src/compiler/nir/nir_clone.c

   1 /*
   2  * Copyright © 2015 Red Hat
   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.h"
  25 #include "nir_control_flow.h"
  26
  27 /* Secret Decoder Ring:
  28  *   clone_foo():
  29  *        Allocate and clone a foo.
  30  *   __clone_foo():
  31  *        Clone body of foo (ie. parent class, embedded struct, etc)
  32  */
  33
  34 typedef struct {
  35    /* True if we are cloning an entire shader. */
  36    bool global_clone;
  37
  38    /* If true allows the clone operation to fall back to the original pointer
  39     * if no clone pointer is found in the remap table.  This allows us to
  40     * clone a loop body without having to add srcs from outside the loop to
  41     * the remap table. This is useful for loop unrolling.
  42     */
  43    bool allow_remap_fallback;
  44
  45    /* maps orig ptr -> cloned ptr: */
  46    struct hash_table *remap_table;
  47
  48    /* List of phi sources. */
  49    struct list_head phi_srcs;
  50
  51    /* new shader object, used as memctx for just about everything else: */
  52    nir_shader *ns;
  53 } clone_state;
  54
  55 static void
  56 init_clone_state(clone_state *state, struct hash_table *remap_table,
  57                  bool global, bool allow_remap_fallback)
  58 {
  59    state->global_clone = global;
  60    state->allow_remap_fallback = allow_remap_fallback;
  61
  62    if (remap_table) {
  63       state->remap_table = remap_table;
  64    } else {
  65       state->remap_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
  66                                                    _mesa_key_pointer_equal);
  67    }
  68
  69    list_inithead(&state->phi_srcs);
  70 }
  71
  72 static void
  73 free_clone_state(clone_state *state)
  74 {
  75    _mesa_hash_table_destroy(state->remap_table, NULL);
  76 }
  77
  78 static inline void *
  79 _lookup_ptr(clone_state *state, const void *ptr, bool global)
  80 {
  81    struct hash_entry *entry;
  82
  83    if (!ptr)
  84       return NULL;
  85
  86    if (!state->global_clone && global)
  87       return (void *)ptr;
  88
  89    entry = _mesa_hash_table_search(state->remap_table, ptr);
  90    if (!entry) {
  91       assert(state->allow_remap_fallback);
  92       return (void *)ptr;
  93    }
  94
  95    return entry->data;
  96 }
  97
  98 static void
  99 add_remap(clone_state *state, void *nptr, const void *ptr)
 100 {
 101    _mesa_hash_table_insert(state->remap_table, ptr, nptr);
 102 }
 103
 104 static void *
 105 remap_local(clone_state *state, const void *ptr)
 106 {
 107    return _lookup_ptr(state, ptr, false);
 108 }
 109
 110 static void *
 111 remap_global(clone_state *state, const void *ptr)
 112 {
 113    return _lookup_ptr(state, ptr, true);
 114 }
 115
 116 static nir_register *
 117 remap_reg(clone_state *state, const nir_register *reg)
 118 {
 119    return _lookup_ptr(state, reg, reg->is_global);
 120 }
 121
 122 static nir_variable *
 123 remap_var(clone_state *state, const nir_variable *var)
 124 {
 125    return _lookup_ptr(state, var, nir_variable_is_global(var));
 126 }
 127
 128 nir_constant *
 129 nir_constant_clone(const nir_constant *c, nir_variable *nvar)
 130 {
 131    nir_constant *nc = ralloc(nvar, nir_constant);
 132
 133    memcpy(nc->values, c->values, sizeof(nc->values));
 134    nc->num_elements = c->num_elements;
 135    nc->elements = ralloc_array(nvar, nir_constant *, c->num_elements);
 136    for (unsigned i = 0; i < c->num_elements; i++) {
 137       nc->elements[i] = nir_constant_clone(c->elements[i], nvar);
 138    }
 139
 140    return nc;
 141 }
 142
 143 /* NOTE: for cloning nir_variables, bypass nir_variable_create to avoid
 144  * having to deal with locals and globals separately:
 145  */
 146 nir_variable *
 147 nir_variable_clone(const nir_variable *var, nir_shader *shader)
 148 {
 149    nir_variable *nvar = rzalloc(shader, nir_variable);
 150
 151    nvar->type = var->type;
 152    nvar->name = ralloc_strdup(nvar, var->name);
 153    nvar->data = var->data;
 154    nvar->num_state_slots = var->num_state_slots;
 155    nvar->state_slots = ralloc_array(nvar, nir_state_slot, var->num_state_slots);
 156    memcpy(nvar->state_slots, var->state_slots,
 157           var->num_state_slots * sizeof(nir_state_slot));
 158    if (var->constant_initializer) {
 159       nvar->constant_initializer =
 160          nir_constant_clone(var->constant_initializer, nvar);
 161    }
 162    nvar->interface_type = var->interface_type;
 163
 164    nvar->num_members = var->num_members;
 165    if (var->num_members) {
 166       nvar->members = ralloc_array(nvar, struct nir_variable_data,
 167                                    var->num_members);
 168       memcpy(nvar->members, var->members,
 169              var->num_members * sizeof(*var->members));
 170    }
 171
 172    return nvar;
 173 }
 174
 175 static nir_variable *
 176 clone_variable(clone_state *state, const nir_variable *var)
 177 {
 178    nir_variable *nvar = nir_variable_clone(var, state->ns);
 179    add_remap(state, nvar, var);
 180
 181    return nvar;
 182 }
 183
 184 /* clone list of nir_variable: */
 185 static void
 186 clone_var_list(clone_state *state, struct exec_list *dst,
 187                const struct exec_list *list)
 188 {
 189    exec_list_make_empty(dst);
 190    foreach_list_typed(nir_variable, var, node, list) {
 191       nir_variable *nvar = clone_variable(state, var);
 192       exec_list_push_tail(dst, &nvar->node);
 193    }
 194 }
 195
 196 /* NOTE: for cloning nir_registers, bypass nir_global/local_reg_create()
 197  * to avoid having to deal with locals and globals separately:
 198  */
 199 static nir_register *
 200 clone_register(clone_state *state, const nir_register *reg)
 201 {
 202    nir_register *nreg = rzalloc(state->ns, nir_register);
 203    add_remap(state, nreg, reg);
 204
 205    nreg->num_components = reg->num_components;
 206    nreg->bit_size = reg->bit_size;
 207    nreg->num_array_elems = reg->num_array_elems;
 208    nreg->index = reg->index;
 209    nreg->name = ralloc_strdup(nreg, reg->name);
 210    nreg->is_global = reg->is_global;
 211    nreg->is_packed = reg->is_packed;
 212
 213    /* reconstructing uses/defs/if_uses handled by nir_instr_insert() */
 214    list_inithead(&nreg->uses);
 215    list_inithead(&nreg->defs);
 216    list_inithead(&nreg->if_uses);
 217
 218    return nreg;
 219 }
 220
 221 /* clone list of nir_register: */
 222 static void
 223 clone_reg_list(clone_state *state, struct exec_list *dst,
 224                const struct exec_list *list)
 225 {
 226    exec_list_make_empty(dst);
 227    foreach_list_typed(nir_register, reg, node, list) {
 228       nir_register *nreg = clone_register(state, reg);
 229       exec_list_push_tail(dst, &nreg->node);
 230    }
 231 }
 232
 233 static void
 234 __clone_src(clone_state *state, void *ninstr_or_if,
 235             nir_src *nsrc, const nir_src *src)
 236 {
 237    nsrc->is_ssa = src->is_ssa;
 238    if (src->is_ssa) {
 239       nsrc->ssa = remap_local(state, src->ssa);
 240    } else {
 241       nsrc->reg.reg = remap_reg(state, src->reg.reg);
 242       if (src->reg.indirect) {
 243          nsrc->reg.indirect = ralloc(ninstr_or_if, nir_src);
 244          __clone_src(state, ninstr_or_if, nsrc->reg.indirect, src->reg.indirect);
 245       }
 246       nsrc->reg.base_offset = src->reg.base_offset;
 247    }
 248 }
 249
 250 static void
 251 __clone_dst(clone_state *state, nir_instr *ninstr,
 252             nir_dest *ndst, const nir_dest *dst)
 253 {
 254    ndst->is_ssa = dst->is_ssa;
 255    if (dst->is_ssa) {
 256       nir_ssa_dest_init(ninstr, ndst, dst->ssa.num_components,
 257                         dst->ssa.bit_size, dst->ssa.name);
 258       add_remap(state, &ndst->ssa, &dst->ssa);
 259    } else {
 260       ndst->reg.reg = remap_reg(state, dst->reg.reg);
 261       if (dst->reg.indirect) {
 262          ndst->reg.indirect = ralloc(ninstr, nir_src);
 263          __clone_src(state, ninstr, ndst->reg.indirect, dst->reg.indirect);
 264       }
 265       ndst->reg.base_offset = dst->reg.base_offset;
 266    }
 267 }
 268
 269 static nir_alu_instr *
 270 clone_alu(clone_state *state, const nir_alu_instr *alu)
 271 {
 272    nir_alu_instr *nalu = nir_alu_instr_create(state->ns, alu->op);
 273    nalu->exact = alu->exact;
 274
 275    __clone_dst(state, &nalu->instr, &nalu->dest.dest, &alu->dest.dest);
 276    nalu->dest.saturate = alu->dest.saturate;
 277    nalu->dest.write_mask = alu->dest.write_mask;
 278
 279    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 280       __clone_src(state, &nalu->instr, &nalu->src[i].src, &alu->src[i].src);
 281       nalu->src[i].negate = alu->src[i].negate;
 282       nalu->src[i].abs = alu->src[i].abs;
 283       memcpy(nalu->src[i].swizzle, alu->src[i].swizzle,
 284              sizeof(nalu->src[i].swizzle));
 285    }
 286
 287    return nalu;
 288 }
 289
 290 static nir_deref_instr *
 291 clone_deref_instr(clone_state *state, const nir_deref_instr *deref)
 292 {
 293    nir_deref_instr *nderef =
 294       nir_deref_instr_create(state->ns, deref->deref_type);
 295
 296    __clone_dst(state, &nderef->instr, &nderef->dest, &deref->dest);
 297
 298    nderef->mode = deref->mode;
 299    nderef->type = deref->type;
 300
 301    if (deref->deref_type == nir_deref_type_var) {
 302       nderef->var = remap_var(state, deref->var);
 303       return nderef;
 304    }
 305
 306    __clone_src(state, &nderef->instr, &nderef->parent, &deref->parent);
 307
 308    switch (deref->deref_type) {
 309    case nir_deref_type_struct:
 310       nderef->strct.index = deref->strct.index;
 311       break;
 312
 313    case nir_deref_type_array:
 314    case nir_deref_type_ptr_as_array:
 315       __clone_src(state, &nderef->instr,
 316                   &nderef->arr.index, &deref->arr.index);
 317       break;
 318
 319    case nir_deref_type_array_wildcard:
 320       /* Nothing to do */
 321       break;
 322
 323    case nir_deref_type_cast:
 324       nderef->cast.ptr_stride = deref->cast.ptr_stride;
 325       break;
 326
 327    default:
 328       unreachable("Invalid instruction deref type");
 329    }
 330
 331    return nderef;
 332 }
 333
 334 static nir_intrinsic_instr *
 335 clone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr)
 336 {
 337    nir_intrinsic_instr *nitr =
 338       nir_intrinsic_instr_create(state->ns, itr->intrinsic);
 339
 340    unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs;
 341
 342    if (nir_intrinsic_infos[itr->intrinsic].has_dest)
 343       __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest);
 344
 345    nitr->num_components = itr->num_components;
 346    memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index));
 347
 348    for (unsigned i = 0; i < num_srcs; i++)
 349       __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]);
 350
 351    return nitr;
 352 }
 353
 354 static nir_load_const_instr *
 355 clone_load_const(clone_state *state, const nir_load_const_instr *lc)
 356 {
 357    nir_load_const_instr *nlc =
 358       nir_load_const_instr_create(state->ns, lc->def.num_components,
 359                                   lc->def.bit_size);
 360
 361    memcpy(&nlc->value, &lc->value, sizeof(nlc->value));
 362
 363    add_remap(state, &nlc->def, &lc->def);
 364
 365    return nlc;
 366 }
 367
 368 static nir_ssa_undef_instr *
 369 clone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa)
 370 {
 371    nir_ssa_undef_instr *nsa =
 372       nir_ssa_undef_instr_create(state->ns, sa->def.num_components,
 373                                  sa->def.bit_size);
 374
 375    add_remap(state, &nsa->def, &sa->def);
 376
 377    return nsa;
 378 }
 379
 380 static nir_tex_instr *
 381 clone_tex(clone_state *state, const nir_tex_instr *tex)
 382 {
 383    nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs);
 384
 385    ntex->sampler_dim = tex->sampler_dim;
 386    ntex->dest_type = tex->dest_type;
 387    ntex->op = tex->op;
 388    __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest);
 389    for (unsigned i = 0; i < ntex->num_srcs; i++) {
 390       ntex->src[i].src_type = tex->src[i].src_type;
 391       __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src);
 392    }
 393    ntex->coord_components = tex->coord_components;
 394    ntex->is_array = tex->is_array;
 395    ntex->is_shadow = tex->is_shadow;
 396    ntex->is_new_style_shadow = tex->is_new_style_shadow;
 397    ntex->component = tex->component;
 398
 399    ntex->texture_index = tex->texture_index;
 400    ntex->texture_array_size = tex->texture_array_size;
 401    ntex->sampler_index = tex->sampler_index;
 402
 403    return ntex;
 404 }
 405
 406 static nir_phi_instr *
 407 clone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk)
 408 {
 409    nir_phi_instr *nphi = nir_phi_instr_create(state->ns);
 410
 411    __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest);
 412
 413    /* Cloning a phi node is a bit different from other instructions.  The
 414     * sources of phi instructions are the only time where we can use an SSA
 415     * def before it is defined.  In order to handle this, we just copy over
 416     * the sources from the old phi instruction directly and then fix them up
 417     * in a second pass once all the instrutions in the function have been
 418     * properly cloned.
 419     *
 420     * In order to ensure that the copied sources (which are the same as the
 421     * old phi instruction's sources for now) don't get inserted into the old
 422     * shader's use-def lists, we have to add the phi instruction *before* we
 423     * set up its sources.
 424     */
 425    nir_instr_insert_after_block(nblk, &nphi->instr);
 426
 427    foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 428       nir_phi_src *nsrc = ralloc(nphi, nir_phi_src);
 429
 430       /* Just copy the old source for now. */
 431       memcpy(nsrc, src, sizeof(*src));
 432
 433       /* Since we're not letting nir_insert_instr handle use/def stuff for us,
 434        * we have to set the parent_instr manually.  It doesn't really matter
 435        * when we do it, so we might as well do it here.
 436        */
 437       nsrc->src.parent_instr = &nphi->instr;
 438
 439       /* Stash it in the list of phi sources.  We'll walk this list and fix up
 440        * sources at the very end of clone_function_impl.
 441        */
 442       list_add(&nsrc->src.use_link, &state->phi_srcs);
 443
 444       exec_list_push_tail(&nphi->srcs, &nsrc->node);
 445    }
 446
 447    return nphi;
 448 }
 449
 450 static nir_jump_instr *
 451 clone_jump(clone_state *state, const nir_jump_instr *jmp)
 452 {
 453    nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type);
 454
 455    return njmp;
 456 }
 457
 458 static nir_call_instr *
 459 clone_call(clone_state *state, const nir_call_instr *call)
 460 {
 461    nir_function *ncallee = remap_global(state, call->callee);
 462    nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee);
 463
 464    for (unsigned i = 0; i < ncall->num_params; i++)
 465       __clone_src(state, ncall, &ncall->params[i], &call->params[i]);
 466
 467    return ncall;
 468 }
 469
 470 static nir_instr *
 471 clone_instr(clone_state *state, const nir_instr *instr)
 472 {
 473    switch (instr->type) {
 474    case nir_instr_type_alu:
 475       return &clone_alu(state, nir_instr_as_alu(instr))->instr;
 476    case nir_instr_type_deref:
 477       return &clone_deref_instr(state, nir_instr_as_deref(instr))->instr;
 478    case nir_instr_type_intrinsic:
 479       return &clone_intrinsic(state, nir_instr_as_intrinsic(instr))->instr;
 480    case nir_instr_type_load_const:
 481       return &clone_load_const(state, nir_instr_as_load_const(instr))->instr;
 482    case nir_instr_type_ssa_undef:
 483       return &clone_ssa_undef(state, nir_instr_as_ssa_undef(instr))->instr;
 484    case nir_instr_type_tex:
 485       return &clone_tex(state, nir_instr_as_tex(instr))->instr;
 486    case nir_instr_type_phi:
 487       unreachable("Cannot clone phis with clone_instr");
 488    case nir_instr_type_jump:
 489       return &clone_jump(state, nir_instr_as_jump(instr))->instr;
 490    case nir_instr_type_call:
 491       return &clone_call(state, nir_instr_as_call(instr))->instr;
 492    case nir_instr_type_parallel_copy:
 493       unreachable("Cannot clone parallel copies");
 494    default:
 495       unreachable("bad instr type");
 496       return NULL;
 497    }
 498 }
 499
 500 static nir_block *
 501 clone_block(clone_state *state, struct exec_list *cf_list, const nir_block *blk)
 502 {
 503    /* Don't actually create a new block.  Just use the one from the tail of
 504     * the list.  NIR guarantees that the tail of the list is a block and that
 505     * no two blocks are side-by-side in the IR;  It should be empty.
 506     */
 507    nir_block *nblk =
 508       exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node);
 509    assert(nblk->cf_node.type == nir_cf_node_block);
 510    assert(exec_list_is_empty(&nblk->instr_list));
 511
 512    /* We need this for phi sources */
 513    add_remap(state, nblk, blk);
 514
 515    nir_foreach_instr(instr, blk) {
 516       if (instr->type == nir_instr_type_phi) {
 517          /* Phi instructions are a bit of a special case when cloning because
 518           * we don't want inserting the instruction to automatically handle
 519           * use/defs for us.  Instead, we need to wait until all the
 520           * blocks/instructions are in so that we can set their sources up.
 521           */
 522          clone_phi(state, nir_instr_as_phi(instr), nblk);
 523       } else {
 524          nir_instr *ninstr = clone_instr(state, instr);
 525          nir_instr_insert_after_block(nblk, ninstr);
 526       }
 527    }
 528
 529    return nblk;
 530 }
 531
 532 static void
 533 clone_cf_list(clone_state *state, struct exec_list *dst,
 534               const struct exec_list *list);
 535
 536 static nir_if *
 537 clone_if(clone_state *state, struct exec_list *cf_list, const nir_if *i)
 538 {
 539    nir_if *ni = nir_if_create(state->ns);
 540
 541    __clone_src(state, ni, &ni->condition, &i->condition);
 542
 543    nir_cf_node_insert_end(cf_list, &ni->cf_node);
 544
 545    clone_cf_list(state, &ni->then_list, &i->then_list);
 546    clone_cf_list(state, &ni->else_list, &i->else_list);
 547
 548    return ni;
 549 }
 550
 551 static nir_loop *
 552 clone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop)
 553 {
 554    nir_loop *nloop = nir_loop_create(state->ns);
 555
 556    nir_cf_node_insert_end(cf_list, &nloop->cf_node);
 557
 558    clone_cf_list(state, &nloop->body, &loop->body);
 559
 560    return nloop;
 561 }
 562
 563 /* clone list of nir_cf_node: */
 564 static void
 565 clone_cf_list(clone_state *state, struct exec_list *dst,
 566               const struct exec_list *list)
 567 {
 568    foreach_list_typed(nir_cf_node, cf, node, list) {
 569       switch (cf->type) {
 570       case nir_cf_node_block:
 571          clone_block(state, dst, nir_cf_node_as_block(cf));
 572          break;
 573       case nir_cf_node_if:
 574          clone_if(state, dst, nir_cf_node_as_if(cf));
 575          break;
 576       case nir_cf_node_loop:
 577          clone_loop(state, dst, nir_cf_node_as_loop(cf));
 578          break;
 579       default:
 580          unreachable("bad cf type");
 581       }
 582    }
 583 }
 584
 585 /* After we've cloned almost everything, we have to walk the list of phi
 586  * sources and fix them up.  Thanks to loops, the block and SSA value for a
 587  * phi source may not be defined when we first encounter it.  Instead, we
 588  * add it to the phi_srcs list and we fix it up here.
 589  */
 590 static void
 591 fixup_phi_srcs(clone_state *state)
 592 {
 593    list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) {
 594       src->pred = remap_local(state, src->pred);
 595
 596       /* Remove from this list */
 597       list_del(&src->src.use_link);
 598
 599       if (src->src.is_ssa) {
 600          src->src.ssa = remap_local(state, src->src.ssa);
 601          list_addtail(&src->src.use_link, &src->src.ssa->uses);
 602       } else {
 603          src->src.reg.reg = remap_reg(state, src->src.reg.reg);
 604          list_addtail(&src->src.use_link, &src->src.reg.reg->uses);
 605       }
 606    }
 607    assert(list_empty(&state->phi_srcs));
 608 }
 609
 610 void
 611 nir_cf_list_clone(nir_cf_list *dst, nir_cf_list *src, nir_cf_node *parent,
 612                   struct hash_table *remap_table)
 613 {
 614    exec_list_make_empty(&dst->list);
 615    dst->impl = src->impl;
 616
 617    if (exec_list_is_empty(&src->list))
 618       return;
 619
 620    clone_state state;
 621    init_clone_state(&state, remap_table, false, true);
 622
 623    /* We use the same shader */
 624    state.ns = src->impl->function->shader;
 625
 626    /* The control-flow code assumes that the list of cf_nodes always starts
 627     * and ends with a block.  We start by adding an empty block.
 628     */
 629    nir_block *nblk = nir_block_create(state.ns);
 630    nblk->cf_node.parent = parent;
 631    exec_list_push_tail(&dst->list, &nblk->cf_node.node);
 632
 633    clone_cf_list(&state, &dst->list, &src->list);
 634
 635    fixup_phi_srcs(&state);
 636 }
 637
 638 static nir_function_impl *
 639 clone_function_impl(clone_state *state, const nir_function_impl *fi)
 640 {
 641    nir_function_impl *nfi = nir_function_impl_create_bare(state->ns);
 642
 643    clone_var_list(state, &nfi->locals, &fi->locals);
 644    clone_reg_list(state, &nfi->registers, &fi->registers);
 645    nfi->reg_alloc = fi->reg_alloc;
 646
 647    assert(list_empty(&state->phi_srcs));
 648
 649    clone_cf_list(state, &nfi->body, &fi->body);
 650
 651    fixup_phi_srcs(state);
 652
 653    /* All metadata is invalidated in the cloning process */
 654    nfi->valid_metadata = 0;
 655
 656    return nfi;
 657 }
 658
 659 nir_function_impl *
 660 nir_function_impl_clone(const nir_function_impl *fi)
 661 {
 662    clone_state state;
 663    init_clone_state(&state, NULL, false, false);
 664
 665    /* We use the same shader */
 666    state.ns = fi->function->shader;
 667
 668    nir_function_impl *nfi = clone_function_impl(&state, fi);
 669
 670    free_clone_state(&state);
 671
 672    return nfi;
 673 }
 674
 675 static nir_function *
 676 clone_function(clone_state *state, const nir_function *fxn, nir_shader *ns)
 677 {
 678    assert(ns == state->ns);
 679    nir_function *nfxn = nir_function_create(ns, fxn->name);
 680
 681    /* Needed for call instructions */
 682    add_remap(state, nfxn, fxn);
 683
 684    nfxn->num_params = fxn->num_params;
 685    nfxn->params = ralloc_array(state->ns, nir_parameter, fxn->num_params);
 686    memcpy(nfxn->params, fxn->params, sizeof(nir_parameter) * fxn->num_params);
 687
 688    /* At first glance, it looks like we should clone the function_impl here.
 689     * However, call instructions need to be able to reference at least the
 690     * function and those will get processed as we clone the function_impls.
 691     * We stop here and do function_impls as a second pass.
 692     */
 693
 694    return nfxn;
 695 }
 696
 697 nir_shader *
 698 nir_shader_clone(void *mem_ctx, const nir_shader *s)
 699 {
 700    clone_state state;
 701    init_clone_state(&state, NULL, true, false);
 702
 703    nir_shader *ns = nir_shader_create(mem_ctx, s->info.stage, s->options, NULL);
 704    state.ns = ns;
 705
 706    clone_var_list(&state, &ns->uniforms, &s->uniforms);
 707    clone_var_list(&state, &ns->inputs,   &s->inputs);
 708    clone_var_list(&state, &ns->outputs,  &s->outputs);
 709    clone_var_list(&state, &ns->shared,   &s->shared);
 710    clone_var_list(&state, &ns->globals,  &s->globals);
 711    clone_var_list(&state, &ns->system_values, &s->system_values);
 712
 713    /* Go through and clone functions */
 714    foreach_list_typed(nir_function, fxn, node, &s->functions)
 715       clone_function(&state, fxn, ns);
 716
 717    /* Only after all functions are cloned can we clone the actual function
 718     * implementations.  This is because nir_call_instrs need to reference the
 719     * functions of other functions and we don't know what order the functions
 720     * will have in the list.
 721     */
 722    nir_foreach_function(fxn, s) {
 723       nir_function *nfxn = remap_global(&state, fxn);
 724       nfxn->impl = clone_function_impl(&state, fxn->impl);
 725       nfxn->impl->function = nfxn;
 726    }
 727
 728    clone_reg_list(&state, &ns->registers, &s->registers);
 729    ns->reg_alloc = s->reg_alloc;
 730
 731    ns->info = s->info;
 732    ns->info.name = ralloc_strdup(ns, ns->info.name);
 733    if (ns->info.label)
 734       ns->info.label = ralloc_strdup(ns, ns->info.label);
 735
 736    ns->num_inputs = s->num_inputs;
 737    ns->num_uniforms = s->num_uniforms;
 738    ns->num_outputs = s->num_outputs;
 739    ns->num_shared = s->num_shared;
 740
 741    ns->constant_data_size = s->constant_data_size;
 742    if (s->constant_data_size > 0) {
 743       ns->constant_data = ralloc_size(ns, s->constant_data_size);
 744       memcpy(ns->constant_data, s->constant_data, s->constant_data_size);
 745    }
 746
 747    free_clone_state(&state);
 748
 749    return ns;
 750 }