ssa_undef_instr_create(ctx->nir, val & 0x7, val >> 3);
 589
 590    read_add_object(ctx, &undef->def);
 591    return undef;
 592 }
 593
 594 union packed_tex_data {
 595    uint32_t u32;
 596    struct {
 597       enum glsl_sampler_dim sampler_dim:4;
 598       nir_alu_type dest_type:8;
 599       unsigned coord_components:3;
 600       unsigned is_array:1;
 601       unsigned is_shadow:1;
 602       unsigned is_new_style_shadow:1;
 603       unsigned component:2;
 604       unsigned unused:10; /* Mark unused for valgrind. */
 605    } u;
 606 };
 607
 608 static void
 609 write_tex(write_ctx *ctx, const nir_tex_instr *tex)
 610 {
 611    blob_write_uint32(ctx->blob, tex->num_srcs);
 612    blob_write_uint32(ctx->blob, tex->op);
 613    blob_write_uint32(ctx->blob, tex->texture_index);
 614    blob_write_uint32(ctx->blob, tex->texture_array_size);
 615    blob_write_uint32(ctx->blob, tex->sampler_index);
 616    blob_write_bytes(ctx->blob, tex->tg4_offsets, sizeof(tex->tg4_offsets));
 617
 618    STATIC_ASSERT(sizeof(union packed_tex_data) == sizeof(uint32_t));
 619    union packed_tex_data packed = {
 620       .u.sampler_dim = tex->sampler_dim,
 621       .u.dest_type = tex->dest_type,
 622       .u.coord_components = tex->coord_components,
 623       .u.is_array = tex->is_array,
 624       .u.is_shadow = tex->is_shadow,
 625       .u.is_new_style_shadow = tex->is_new_style_shadow,
 626       .u.component = tex->component,
 627    };
 628    blob_write_uint32(ctx->blob, packed.u32);
 629
 630    write_dest(ctx, &tex->dest);
 631    for (unsigned i = 0; i < tex->num_srcs; i++) {
 632       blob_write_uint32(ctx->blob, tex->src[i].src_type);
 633       write_src(ctx, &tex->src[i].src);
 634    }
 635 }
 636
 637 static nir_tex_instr *
 638 read_tex(read_ctx *ctx)
 639 {
 640    unsigned num_srcs = blob_read_uint32(ctx->blob);
 641    nir_tex_instr *tex = nir_tex_instr_create(ctx->nir, num_srcs);
 642
 643    tex->op = blob_read_uint32(ctx->blob);
 644    tex->texture_index = blob_read_uint32(ctx->blob);
 645    tex->texture_array_size = blob_read_uint32(ctx->blob);
 646    tex->sampler_index = blob_read_uint32(ctx->blob);
 647    blob_copy_bytes(ctx->blob, tex->tg4_offsets, sizeof(tex->tg4_offsets));
 648
 649    union packed_tex_data packed;
 650    packed.u32 = blob_read_uint32(ctx->blob);
 651    tex->sampler_dim = packed.u.sampler_dim;
 652    tex->dest_type = packed.u.dest_type;
 653    tex->coord_components = packed.u.coord_components;
 654    tex->is_array = packed.u.is_array;
 655    tex->is_shadow = packed.u.is_shadow;
 656    tex->is_new_style_shadow = packed.u.is_new_style_shadow;
 657    tex->component = packed.u.component;
 658
 659    read_dest(ctx, &tex->dest, &tex->instr);
 660    for (unsigned i = 0; i < tex->num_srcs; i++) {
 661       tex->src[i].src_type = blob_read_uint32(ctx->blob);
 662       read_src(ctx, &tex->src[i].src, &tex->instr);
 663    }
 664
 665    return tex;
 666 }
 667
 668 static void
 669 write_phi(write_ctx *ctx, const nir_phi_instr *phi)
 670 {
 671    /* Phi nodes are special, since they may reference SSA definitions and
 672     * basic blocks that don't exist yet. We leave two empty uintptr_t's here,
 673     * and then store enough information so that a later fixup pass can fill
 674     * them in correctly.
 675     */
 676    write_dest(ctx, &phi->dest);
 677
 678    blob_write_uint32(ctx->blob, exec_list_length(&phi->srcs));
 679
 680    nir_foreach_phi_src(src, phi) {
 681       assert(src->src.is_ssa);
 682       size_t blob_offset = blob_reserve_intptr(ctx->blob);
 683       MAYBE_UNUSED size_t blob_offset2 = blob_reserve_intptr(ctx->blob);
 684       assert(blob_offset + sizeof(uintptr_t) == blob_offset2);
 685       write_phi_fixup fixup = {
 686          .blob_offset = blob_offset,
 687          .src = src->src.ssa,
 688          .block = src->pred,
 689       };
 690       util_dynarray_append(&ctx->phi_fixups, write_phi_fixup, fixup);
 691    }
 692 }
 693
 694 static void
 695 write_fixup_phis(write_ctx *ctx)
 696 {
 697    util_dynarray_foreach(&ctx->phi_fixups, write_phi_fixup, fixup) {
 698       uintptr_t *blob_ptr = (uintptr_t *)(ctx->blob->data + fixup->blob_offset);
 699       blob_ptr[0] = write_lookup_object(ctx, fixup->src);
 700       blob_ptr[1] = write_lookup_object(ctx, fixup->block);
 701    }
 702
 703    util_dynarray_clear(&ctx->phi_fixups);
 704 }
 705
 706 static nir_phi_instr *
 707 read_phi(read_ctx *ctx, nir_block *blk)
 708 {
 709    nir_phi_instr *phi = nir_phi_instr_create(ctx->nir);
 710
 711    read_dest(ctx, &phi->dest, &phi->instr);
 712
 713    unsigned num_srcs = blob_read_uint32(ctx->blob);
 714
 715    /* For similar reasons as before, we just store the index directly into the
 716     * pointer, and let a later pass resolve the phi sources.
 717     *
 718     * In order to ensure that the copied sources (which are just the indices
 719     * from the blob for now) don't get inserted into the old shader's use-def
 720     * lists, we have to add the phi instruction *before* we set up its
 721     * sources.
 722     */
 723    nir_instr_insert_after_block(blk, &phi->instr);
 724
 725    for (unsigned i = 0; i < num_srcs; i++) {
 726       nir_phi_src *src = ralloc(phi, nir_phi_src);
 727
 728       src->src.is_ssa = true;
 729       src->src.ssa = (nir_ssa_def *) blob_read_intptr(ctx->blob);
 730       src->pred = (nir_block *) blob_read_intptr(ctx->blob);
 731
 732       /* Since we're not letting nir_insert_instr handle use/def stuff for us,
 733        * we have to set the parent_instr manually.  It doesn't really matter
 734        * when we do it, so we might as well do it here.
 735        */
 736       src->src.parent_instr = &phi->instr;
 737
 738       /* Stash it in the list of phi sources.  We'll walk this list and fix up
 739        * sources at the very end of read_function_impl.
 740        */
 741       list_add(&src->src.use_link, &ctx->phi_srcs);
 742
 743       exec_list_push_tail(&phi->srcs, &src->node);
 744    }
 745
 746    return phi;
 747 }
 748
 749 static void
 750 read_fixup_phis(read_ctx *ctx)
 751 {
 752    list_for_each_entry_safe(nir_phi_src, src, &ctx->phi_srcs, src.use_link) {
 753       src->pred = read_lookup_object(ctx, (uintptr_t)src->pred);
 754       src->src.ssa = read_lookup_object(ctx, (uintptr_t)src->src.ssa);
 755
 756       /* Remove from this list */
 757       list_del(&src->src.use_link);
 758
 759       list_addtail(&src->src.use_link, &src->src.ssa->uses);
 760    }
 761    assert(list_empty(&ctx->phi_srcs));
 762 }
 763
 764 static void
 765 write_jump(write_ctx *ctx, const nir_jump_instr *jmp)
 766 {
 767    blob_write_uint32(ctx->blob, jmp->type);
 768 }
 769
 770 static nir_jump_instr *
 771 read_jump(read_ctx *ctx)
 772 {
 773    nir_jump_type type = blob_read_uint32(ctx->blob);
 774    nir_jump_instr *jmp = nir_jump_instr_create(ctx->nir, type);
 775    return jmp;
 776 }
 777
 778 static void
 779 write_call(write_ctx *ctx, const nir_call_instr *call)
 780 {
 781    blob_write_intptr(ctx->blob, write_lookup_object(ctx, call->callee));
 782
 783    for (unsigned i = 0; i < call->num_params; i++)
 784       write_src(ctx, &call->params[i]);
 785 }
 786
 787 static nir_call_instr *
 788 read_call(read_ctx *ctx)
 789 {
 790    nir_function *callee = read_object(ctx);
 791    nir_call_instr *call = nir_call_instr_create(ctx->nir, callee);
 792
 793    for (unsigned i = 0; i < call->num_params; i++)
 794       read_src(ctx, &call->params[i], call);
 795
 796    return call;
 797 }
 798
 799 static void
 800 write_instr(write_ctx *ctx, const nir_instr *instr)
 801 {
 802    blob_write_uint32(ctx->blob, instr->type);
 803    switch (instr->type) {
 804    case nir_instr_type_alu:
 805       write_alu(ctx, nir_instr_as_alu(instr));
 806       break;
 807    case nir_instr_type_deref:
 808       write_deref(ctx, nir_instr_as_deref(instr));
 809       break;
 810    case nir_instr_type_intrinsic:
 811       write_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 812       break;
 813    case nir_instr_type_load_const:
 814       write_load_const(ctx, nir_instr_as_load_const(instr));
 815       break;
 816    case nir_instr_type_ssa_undef:
 817       write_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
 818       break;
 819    case nir_instr_type_tex:
 820       write_tex(ctx, nir_instr_as_tex(instr));
 821       break;
 822    case nir_instr_type_phi:
 823       write_phi(ctx, nir_instr_as_phi(instr));
 824       break;
 825    case nir_instr_type_jump:
 826       write_jump(ctx, nir_instr_as_jump(instr));
 827       break;
 828    case nir_instr_type_call:
 829       write_call(ctx, nir_instr_as_call(instr));
 830       break;
 831    case nir_instr_type_parallel_copy:
 832       unreachable("Cannot write parallel copies");
 833    default:
 834       unreachable("bad instr type");
 835    }
 836 }
 837
 838 static void
 839 read_instr(read_ctx *ctx, nir_block *block)
 840 {
 841    nir_instr_type type = blob_read_uint32(ctx->blob);
 842    nir_instr *instr;
 843    switch (type) {
 844    case nir_instr_type_alu:
 845       instr = &read_alu(ctx)->instr;
 846       break;
 847    case nir_instr_type_deref:
 848       instr = &read_deref(ctx)->instr;
 849       break;
 850    case nir_instr_type_intrinsic:
 851       instr = &read_intrinsic(ctx)->instr;
 852       break;
 853    case nir_instr_type_load_const:
 854       instr = &read_load_const(ctx)->instr;
 855       break;
 856    case nir_instr_type_ssa_undef:
 857       instr = &read_ssa_undef(ctx)->instr;
 858       break;
 859    case nir_instr_type_tex:
 860       instr = &read_tex(ctx)->instr;
 861       break;
 862    case nir_instr_type_phi:
 863       /* Phi instructions are a bit of a special case when reading because we
 864        * don't want inserting the instruction to automatically handle use/defs
 865        * for us.  Instead, we need to wait until all the blocks/instructions
 866        * are read so that we can set their sources up.
 867        */
 868       read_phi(ctx, block);
 869       return;
 870    case nir_instr_type_jump:
 871       instr = &read_jump(ctx)->instr;
 872       break;
 873    case nir_instr_type_call:
 874       instr = &read_call(ctx)->instr;
 875       break;
 876    case nir_instr_type_parallel_copy:
 877       unreachable("Cannot read parallel copies");
 878    default:
 879       unreachable("bad instr type");
 880    }
 881
 882    nir_instr_insert_after_block(block, instr);
 883 }
 884
 885 static void
 886 write_block(write_ctx *ctx, const nir_block *block)
 887 {
 888    write_add_object(ctx, block);
 889    blob_write_uint32(ctx->blob, exec_list_length(&block->instr_list));
 890    nir_foreach_instr(instr, block)
 891       write_instr(ctx, instr);
 892 }
 893
 894 static void
 895 read_block(read_ctx *ctx, struct exec_list *cf_list)
 896 {
 897    /* Don't actually create a new block.  Just use the one from the tail of
 898     * the list.  NIR guarantees that the tail of the list is a block and that
 899     * no two blocks are side-by-side in the IR;  It should be empty.
 900     */
 901    nir_block *block =
 902       exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node);
 903
 904    read_add_object(ctx, block);
 905    unsigned num_instrs = blob_read_uint32(ctx->blob);
 906    for (unsigned i = 0; i < num_instrs; i++) {
 907       read_instr(ctx, block);
 908    }
 909 }
 910
 911 static void
 912 write_cf_list(write_ctx *ctx, const struct exec_list *cf_list);
 913
 914 static void
 915 read_cf_list(read_ctx *ctx, struct exec_list *cf_list);
 916
 917 static void
 918 write_if(write_ctx *ctx, nir_if *nif)
 919 {
 920    write_src(ctx, &nif->condition);
 921
 922    write_cf_list(ctx, &nif->then_list);
 923    write_cf_list(ctx, &nif->else_list);
 924 }
 925
 926 static void
 927 read_if(read_ctx *ctx, struct exec_list *cf_list)
 928 {
 929    nir_if *nif = nir_if_create(ctx->nir);
 930
 931    read_src(ctx, &nif->condition, nif);
 932
 933    nir_cf_node_insert_end(cf_list, &nif->cf_node);
 934
 935    read_cf_list(ctx, &nif->then_list);
 936    read_cf_list(ctx, &nif->else_list);
 937 }
 938
 939 static void
 940 write_loop(write_ctx *ctx, nir_loop *loop)
 941 {
 942    write_cf_list(ctx, &loop->body);
 943 }
 944
 945 static void
 946 read_loop(read_ctx *ctx, struct exec_list *cf_list)
 947 {
 948    nir_loop *loop = nir_loop_create(ctx->nir);
 949
 950    nir_cf_node_insert_end(cf_list, &loop->cf_node);
 951
 952    read_cf_list(ctx, &loop->body);
 953 }
 954
 955 static void
 956 write_cf_node(write_ctx *ctx, nir_cf_node *cf)
 957 {
 958    blob_write_uint32(ctx->blob, cf->type);
 959
 960    switch (cf->type) {
 961    case nir_cf_node_block:
 962       write_block(ctx, nir_cf_node_as_block(cf));
 963       break;
 964    case nir_cf_node_if:
 965       write_if(ctx, nir_cf_node_as_if(cf));
 966       break;
 967    case nir_cf_node_loop:
 968       write_loop(ctx, nir_cf_node_as_loop(cf));
 969       break;
 970    default:
 971       unreachable("bad cf type");
 972    }
 973 }
 974
 975 static void
 976 read_cf_node(read_ctx *ctx, struct exec_list *list)
 977 {
 978    nir_cf_node_type type = blob_read_uint32(ctx->blob);
 979
 980    switch (type) {
 981    case nir_cf_node_block:
 982       read_block(ctx, list);
 983       break;
 984    case nir_cf_node_if:
 985       read_if(ctx, list);
 986       break;
 987    case nir_cf_node_loop:
 988       read_loop(ctx, list);
 989       break;
 990    default:
 991       unreachable("bad cf type");
 992    }
 993 }
 994
 995 static void
 996 write_cf_list(write_ctx *ctx, const struct exec_list *cf_list)
 997 {
 998    blob_write_uint32(ctx->blob, exec_list_length(cf_list));
 999    foreach_list_typed(nir_cf_node, cf, node, cf_list) {
1000       write_cf_node(ctx, cf);
1001    }
1002 }
1003
1004 static void
1005 read_cf_list(read_ctx *ctx, struct exec_list *cf_list)
1006 {
1007    uint32_t num_cf_nodes = blob_read_uint32(ctx->blob);
1008    for (unsigned i = 0; i < num_cf_nodes; i++)
1009       read_cf_node(ctx, cf_list);
1010 }
1011
1012 static void
1013 write_function_impl(write_ctx *ctx, const nir_function_impl *fi)
1014 {
1015    write_var_list(ctx, &fi->locals);
1016    write_reg_list(ctx, &fi->registers);
1017    blob_write_uint32(ctx->blob, fi->reg_alloc);
1018
1019    write_cf_list(ctx, &fi->body);
1020    write_fixup_phis(ctx);
1021 }
1022
1023 static nir_function_impl *
1024 read_function_impl(read_ctx *ctx, nir_function *fxn)
1025 {
1026    nir_function_impl *fi = nir_function_impl_create_bare(ctx->nir);
1027    fi->function = fxn;
1028
1029    read_var_list(ctx, &fi->locals);
1030    read_reg_list(ctx, &fi->registers);
1031    fi->reg_alloc = blob_read_uint32(ctx->blob);
1032
1033    read_cf_list(ctx, &fi->body);
1034    read_fixup_phis(ctx);
1035
1036    fi->valid_metadata = 0;
1037
1038    return fi;
1039 }
1040
1041 static void
1042 write_function(write_ctx *ctx, const nir_function *fxn)
1043 {
1044    blob_write_uint32(ctx->blob, !!(fxn->name));
1045    if (fxn->name)
1046       blob_write_string(ctx->blob, fxn->name);
1047
1048    write_add_object(ctx, fxn);
1049
1050    blob_write_uint32(ctx->blob, fxn->num_params);
1051    for (unsigned i = 0; i < fxn->num_params; i++) {
1052       uint32_t val =
1053          ((uint32_t)fxn->params[i].num_components) |
1054          ((uint32_t)fxn->params[i].bit_size) << 8;
1055       blob_write_uint32(ctx->blob, val);
1056    }
1057
1058    blob_write_uint32(ctx->blob, fxn->is_entrypoint);
1059
1060    /* At first glance, it looks like we should write the function_impl here.
1061     * However, call instructions need to be able to reference at least the
1062     * function and those will get processed as we write the function_impls.
1063     * We stop here and write function_impls as a second pass.
1064     */
1065 }
1066
1067 static void
1068 read_function(read_ctx *ctx)
1069 {
1070    bool has_name = blob_read_uint32(ctx->blob);
1071    char *name = has_name ? blob_read_string(ctx->blob) : NULL;
1072
1073    nir_function *fxn = nir_function_create(ctx->nir, name);
1074
1075    read_add_object(ctx, fxn);
1076
1077    fxn->num_params = blob_read_uint32(ctx->blob);
1078    fxn->params = ralloc_array(fxn, nir_parameter, fxn->num_params);
1079    for (unsigned i = 0; i < fxn->num_params; i++) {
1080       uint32_t val = blob_read_uint32(ctx->blob);
1081       fxn->params[i].num_components = val & 0xff;
1082       fxn->params[i].bit_size = (val >> 8) & 0xff;
1083    }
1084
1085    fxn->is_entrypoint = blob_read_uint32(ctx->blob);
1086 }
1087
1088 void
1089 nir_serialize(struct blob *blob, const nir_shader *nir)
1090 {
1091    write_ctx ctx;
1092    ctx.remap_table = _mesa_pointer_hash_table_create(NULL);
1093    ctx.next_idx = 0;
1094    ctx.blob = blob;
1095    ctx.nir = nir;
1096    util_dynarray_init(&ctx.phi_fixups, NULL);
1097
1098    size_t idx_size_offset = blob_reserve_intptr(blob);
1099
1100    struct shader_info info = nir->info;
1101    uint32_t strings = 0;
1102    if (info.name)
1103       strings |= 0x1;
1104    if (info.label)
1105       strings |= 0x2;
1106    blob_write_uint32(blob, strings);
1107    if (info.name)
1108       blob_write_string(blob, info.name);
1109    if (info.label)
1110       blob_write_string(blob, info.label);
1111    info.name = info.label = NULL;
1112    blob_write_bytes(blob, (uint8_t *) &info, sizeof(info));
1113
1114    write_var_list(&ctx, &nir->uniforms);
1115    write_var_list(&ctx, &nir->inputs);
1116    write_var_list(&ctx, &nir->outputs);
1117    write_var_list(&ctx, &nir->shared);
1118    write_var_list(&ctx, &nir->globals);
1119    write_var_list(&ctx, &nir->system_values);
1120
1121    blob_write_uint32(blob, nir->num_inputs);
1122    blob_write_uint32(blob, nir->num_uniforms);
1123    blob_write_uint32(blob, nir->num_outputs);
1124    blob_write_uint32(blob, nir->num_shared);
1125    blob_write_uint32(blob, nir->scratch_size);
1126
1127    blob_write_uint32(blob, exec_list_length(&nir->functions));
1128    nir_foreach_function(fxn, nir) {
1129       write_function(&ctx, fxn);
1130    }
1131
1132    nir_foreach_function(fxn, nir) {
1133       write_function_impl(&ctx, fxn->impl);
1134    }
1135
1136    blob_write_uint32(blob, nir->constant_data_size);
1137    if (nir->constant_data_size > 0)
1138       blob_write_bytes(blob, nir->constant_data, nir->constant_data_size);
1139
1140    *(uintptr_t *)(blob->data + idx_size_offset) = ctx.next_idx;
1141
1142    _mesa_hash_table_destroy(ctx.remap_table, NULL);
1143    util_dynarray_fini(&ctx.phi_fixups);
1144 }
1145
1146 nir_shader *
1147 nir_deserialize(void *mem_ctx,
1148                 const struct nir_shader_compiler_options *options,
1149                 struct blob_reader *blob)
1150 {
1151    read_ctx ctx;
1152    ctx.blob = blob;
1153    list_inithead(&ctx.phi_srcs);
1154    ctx.idx_table_len = blob_read_intptr(blob);
1155    ctx.idx_table = calloc(ctx.idx_table_len, sizeof(uintptr_t));
1156    ctx.next_idx = 0;
1157
1158    uint32_t strings = blob_read_uint32(blob);
1159    char *name = (strings & 0x1) ? blob_read_string(blob) : NULL;
1160    char *label = (strings & 0x2) ? blob_read_string(blob) : NULL;
1161
1162    struct shader_info info;
1163    blob_copy_bytes(blob, (uint8_t *) &info, sizeof(info));
1164
1165    ctx.nir = nir_shader_create(mem_ctx, info.stage, options, NULL);
1166
1167    info.name = name ? ralloc_strdup(ctx.nir, name) : NULL;
1168    info.label = label ? ralloc_strdup(ctx.nir, label) : NULL;
1169
1170    ctx.nir->info = info;
1171
1172    read_var_list(&ctx, &ctx.nir->uniforms);
1173    read_var_list(&ctx, &ctx.nir->inputs);
1174    read_var_list(&ctx, &ctx.nir->outputs);
1175    read_var_list(&ctx, &ctx.nir->shared);
1176    read_var_list(&ctx, &ctx.nir->globals);
1177    read_var_list(&ctx, &ctx.nir->system_values);
1178
1179    ctx.nir->num_inputs = blob_read_uint32(blob);
1180    ctx.nir->num_uniforms = blob_read_uint32(blob);
1181    ctx.nir->num_outputs = blob_read_uint32(blob);
1182    ctx.nir->num_shared = blob_read_uint32(blob);
1183    ctx.nir->scratch_size = blob_read_uint32(blob);
1184
1185    unsigned num_functions = blob_read_uint32(blob);
1186    for (unsigned i = 0; i < num_functions; i++)
1187       read_function(&ctx);
1188
1189    nir_foreach_function(fxn, ctx.nir)
1190       fxn->impl = read_function_impl(&ctx, fxn);
1191
1192    ctx.nir->constant_data_size = blob_read_uint32(blob);
1193    if (ctx.nir->constant_data_size > 0) {
1194       ctx.nir->constant_data =
1195          ralloc_size(ctx.nir, ctx.nir->constant_data_size);
1196       blob_copy_bytes(blob, ctx.nir->constant_data,
1197                       ctx.nir->constant_data_size);
1198    }
1199
1200    free(ctx.idx_table);
1201
1202    return ctx.nir;
1203 }
1204
1205 void
1206 nir_shader_serialize_deserialize(nir_shader *shader)
1207 {
1208    const struct nir_shader_compiler_options *options = shader->options;
1209
1210    struct blob writer;
1211    blob_init(&writer);
1212    nir_serialize(&writer, shader);
1213
1214    /* Delete all of dest's ralloc children but leave dest alone */
1215    void *dead_ctx = ralloc_context(NULL);
1216    ralloc_adopt(dead_ctx, shader);
1217    ralloc_free(dead_ctx);
1218
1219    dead_ctx = ralloc_context(NULL);
1220
1221    struct blob_reader reader;
1222    blob_reader_init(&reader, writer.data, writer.size);
1223    nir_shader *copy = nir_deserialize(dead_ctx, options, &reader);
1224
1225    blob_finish(&writer);
1226
1227    nir_shader_replace(shader, copy);
1228    ralloc_free(dead_ctx);
1229 }