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