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(ctx->nir, 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    blob_write_string(ctx->blob, var->name);
 141    blob_write_bytes(ctx->blob, (uint8_t *) &var->data, sizeof(var->data));
 142    blob_write_uint32(ctx->blob, var->num_state_slots);
 143    blob_write_bytes(ctx->blob, (uint8_t *) var->state_slots,
 144                     var->num_state_slots * sizeof(nir_state_slot));
 145    blob_write_uint32(ctx->blob, !!(var->constant_initializer));
 146    if (var->constant_initializer)
 147       write_constant(ctx, var->constant_initializer);
 148    blob_write_uint32(ctx->blob, !!(var->interface_type));
 149    if (var->interface_type)
 150       encode_type_to_blob(ctx->blob, var->interface_type);
 151 }
 152
 153 static nir_variable *
 154 read_variable(read_ctx *ctx)
 155 {
 156    nir_variable *var = rzalloc(ctx->nir, nir_variable);
 157    read_add_object(ctx, var);
 158
 159    var->type = decode_type_from_blob(ctx->blob);
 160    bool has_name = blob_read_uint32(ctx->blob);
 161    if (has_name) {
 162       const char *name = blob_read_string(ctx->blob);
 163       var->name = ralloc_strdup(var, name);
 164    } else {
 165       var->name = NULL;
 166    }
 167    blob_copy_bytes(ctx->blob, (uint8_t *) &var->data, sizeof(var->data));
 168    var->num_state_slots = blob_read_uint32(ctx->blob);
 169    var->state_slots = ralloc_array(var, nir_state_slot, var->num_state_slots);
 170    blob_copy_bytes(ctx->blob, (uint8_t *) var->state_slots,
 171                    var->num_state_slots * sizeof(nir_state_slot));
 172    bool has_const_initializer = blob_read_uint32(ctx->blob);
 173    if (has_const_initializer)
 174       var->constant_initializer = read_constant(ctx, var);
 175    else
 176       var->constant_initializer = NULL;
 177    bool has_interface_type = blob_read_uint32(ctx->blob);
 178    if (has_interface_type)
 179       var->interface_type = decode_type_from_blob(ctx->blob);
 180    else
 181       var->interface_type = NULL;
 182
 183    return var;
 184 }
 185
 186 static void
 187 write_var_list(write_ctx *ctx, const struct exec_list *src)
 188 {
 189    blob_write_uint32(ctx->blob, exec_list_length(src));
 190    foreach_list_typed(nir_variable, var, node, src) {
 191       write_variable(ctx, var);
 192    }
 193 }
 194
 195 static void
 196 read_var_list(read_ctx *ctx, struct exec_list *dst)
 197 {
 198    exec_list_make_empty(dst);
 199    unsigned num_vars = blob_read_uint32(ctx->blob);
 200    for (unsigned i = 0; i < num_vars; i++) {
 201       nir_variable *var = read_variable(ctx);
 202       exec_list_push_tail(dst, &var->node);
 203    }
 204 }
 205
 206 static void
 207 write_register(write_ctx *ctx, const nir_register *reg)
 208 {
 209    write_add_object(ctx, reg);
 210    blob_write_uint32(ctx->blob, reg->num_components);
 211    blob_write_uint32(ctx->blob, reg->bit_size);
 212    blob_write_uint32(ctx->blob, reg->num_array_elems);
 213    blob_write_uint32(ctx->blob, reg->index);
 214    blob_write_uint32(ctx->blob, !!(reg->name));
 215    if (reg->name)
 216       blob_write_string(ctx->blob, reg->name);
 217    blob_write_uint32(ctx->blob, reg->is_global << 1 | reg->is_packed);
 218 }
 219
 220 static nir_register *
 221 read_register(read_ctx *ctx)
 222 {
 223    nir_register *reg = ralloc(ctx->nir, nir_register);
 224    read_add_object(ctx, reg);
 225    reg->num_components = blob_read_uint32(ctx->blob);
 226    reg->bit_size = blob_read_uint32(ctx->blob);
 227    reg->num_array_elems = blob_read_uint32(ctx->blob);
 228    reg->index = blob_read_uint32(ctx->blob);
 229    bool has_name = blob_read_uint32(ctx->blob);
 230    if (has_name) {
 231       const char *name = blob_read_string(ctx->blob);
 232       reg->name = ralloc_strdup(reg, name);
 233    } else {
 234       reg->name = NULL;
 235    }
 236    unsigned flags = blob_read_uint32(ctx->blob);
 237    reg->is_global = flags & 0x2;
 238    reg->is_packed = flags & 0x1;
 239
 240    list_inithead(&reg->uses);
 241    list_inithead(&reg->defs);
 242    list_inithead(&reg->if_uses);
 243
 244    return reg;
 245 }
 246
 247 static void
 248 write_reg_list(write_ctx *ctx, const struct exec_list *src)
 249 {
 250    blob_write_uint32(ctx->blob, exec_list_length(src));
 251    foreach_list_typed(nir_register, reg, node, src)
 252       write_register(ctx, reg);
 253 }
 254
 255 static void
 256 read_reg_list(read_ctx *ctx, struct exec_list *dst)
 257 {
 258    exec_list_make_empty(dst);
 259    unsigned num_regs = blob_read_uint32(ctx->blob);
 260    for (unsigned i = 0; i < num_regs; i++) {
 261       nir_register *reg = read_register(ctx);
 262       exec_list_push_tail(dst, &reg->node);
 263    }
 264 }
 265
 266 static void
 267 write_src(write_ctx *ctx, const nir_src *src)
 268 {
 269    /* Since sources are very frequent, we try to save some space when storing
 270     * them. In particular, we store whether the source is a register and
 271     * whether the register has an indirect index in the low two bits. We can
 272     * assume that the high two bits of the index are zero, since otherwise our
 273     * address space would've been exhausted allocating the remap table!
 274     */
 275    if (src->is_ssa) {
 276       uintptr_t idx = write_lookup_object(ctx, src->ssa) << 2;
 277       idx |= 1;
 278       blob_write_intptr(ctx->blob, idx);
 279    } else {
 280       uintptr_t idx = write_lookup_object(ctx, src->reg.reg) << 2;
 281       if (src->reg.indirect)
 282          idx |= 2;
 283       blob_write_intptr(ctx->blob, idx);
 284       blob_write_uint32(ctx->blob, src->reg.base_offset);
 285       if (src->reg.indirect) {
 286          write_src(ctx, src->reg.indirect);
 287       }
 288    }
 289 }
 290
 291 static void
 292 read_src(read_ctx *ctx, nir_src *src, void *mem_ctx)
 293 {
 294    uintptr_t val = blob_read_intptr(ctx->blob);
 295    uintptr_t idx = val >> 2;
 296    src->is_ssa = val & 0x1;
 297    if (src->is_ssa) {
 298       src->ssa = read_lookup_object(ctx, idx);
 299    } else {
 300       bool is_indirect = val & 0x2;
 301       src->reg.reg = read_lookup_object(ctx, idx);
 302       src->reg.base_offset = blob_read_uint32(ctx->blob);
 303       if (is_indirect) {
 304          src->reg.indirect = ralloc(mem_ctx, nir_src);
 305          read_src(ctx, src->reg.indirect, mem_ctx);
 306       } else {
 307          src->reg.indirect = NULL;
 308       }
 309    }
 310 }
 311
 312 static void
 313 write_dest(write_ctx *ctx, const nir_dest *dst)
 314 {
 315    uint32_t val = dst->is_ssa;
 316    if (dst->is_ssa) {
 317       val |= !!(dst->ssa.name) << 1;
 318       val |= dst->ssa.num_components << 2;
 319       val |= dst->ssa.bit_size << 5;
 320    } else {
 321       val |= !!(dst->reg.indirect) << 1;
 322    }
 323    blob_write_uint32(ctx->blob, val);
 324    if (dst->is_ssa) {
 325       write_add_object(ctx, &dst->ssa);
 326       if (dst->ssa.name)
 327          blob_write_string(ctx->blob, dst->ssa.name);
 328    } else {
 329       blob_write_intptr(ctx->blob, write_lookup_object(ctx, dst->reg.reg));
 330       blob_write_uint32(ctx->blob, dst->reg.base_offset);
 331       if (dst->reg.indirect)
 332          write_src(ctx, dst->reg.indirect);
 333    }
 334 }
 335
 336 static void
 337 read_dest(read_ctx *ctx, nir_dest *dst, nir_instr *instr)
 338 {
 339    uint32_t val = blob_read_uint32(ctx->blob);
 340    bool is_ssa = val & 0x1;
 341    if (is_ssa) {
 342       bool has_name = val & 0x2;
 343       unsigned num_components = (val >> 2) & 0x7;
 344       unsigned bit_size = val >> 5;
 345       char *name = has_name ? blob_read_string(ctx->blob) : NULL;
 346       nir_ssa_dest_init(instr, dst, num_components, bit_size, name);
 347       read_add_object(ctx, &dst->ssa);
 348    } else {
 349       bool is_indirect = val & 0x2;
 350       dst->reg.reg = read_object(ctx);
 351       dst->reg.base_offset = blob_read_uint32(ctx->blob);
 352       if (is_indirect) {
 353          dst->reg.indirect = ralloc(instr, nir_src);
 354          read_src(ctx, dst->reg.indirect, instr);
 355       }
 356    }
 357 }
 358
 359 static void
 360 write_deref_chain(write_ctx *ctx, const nir_deref_var *deref_var)
 361 {
 362    write_object(ctx, deref_var->var);
 363
 364    uint32_t len = 0;
 365    for (const nir_deref *d = deref_var->deref.child; d; d = d->child)
 366       len++;
 367    blob_write_uint32(ctx->blob, len);
 368
 369    for (const nir_deref *d = deref_var->deref.child; d; d = d->child) {
 370       blob_write_uint32(ctx->blob, d->deref_type);
 371       switch (d->deref_type) {
 372       case nir_deref_type_array: {
 373          const nir_deref_array *deref_array = nir_deref_as_array(d);
 374          blob_write_uint32(ctx->blob, deref_array->deref_array_type);
 375          blob_write_uint32(ctx->blob, deref_array->base_offset);
 376          if (deref_array->deref_array_type == nir_deref_array_type_indirect)
 377             write_src(ctx, &deref_array->indirect);
 378          break;
 379       }
 380       case nir_deref_type_struct: {
 381          const nir_deref_struct *deref_struct = nir_deref_as_struct(d);
 382          blob_write_uint32(ctx->blob, deref_struct->index);
 383          break;
 384       }
 385       case nir_deref_type_var:
 386          unreachable("Invalid deref type");
 387       }
 388
 389       encode_type_to_blob(ctx->blob, d->type);
 390    }
 391 }
 392
 393 static nir_deref_var *
 394 read_deref_chain(read_ctx *ctx, void *mem_ctx)
 395 {
 396    nir_variable *var = read_object(ctx);
 397    nir_deref_var *deref_var = nir_deref_var_create(mem_ctx, var);
 398
 399    uint32_t len = blob_read_uint32(ctx->blob);
 400
 401    nir_deref *tail = &deref_var->deref;
 402    for (uint32_t i = 0; i < len; i++) {
 403       nir_deref_type deref_type = blob_read_uint32(ctx->blob);
 404       nir_deref *deref = NULL;
 405       switch (deref_type) {
 406       case nir_deref_type_array: {
 407          nir_deref_array *deref_array = nir_deref_array_create(tail);
 408          deref_array->deref_array_type = blob_read_uint32(ctx->blob);
 409          deref_array->base_offset = blob_read_uint32(ctx->blob);
 410          if (deref_array->deref_array_type == nir_deref_array_type_indirect)
 411             read_src(ctx, &deref_array->indirect, mem_ctx);
 412          deref = &deref_array->deref;
 413          break;
 414       }
 415       case nir_deref_type_struct: {
 416          uint32_t index = blob_read_uint32(ctx->blob);
 417          nir_deref_struct *deref_struct = nir_deref_struct_create(tail, index);
 418          deref = &deref_struct->deref;
 419          break;
 420       }
 421       case nir_deref_type_var:
 422          unreachable("Invalid deref type");
 423       }
 424
 425       deref->type = decode_type_from_blob(ctx->blob);
 426
 427       tail->child = deref;
 428       tail = deref;
 429    }
 430
 431    return deref_var;
 432 }
 433
 434 static void
 435 write_alu(write_ctx *ctx, const nir_alu_instr *alu)
 436 {
 437    blob_write_uint32(ctx->blob, alu->op);
 438    uint32_t flags = alu->exact;
 439    flags |= alu->dest.saturate << 1;
 440    flags |= alu->dest.write_mask << 2;
 441    blob_write_uint32(ctx->blob, flags);
 442
 443    write_dest(ctx, &alu->dest.dest);
 444
 445    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 446       write_src(ctx, &alu->src[i].src);
 447       flags = alu->src[i].negate;
 448       flags |= alu->src[i].abs << 1;
 449       for (unsigned j = 0; j < 4; j++)
 450          flags |= alu->src[i].swizzle[j] << (2 + 2 * j);
 451       blob_write_uint32(ctx->blob, flags);
 452    }
 453 }
 454
 455 static nir_alu_instr *
 456 read_alu(read_ctx *ctx)
 457 {
 458    nir_op op = blob_read_uint32(ctx->blob);
 459    nir_alu_instr *alu = nir_alu_instr_create(ctx->nir, op);
 460
 461    uint32_t flags = blob_read_uint32(ctx->blob);
 462    alu->exact = flags & 1;
 463    alu->dest.saturate = flags & 2;
 464    alu->dest.write_mask = flags >> 2;
 465
 466    read_dest(ctx, &alu->dest.dest, &alu->instr);
 467
 468    for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++) {
 469       read_src(ctx, &alu->src[i].src, &alu->instr);
 470       flags = blob_read_uint32(ctx->blob);
 471       alu->src[i].negate = flags & 1;
 472       alu->src[i].abs = flags & 2;
 473       for (unsigned j = 0; j < 4; j++)
 474          alu->src[i].swizzle[j] = (flags >> (2 * j + 2)) & 3;
 475    }
 476
 477    return alu;
 478 }
 479
 480 static void
 481 write_intrinsic(write_ctx *ctx, const nir_intrinsic_instr *intrin)
 482 {
 483    blob_write_uint32(ctx->blob, intrin->intrinsic);
 484
 485    unsigned num_variables = nir_intrinsic_infos[intrin->intrinsic].num_variables;
 486    unsigned num_srcs = nir_intrinsic_infos[intrin->intrinsic].num_srcs;
 487    unsigned num_indices = nir_intrinsic_infos[intrin->intrinsic].num_indices;
 488
 489    blob_write_uint32(ctx->blob, intrin->num_components);
 490
 491    if (nir_intrinsic_infos[intrin->intrinsic].has_dest)
 492       write_dest(ctx, &intrin->dest);
 493
 494    for (unsigned i = 0; i < num_variables; i++)
 495       write_deref_chain(ctx, intrin->variables[i]);
 496
 497    for (unsigned i = 0; i < num_srcs; i++)
 498       write_src(ctx, &intrin->src[i]);
 499
 500    for (unsigned i = 0; i < num_indices; i++)
 501       blob_write_uint32(ctx->blob, intrin->const_index[i]);
 502 }
 503
 504 static nir_intrinsic_instr *
 505 read_intrinsic(read_ctx *ctx)
 506 {
 507    nir_intrinsic_op op = blob_read_uint32(ctx->blob);
 508
 509    nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(ctx->nir, op);
 510
 511    unsigned num_variables = nir_intrinsic_infos[op].num_variables;
 512    unsigned num_srcs = nir_intrinsic_infos[op].num_srcs;
 513    unsigned num_indices = nir_intrinsic_infos[op].num_indices;
 514
 515    intrin->num_components = blob_read_uint32(ctx->blob);
 516
 517    if (nir_intrinsic_infos[op].has_dest)
 518       read_dest(ctx, &intrin->dest, &intrin->instr);
 519
 520    for (unsigned i = 0; i < num_variables; i++)
 521       intrin->variables[i] = read_deref_chain(ctx, &intrin->instr);
 522
 523    for (unsigned i = 0; i < num_srcs; i++)
 524       read_src(ctx, &intrin->src[i], &intrin->instr);
 525
 526    for (unsigned i = 0; i < num_indices; i++)
 527       intrin->const_index[i] = blob_read_uint32(ctx->blob);
 528
 529    return intrin;
 530 }
 531
 532 static void
 533 write_load_const(write_ctx *ctx, const nir_load_const_instr *lc)
 534 {
 535    uint32_t val = lc->def.num_components;
 536    val |= lc->def.bit_size << 3;
 537    blob_write_uint32(ctx->blob, val);
 538    blob_write_bytes(ctx->blob, (uint8_t *) &lc->value, sizeof(lc->value));
 539    write_add_object(ctx, &lc->def);
 540 }
 541
 542 static nir_load_const_instr *
 543 read_load_const(read_ctx *ctx)
 544 {
 545    uint32_t val = blob_read_uint32(ctx->blob);
 546
 547    nir_load_const_instr *lc =
 548       nir_load_const_instr_create(ctx->nir, val & 0x7, val >> 3);
 549
 550    blob_copy_bytes(ctx->blob, (uint8_t *) &lc->value, sizeof(lc->value));
 551    read_add_object(ctx, &lc->def);
 552    return lc;
 553 }
 554
 555 static void
 556 write_ssa_undef(write_ctx *ctx, const nir_ssa_undef_instr *undef)
 557 {
 558    uint32_t val = undef->def.num_components;
 559    val |= undef->def.bit_size << 3;
 560    blob_write_uint32(ctx->blob, val);
 561    write_add_object(ctx, &undef->def);
 562 }
 563
 564 static nir_ssa_undef_instr *
 565 read_ssa_undef(read_ctx *ctx)
 566 {
 567    uint32_t val = blob_read_uint32(ctx->blob);
 568
 569    nir_ssa_undef_instr *undef =
 570       nir_ssa_undef_instr_create(ctx->nir, val & 0x7, val >> 3);
 571
 572    read_add_object(ctx, &undef->def);
 573    return undef;
 574 }
 575
 576 union packed_tex_data {
 577    uint32_t u32;
 578    struct {
 579       enum glsl_sampler_dim sampler_dim:4;
 580       nir_alu_type dest_type:8;
 581       unsigned coord_components:3;
 582       unsigned is_array:1;
 583       unsigned is_shadow:1;
 584       unsigned is_new_style_shadow:1;
 585       unsigned component:2;
 586       unsigned has_texture_deref:1;
 587       unsigned has_sampler_deref:1;
 588    } u;
 589 };
 590
 591 static void
 592 write_tex(write_ctx *ctx, const nir_tex_instr *tex)
 593 {
 594    blob_write_uint32(ctx->blob, tex->num_srcs);
 595    blob_write_uint32(ctx->blob, tex->op);
 596    blob_write_uint32(ctx->blob, tex->texture_index);
 597    blob_write_uint32(ctx->blob, tex->texture_array_size);
 598    blob_write_uint32(ctx->blob, tex->sampler_index);
 599
 600    STATIC_ASSERT(sizeof(union packed_tex_data) == sizeof(uint32_t));
 601    union packed_tex_data packed = {
 602       .u.sampler_dim = tex->sampler_dim,
 603       .u.dest_type = tex->dest_type,
 604       .u.coord_components = tex->coord_components,
 605       .u.is_array = tex->is_array,
 606       .u.is_shadow = tex->is_shadow,
 607       .u.is_new_style_shadow = tex->is_new_style_shadow,
 608       .u.component = tex->component,
 609       .u.has_texture_deref = tex->texture != NULL,
 610       .u.has_sampler_deref = tex->sampler != NULL,
 611    };
 612    blob_write_uint32(ctx->blob, packed.u32);
 613
 614    write_dest(ctx, &tex->dest);
 615    for (unsigned i = 0; i < tex->num_srcs; i++) {
 616       blob_write_uint32(ctx->blob, tex->src[i].src_type);
 617       write_src(ctx, &tex->src[i].src);
 618    }
 619
 620    if (tex->texture)
 621       write_deref_chain(ctx, tex->texture);
 622    if (tex->sampler)
 623       write_deref_chain(ctx, tex->sampler);
 624 }
 625
 626 static nir_tex_instr *
 627 read_tex(read_ctx *ctx)
 628 {
 629    unsigned num_srcs = blob_read_uint32(ctx->blob);
 630    nir_tex_instr *tex = nir_tex_instr_create(ctx->nir, num_srcs);
 631
 632    tex->op = blob_read_uint32(ctx->blob);
 633    tex->texture_index = blob_read_uint32(ctx->blob);
 634    tex->texture_array_size = blob_read_uint32(ctx->blob);
 635    tex->sampler_index = blob_read_uint32(ctx->blob);
 636
 637    union packed_tex_data packed;
 638    packed.u32 = blob_read_uint32(ctx->blob);
 639    tex->sampler_dim = packed.u.sampler_dim;
 640    tex->dest_type = packed.u.dest_type;
 641    tex->coord_components = packed.u.coord_components;
 642    tex->is_array = packed.u.is_array;
 643    tex->is_shadow = packed.u.is_shadow;
 644    tex->is_new_style_shadow = packed.u.is_new_style_shadow;
 645    tex->component = packed.u.component;
 646
 647    read_dest(ctx, &tex->dest, &tex->instr);
 648    for (unsigned i = 0; i < tex->num_srcs; i++) {
 649       tex->src[i].src_type = blob_read_uint32(ctx->blob);
 650       read_src(ctx, &tex->src[i].src, &tex->instr);
 651    }
 652
 653    tex->texture = packed.u.has_texture_deref ?
 654                   read_deref_chain(ctx, &tex->instr) : NULL;
 655    tex->sampler = packed.u.has_sampler_deref ?
 656                   read_deref_chain(ctx, &tex->instr) : NULL;
 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_deref_chain(ctx, call->params[i]);
 778
 779    write_deref_chain(ctx, call->return_deref);
 780 }
 781
 782 static nir_call_instr *
 783 read_call(read_ctx *ctx)
 784 {
 785    nir_function *callee = read_object(ctx);
 786    nir_call_instr *call = nir_call_instr_create(ctx->nir, callee);
 787
 788    for (unsigned i = 0; i < call->num_params; i++)
 789       call->params[i] = read_deref_chain(ctx, &call->instr);
 790
 791    call->return_deref = read_deref_chain(ctx, &call->instr);
 792
 793    return call;
 794 }
 795
 796 static void
 797 write_instr(write_ctx *ctx, const nir_instr *instr)
 798 {
 799    blob_write_uint32(ctx->blob, instr->type);
 800    switch (instr->type) {
 801    case nir_instr_type_alu:
 802       write_alu(ctx, nir_instr_as_alu(instr));
 803       break;
 804    case nir_instr_type_intrinsic:
 805       write_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 806       break;
 807    case nir_instr_type_load_const:
 808       write_load_const(ctx, nir_instr_as_load_const(instr));
 809       break;
 810    case nir_instr_type_ssa_undef:
 811       write_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
 812       break;
 813    case nir_instr_type_tex:
 814       write_tex(ctx, nir_instr_as_tex(instr));
 815       break;
 816    case nir_instr_type_phi:
 817       write_phi(ctx, nir_instr_as_phi(instr));
 818       break;
 819    case nir_instr_type_jump:
 820       write_jump(ctx, nir_instr_as_jump(instr));
 821       break;
 822    case nir_instr_type_call:
 823       write_call(ctx, nir_instr_as_call(instr));
 824       break;
 825    case nir_instr_type_parallel_copy:
 826       unreachable("Cannot write parallel copies");
 827    default:
 828       unreachable("bad instr type");
 829    }
 830 }
 831
 832 static void
 833 read_instr(read_ctx *ctx, nir_block *block)
 834 {
 835    nir_instr_type type = blob_read_uint32(ctx->blob);
 836    nir_instr *instr;
 837    switch (type) {
 838    case nir_instr_type_alu:
 839       instr = &read_alu(ctx)->instr;
 840       break;
 841    case nir_instr_type_intrinsic:
 842       instr = &read_intrinsic(ctx)->instr;
 843       break;
 844    case nir_instr_type_load_const:
 845       instr = &read_load_const(ctx)->instr;
 846       break;
 847    case nir_instr_type_ssa_undef:
 848       instr = &read_ssa_undef(ctx)->instr;
 849       break;
 850    case nir_instr_type_tex:
 851       instr = &read_tex(ctx)->instr;
 852       break;
 853    case nir_instr_type_phi:
 854       /* Phi instructions are a bit of a special case when reading because we
 855        * don't want inserting the instruction to automatically handle use/defs
 856        * for us.  Instead, we need to wait until all the blocks/instructions
 857        * are read so that we can set their sources up.
 858        */
 859       read_phi(ctx, block);
 860       return;
 861    case nir_instr_type_jump:
 862       instr = &read_jump(ctx)->instr;
 863       break;
 864    case nir_instr_type_call:
 865       instr = &read_call(ctx)->instr;
 866       break;
 867    case nir_instr_type_parallel_copy:
 868       unreachable("Cannot read parallel copies");
 869    default:
 870       unreachable("bad instr type");
 871    }
 872
 873    nir_instr_insert_after_block(block, instr);
 874 }
 875
 876 static void
 877 write_block(write_ctx *ctx, const nir_block *block)
 878 {
 879    write_add_object(ctx, block);
 880    blob_write_uint32(ctx->blob, exec_list_length(&block->instr_list));
 881    nir_foreach_instr(instr, block)
 882       write_instr(ctx, instr);
 883 }
 884
 885 static void
 886 read_block(read_ctx *ctx, struct exec_list *cf_list)
 887 {
 888    /* Don't actually create a new block.  Just use the one from the tail of
 889     * the list.  NIR guarantees that the tail of the list is a block and that
 890     * no two blocks are side-by-side in the IR;  It should be empty.
 891     */
 892    nir_block *block =
 893       exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node);
 894
 895    read_add_object(ctx, block);
 896    unsigned num_instrs = blob_read_uint32(ctx->blob);
 897    for (unsigned i = 0; i < num_instrs; i++) {
 898       read_instr(ctx, block);
 899    }
 900 }
 901
 902 static void
 903 write_cf_list(write_ctx *ctx, const struct exec_list *cf_list);
 904
 905 static void
 906 read_cf_list(read_ctx *ctx, struct exec_list *cf_list);
 907
 908 static void
 909 write_if(write_ctx *ctx, nir_if *nif)
 910 {
 911    write_src(ctx, &nif->condition);
 912
 913    write_cf_list(ctx, &nif->then_list);
 914    write_cf_list(ctx, &nif->else_list);
 915 }
 916
 917 static void
 918 read_if(read_ctx *ctx, struct exec_list *cf_list)
 919 {
 920    nir_if *nif = nir_if_create(ctx->nir);
 921
 922    read_src(ctx, &nif->condition, nif);
 923
 924    nir_cf_node_insert_end(cf_list, &nif->cf_node);
 925
 926    read_cf_list(ctx, &nif->then_list);
 927    read_cf_list(ctx, &nif->else_list);
 928 }
 929
 930 static void
 931 write_loop(write_ctx *ctx, nir_loop *loop)
 932 {
 933    write_cf_list(ctx, &loop->body);
 934 }
 935
 936 static void
 937 read_loop(read_ctx *ctx, struct exec_list *cf_list)
 938 {
 939    nir_loop *loop = nir_loop_create(ctx->nir);
 940
 941    nir_cf_node_insert_end(cf_list, &loop->cf_node);
 942
 943    read_cf_list(ctx, &loop->body);
 944 }
 945
 946 static void
 947 write_cf_node(write_ctx *ctx, nir_cf_node *cf)
 948 {
 949    blob_write_uint32(ctx->blob, cf->type);
 950
 951    switch (cf->type) {
 952    case nir_cf_node_block:
 953       write_block(ctx, nir_cf_node_as_block(cf));
 954       break;
 955    case nir_cf_node_if:
 956       write_if(ctx, nir_cf_node_as_if(cf));
 957       break;
 958    case nir_cf_node_loop:
 959       write_loop(ctx, nir_cf_node_as_loop(cf));
 960       break;
 961    default:
 962       unreachable("bad cf type");
 963    }
 964 }
 965
 966 static void
 967 read_cf_node(read_ctx *ctx, struct exec_list *list)
 968 {
 969    nir_cf_node_type type = blob_read_uint32(ctx->blob);
 970
 971    switch (type) {
 972    case nir_cf_node_block:
 973       read_block(ctx, list);
 974       break;
 975    case nir_cf_node_if:
 976       read_if(ctx, list);
 977       break;
 978    case nir_cf_node_loop:
 979       read_loop(ctx, list);
 980       break;
 981    default:
 982       unreachable("bad cf type");
 983    }
 984 }
 985
 986 static void
 987 write_cf_list(write_ctx *ctx, const struct exec_list *cf_list)
 988 {
 989    blob_write_uint32(ctx->blob, exec_list_length(cf_list));
 990    foreach_list_typed(nir_cf_node, cf, node, cf_list) {
 991       write_cf_node(ctx, cf);
 992    }
 993 }
 994
 995 static void
 996 read_cf_list(read_ctx *ctx, struct exec_list *cf_list)
 997 {
 998    uint32_t num_cf_nodes = blob_read_uint32(ctx->blob);
 999    for (unsigned i = 0; i < num_cf_nodes; i++)
1000       read_cf_node(ctx, cf_list);
1001 }
1002
1003 static void
1004 write_function_impl(write_ctx *ctx, const nir_function_impl *fi)
1005 {
1006    write_var_list(ctx, &fi->locals);
1007    write_reg_list(ctx, &fi->registers);
1008    blob_write_uint32(ctx->blob, fi->reg_alloc);
1009
1010    blob_write_uint32(ctx->blob, fi->num_params);
1011    for (unsigned i = 0; i < fi->num_params; i++) {
1012       write_variable(ctx, fi->params[i]);
1013    }
1014
1015    blob_write_uint32(ctx->blob, !!(fi->return_var));
1016    if (fi->return_var)
1017       write_variable(ctx, fi->return_var);
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    fi->num_params = blob_read_uint32(ctx->blob);
1034    for (unsigned i = 0; i < fi->num_params; i++) {
1035       fi->params[i] = read_variable(ctx);
1036    }
1037
1038    bool has_return = blob_read_uint32(ctx->blob);
1039    if (has_return)
1040       fi->return_var = read_variable(ctx);
1041    else
1042       fi->return_var = NULL;
1043
1044    read_cf_list(ctx, &fi->body);
1045    read_fixup_phis(ctx);
1046
1047    fi->valid_metadata = 0;
1048
1049    return fi;
1050 }
1051
1052 static void
1053 write_function(write_ctx *ctx, const nir_function *fxn)
1054 {
1055    blob_write_uint32(ctx->blob, !!(fxn->name));
1056    if (fxn->name)
1057       blob_write_string(ctx->blob, fxn->name);
1058
1059    write_add_object(ctx, fxn);
1060
1061    blob_write_uint32(ctx->blob, fxn->num_params);
1062    for (unsigned i = 0; i < fxn->num_params; i++) {
1063       blob_write_uint32(ctx->blob, fxn->params[i].param_type);
1064       encode_type_to_blob(ctx->blob, fxn->params[i].type);
1065    }
1066
1067    encode_type_to_blob(ctx->blob, fxn->return_type);
1068
1069    /* At first glance, it looks like we should write the function_impl here.
1070     * However, call instructions need to be able to reference at least the
1071     * function and those will get processed as we write the function_impls.
1072     * We stop here and write function_impls as a second pass.
1073     */
1074 }
1075
1076 static void
1077 read_function(read_ctx *ctx)
1078 {
1079    bool has_name = blob_read_uint32(ctx->blob);
1080    char *name = has_name ? blob_read_string(ctx->blob) : NULL;
1081
1082    nir_function *fxn = nir_function_create(ctx->nir, name);
1083
1084    read_add_object(ctx, fxn);
1085
1086    fxn->num_params = blob_read_uint32(ctx->blob);
1087    for (unsigned i = 0; i < fxn->num_params; i++) {
1088       fxn->params[i].param_type = blob_read_uint32(ctx->blob);
1089       fxn->params[i].type = decode_type_from_blob(ctx->blob);
1090    }
1091
1092    fxn->return_type = decode_type_from_blob(ctx->blob);
1093 }
1094
1095 void
1096 nir_serialize(struct blob *blob, const nir_shader *nir)
1097 {
1098    write_ctx ctx;
1099    ctx.remap_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
1100                                              _mesa_key_pointer_equal);
1101    ctx.next_idx = 0;
1102    ctx.blob = blob;
1103    ctx.nir = nir;
1104    util_dynarray_init(&ctx.phi_fixups, NULL);
1105
1106    size_t idx_size_offset = blob_reserve_intptr(blob);
1107
1108    struct shader_info info = nir->info;
1109    uint32_t strings = 0;
1110    if (info.name)
1111       strings |= 0x1;
1112    if (info.label)
1113       strings |= 0x2;
1114    blob_write_uint32(blob, strings);
1115    if (info.name)
1116       blob_write_string(blob, info.name);
1117    if (info.label)
1118       blob_write_string(blob, info.label);
1119    info.name = info.label = NULL;
1120    blob_write_bytes(blob, (uint8_t *) &info, sizeof(info));
1121
1122    write_var_list(&ctx, &nir->uniforms);
1123    write_var_list(&ctx, &nir->inputs);
1124    write_var_list(&ctx, &nir->outputs);
1125    write_var_list(&ctx, &nir->shared);
1126    write_var_list(&ctx, &nir->globals);
1127    write_var_list(&ctx, &nir->system_values);
1128
1129    write_reg_list(&ctx, &nir->registers);
1130    blob_write_uint32(blob, nir->reg_alloc);
1131    blob_write_uint32(blob, nir->num_inputs);
1132    blob_write_uint32(blob, nir->num_uniforms);
1133    blob_write_uint32(blob, nir->num_outputs);
1134    blob_write_uint32(blob, nir->num_shared);
1135
1136    blob_write_uint32(blob, exec_list_length(&nir->functions));
1137    nir_foreach_function(fxn, nir) {
1138       write_function(&ctx, fxn);
1139    }
1140
1141    nir_foreach_function(fxn, nir) {
1142       write_function_impl(&ctx, fxn->impl);
1143    }
1144
1145    *(uintptr_t *)(blob->data + idx_size_offset) = ctx.next_idx;
1146
1147    _mesa_hash_table_destroy(ctx.remap_table, NULL);
1148    util_dynarray_fini(&ctx.phi_fixups);
1149 }
1150
1151 nir_shader *
1152 nir_deserialize(void *mem_ctx,
1153                 const struct nir_shader_compiler_options *options,
1154                 struct blob_reader *blob)
1155 {
1156    read_ctx ctx;
1157    ctx.blob = blob;
1158    list_inithead(&ctx.phi_srcs);
1159    ctx.idx_table_len = blob_read_intptr(blob);
1160    ctx.idx_table = calloc(ctx.idx_table_len, sizeof(uintptr_t));
1161    ctx.next_idx = 0;
1162
1163    uint32_t strings = blob_read_uint32(blob);
1164    char *name = (strings & 0x1) ? blob_read_string(blob) : NULL;
1165    char *label = (strings & 0x2) ? blob_read_string(blob) : NULL;
1166
1167    struct shader_info info;
1168    blob_copy_bytes(blob, (uint8_t *) &info, sizeof(info));
1169
1170    ctx.nir = nir_shader_create(mem_ctx, info.stage, options, NULL);
1171
1172    info.name = name ? ralloc_strdup(ctx.nir, name) : NULL;
1173    info.label = label ? ralloc_strdup(ctx.nir, label) : NULL;
1174
1175    ctx.nir->info = info;
1176
1177    read_var_list(&ctx, &ctx.nir->uniforms);
1178    read_var_list(&ctx, &ctx.nir->inputs);
1179    read_var_list(&ctx, &ctx.nir->outputs);
1180    read_var_list(&ctx, &ctx.nir->shared);
1181    read_var_list(&ctx, &ctx.nir->globals);
1182    read_var_list(&ctx, &ctx.nir->system_values);
1183
1184    read_reg_list(&ctx, &ctx.nir->registers);
1185    ctx.nir->reg_alloc = blob_read_uint32(blob);
1186    ctx.nir->num_inputs = blob_read_uint32(blob);
1187    ctx.nir->num_uniforms = blob_read_uint32(blob);
1188    ctx.nir->num_outputs = blob_read_uint32(blob);
1189    ctx.nir->num_shared = blob_read_uint32(blob);
1190
1191    unsigned num_functions = blob_read_uint32(blob);
1192    for (unsigned i = 0; i < num_functions; i++)
1193       read_function(&ctx);
1194
1195    nir_foreach_function(fxn, ctx.nir)
1196       fxn->impl = read_function_impl(&ctx, fxn);
1197
1198    free(ctx.idx_table);
1199
1200    return ctx.nir;
1201 }
1202
1203 nir_shader *
1204 nir_shader_serialize_deserialize(void *mem_ctx, nir_shader *s)
1205 {
1206    const struct nir_shader_compiler_options *options = s->options;
1207
1208    struct blob writer;
1209    blob_init(&writer);
1210    nir_serialize(&writer, s);
1211    ralloc_free(s);
1212
1213    struct blob_reader reader;
1214    blob_reader_init(&reader, writer.data, writer.size);
1215    nir_shader *ns = nir_deserialize(mem_ctx, options, &reader);
1216
1217    blob_finish(&writer);
1218
1219    return ns;
1220 }