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