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_pointer_hash_table_create(NULL);
  66    }
  67
  68    list_inithead(&state->phi_srcs);
  69 }
  70
  71 static void
  72 free_clone_state(clone_state *state)
  73 {
  74    _mesa_hash_table_destroy(state->remap_table, NULL);
  75 }
  76
  77 static inline void *
  78 _lookup_ptr(clone_state *state, const void *ptr, bool global)
  79 {
  80    struct hash_entry *entry;
  81
  82    if (!ptr)
  83       return NULL;
  84
  85    if (!state->global_clone && global)
  86       return (void *)ptr;
  87
  88    entry = _mesa_hash_table_search(state->remap_table, ptr);
  89    if (!entry) {
  90       assert(state->allow_remap_fallback);
  91       return (void *)ptr;
  92    }
  93
  94    return entry->data;
  95 }
  96
  97 static void
  98 add_remap(clone_state *state, void *nptr, const void *ptr)
  99 {
 100    _mesa_hash_table_insert(state->remap_table, ptr, nptr);
 101 }
 102
 103 static void *
 104 remap_local(clone_state *state, const void *ptr)
 105 {
 106    return _lookup_ptr(state, ptr, false);
 107 }
 108
 109 static void *
 110 remap_global(clone_state *state, const void *ptr)
 111 {
 112    return _lookup_ptr(state, ptr, true);
 113 }
 114
 115 static nir_register *
 116 remap_reg(clone_state *state, const nir_register *reg)
 117 {
 118    return _lookup_ptr(state, reg, false);
 119 }
 120
 121 static nir_variable *
 122 remap_var(clone_state *state, const nir_variable *var)
 123 {
 124    return _lookup_ptr(state, var, nir_variable_is_global(var));
 125 }
 126
 127 nir_constant *
 128 nir_constant_clone(const nir_constant *c, nir_variable *nvar)
 129 {
 130    nir_constant *nc = ralloc(nvar, nir_constant);
 131
 132    memcpy(nc->values, c->values, sizeof(nc->values));
 133    nc->num_elements = c->num_elements;
 134    nc->elements = ralloc_array(nvar, nir_constant *, c->num_elements);
 135    for (unsigned i = 0; i < c->num_elements; i++) {
 136       nc->elements[i] = nir_constant_clone(c->elements[i], nvar);
 137    }
 138
 139    return nc;
 140 }
 141
 142 /* NOTE: for cloning nir_variables, bypass nir_variable_create to avoid
 143  * having to deal with locals and globals separately:
 144  */
 145 nir_variable *
 146 nir_variable_clone(const nir_variable *var, nir_shader *shader)
 147 {
 148    nir_variable *nvar = rzalloc(shader, nir_variable);
 149
 150    nvar->type = var->type;
 151    nvar->name = ralloc_strdup(nvar, var->name);
 152    nvar->data = var->data;
 153    nvar->num_state_slots = var->num_state_slots;
 154    nvar->state_slots = ralloc_array(nvar, nir_state_slot, var->num_state_slots);
 155    memcpy(nvar->state_slots, var->state_slots,
 156           var->num_state_slots * sizeof(nir_state_slot));
 157    if (var->constant_initializer) {
 158       nvar->constant_initializer =
 159          nir_constant_clone(var->constant_initializer, nvar);
 160    }
 161    nvar->interface_type = var->interface_type;
 162
 163    nvar->num_members = var->num_members;
 164    if (var->num_members) {
 165       nvar->members = ralloc_array(nvar, struct nir_variable_data,
 166                                    var->num_members);
 167       memcpy(nvar->members, var->members,
 168              var->num_members * sizeof(*var->members));
 169    }
 170
 171    return nvar;
 172 }
 173
 174 static nir_variable *
 175 clone_variable(clone_state *state, const nir_variable *var)
 176 {
 177    nir_variable *nvar = nir_variable_clone(var, state->ns);
 178    add_remap(state, nvar, var);
 179
 180    return nvar;
 181 }
 182
 183 /* clone list of nir_variable: */
 184 static void
 185 clone_var_list(clone_state *state, struct exec_list *dst,
 186                const struct exec_list *list)
 187 {
 188    exec_list_make_empty(dst);
 189    foreach_list_typed(nir_variable, var, node, list) {
 190       nir_variable *nvar = clone_variable(state, var);
 191       exec_list_push_tail(dst, &nvar->node);
 192    }
 193 }
 194
 195 /* NOTE: for cloning nir_registers, bypass nir_global/local_reg_create()
 196  * to avoid having to deal with locals and globals separately:
 197  */
 198 static nir_register *
 199 clone_register(clone_state *state, const nir_register *reg)
 200 {
 201    nir_register *nreg = rzalloc(state->ns, nir_register);
 202    add_remap(state, nreg, reg);
 203
 204    nreg->num_components = reg->num_components;
 205    nreg->bit_size = reg->bit_size;
 206    nreg->num_array_elems = reg->num_array_elems;
 207    nreg->index = reg->index;
 208    nreg->name = ralloc_strdup(nreg, reg->name);
 209
 210    /* reconstructing uses/defs/if_uses handled by nir_instr_insert() */
 211    list_inithead(&nreg->uses);
 212    list_inithead(&nreg->defs);
 213    list_inithead(&nreg->if_uses);
 214
 215    return nreg;
 216 }
 217
 218 /* clone list of nir_register: */
 219 static void
 220 clone_reg_list(clone_state *state, struct exec_list *dst,
 221                const struct exec_list *list)
 222 {
 223    exec_list_make_empty(dst);
 224    foreach_list_typed(nir_register, reg, node, list) {
 225       nir_register *nreg = clone_register(state, reg);
 226       exec_list_push_tail(dst, &nreg->node);
 227    }
 228 }
 229
 230 static void
 231 __clone_src(clone_state *state, void *ninstr_or_if,
 232             nir_src *nsrc, const nir_src *src)
 233 {
 234    nsrc->is_ssa = src->is_ssa;
 235    if (src->is_ssa) {
 236       nsrc->ssa = remap_local(state, src->ssa);
 237    } else {
 238       nsrc->reg.reg = remap_reg(state, src->reg.reg);
 239       if (src->reg.indirect) {
 240          nsrc->reg.indirect = ralloc(ninstr_or_if, nir_src);
 241          __clone_src(state, ninstr_or_if, nsrc->reg.indirect, src->reg.indirect);
 242       }
 243       nsrc->reg.base_offset = src->reg.base_offset;
 244    }
 245 }
 246
 247 static void
 248 __clone_dst(clone_state *state, nir_instr *ninstr,
 249             nir_dest *ndst, const nir_dest *dst)
 250 {
 251    ndst->is_ssa = dst->is_ssa;
 252    if (dst->is_ssa) {
 253       nir_ssa_dest_init(ninstr, ndst, dst->ssa.num_components,
 254                         dst->ssa.bit_size, dst->ssa.name);
 255       add_remap(state, &ndst->ssa, &dst->ssa);
 256    } else {
 257       ndst->reg.reg = remap_reg(state, dst->reg.reg);
 258       if (dst->reg.indirect) {
 259          ndst->reg.indirect = ralloc(ninstr, nir_src);
 260          __clone_src(state, ninstr, ndst->reg.indirect, dst->reg.indirect);
 261       }
 262       ndst->reg.base_offset = dst->reg.base_offset;
 263    }
 264 }
 265
 266 static nir_alu_instr *
 267 clone_alu(clone_state *state, const nir_alu_instr *alu)
 268 {
 269    nir_alu_instr *nalu = nir_alu_instr_create(state->ns, alu->op);
 270    nalu->exact = alu->exact;
 271
 272    __clone_dst(state, &nalu->instr, &nalu->dest.dest, &alu->dest.dest);
 273    nalu->dest.saturate = alu->dest.saturate;
 274    nalu->dest.write_mask = alu->dest.write_mask;
 275
 276    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 277       __clone_src(state, &nalu->instr, &nalu->src[i].src, &alu->src[i].src);
 278       nalu->src[i].negate = alu->src[i].negate;
 279       nalu->src[i].abs = alu->src[i].abs;
 280       memcpy(nalu->src[i].swizzle, alu->src[i].swizzle,
 281              sizeof(nalu->src[i].swizzle));
 282    }
 283
 284    return nalu;
 285 }
 286
 287 static nir_deref_instr *
 288 clone_deref_instr(clone_state *state, const nir_deref_instr *deref)
 289 {
 290    nir_deref_instr *nderef =
 291       nir_deref_instr_create(state->ns, deref->deref_type);
 292
 293    __clone_dst(state, &nderef->instr, &nderef->dest, &deref->dest);
 294
 295    nderef->mode = deref->mode;
 296    nderef->type = deref->type;
 297
 298    if (deref->deref_type == nir_deref_type_var) {
 299       nderef->var = remap_var(state, deref->var);
 300       return nderef;
 301    }
 302
 303    __clone_src(state, &nderef->instr, &nderef->parent, &deref->parent);
 304
 305    switch (deref->deref_type) {
 306    case nir_deref_type_struct:
 307       nderef->strct.index = deref->strct.index;
 308       break;
 309
 310    case nir_deref_type_array:
 311    case nir_deref_type_ptr_as_array:
 312       __clone_src(state, &nderef->instr,
 313                   &nderef->arr.index, &deref->arr.index);
 314       break;
 315
 316    case nir_deref_type_array_wildcard:
 317       /* Nothing to do */
 318       break;
 319
 320    case nir_deref_type_cast:
 321       nderef->cast.ptr_stride = deref->cast.ptr_stride;
 322       break;
 323
 324    default:
 325       unreachable("Invalid instruction deref type");
 326    }
 327
 328    return nderef;
 329 }
 330
 331 static nir_intrinsic_instr *
 332 clone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr)
 333 {
 334    nir_intrinsic_instr *nitr =
 335       nir_intrinsic_instr_create(state->ns, itr->intrinsic);
 336
 337    unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs;
 338
 339    if (nir_intrinsic_infos[itr->intrinsic].has_dest)
 340       __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest);
 341
 342    nitr->num_components = itr->num_components;
 343    memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index));
 344
 345    for (unsigned i = 0; i < num_srcs; i++)
 346       __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]);
 347
 348    return nitr;
 349 }
 350
 351 static nir_load_const_instr *
 352 clone_load_const(clone_state *state, const nir_load_const_instr *lc)
 353 {
 354    nir_load_const_instr *nlc =
 355       nir_load_const_instr_create(state->ns, lc->def.num_components,
 356                                   lc->def.bit_size);
 357
 358    memcpy(&nlc->value, &lc->value, sizeof(nlc->value));
 359
 360    add_remap(state, &nlc->def, &lc->def);
 361
 362    return nlc;
 363 }
 364
 365 static nir_ssa_undef_instr *
 366 clone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa)
 367 {
 368    nir_ssa_undef_instr *nsa =
 369       nir_ssa_undef_instr_create(state->ns, sa->def.num_components,
 370                                  sa->def.bit_size);
 371
 372    add_remap(state, &nsa->def, &sa->def);
 373
 374    return nsa;
 375 }
 376
 377 static nir_tex_instr *
 378 clone_tex(clone_state *state, const nir_tex_instr *tex)
 379 {
 380    nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs);
 381
 382    ntex->sampler_dim = tex->sampler_dim;
 383    ntex->dest_type = tex->dest_type;
 384    ntex->op = tex->op;
 385    __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest);
 386    for (unsigned i = 0; i < ntex->num_srcs; i++) {
 387       ntex->src[i].src_type = tex->src[i].src_type;
 388       __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src);
 389    }
 390    ntex->coord_components = tex->coord_components;
 391    ntex->is_array = tex->is_array;
 392    ntex->is_shadow = tex->is_shadow;
 393    ntex->is_new_style_shadow = tex->is_new_style_shadow;
 394    ntex->component = tex->component;
 395    memcpy(ntex->tg4_offsets, tex->tg4_offsets, sizeof(tex->tg4_offsets));
 396
 397    ntex->texture_index = tex->texture_index;
 398    ntex->texture_array_size = tex->texture_array_size;
 399    ntex->sampler_index = tex->sampler_index;
 400
 401    return ntex;
 402 }
 403
 404 static nir_phi_instr *
 405 clone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk)
 406 {
 407    nir_phi_instr *nphi = nir_phi_instr_create(state->ns);
 408
 409    __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest);
 410
 411    /* Cloning a phi node is a bit different from other instructions.  The
 412     * sources of phi instructions are the only time where we can use an SSA
 413     * def before it is defined.  In order to handle this, we just copy over
 414     * the sources from the old phi instruction directly and then fix them up
 415     * in a second pass once all the instrutions in the function have been
 416     * properly cloned.
 417     *
 418     * In order to ensure that the copied sources (which are the same as the
 419     * old phi instruction's sources for now) don't get inserted into the old
 420     * shader's use-def lists, we have to add the phi instruction *before* we
 421     * set up its sources.
 422     */
 423    nir_instr_insert_after_block(nblk, &nphi->instr);
 424
 425    foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 426       nir_phi_src *nsrc = ralloc(nphi, nir_phi_src);
 427
 428       /* Just copy the old source for now. */
 429       memcpy(nsrc, src, sizeof(*src));
 430
 431       /* Since we're not letting nir_insert_instr handle use/def stuff for us,
 432        * we have to set the parent_instr manually.  It doesn't really matter
 433        * when we do it, so we might as well do it here.
 434        */
 435       nsrc->src.parent_instr = &nphi->instr;
 436
 437       /* Stash it in the list of phi sources.  We'll walk this list and fix up
 438        * sources at the very end of clone_function_impl.
 439        */
 440       list_add(&nsrc->src.use_link, &state->phi_srcs);
 441
 442       exec_list_push_tail(&nphi->srcs, &nsrc->node);
 443    }
 444
 445    return nphi;
 446 }
 447
 448 static nir_jump_instr *
 449 clone_jump(clone_state *state, const nir_jump_instr *jmp)
 450 {
 451    nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type);
 452
 453    return njmp;
 454 }
 455
 456 static nir_call_instr *
 457 clone_call(clone_state *state, const nir_call_instr *call)
 458 {
 459    nir_function *ncallee = remap_global(state, call->callee);
 460    nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee);
 461
 462    for (unsigned i = 0; i < ncall->num_params; i++)
 463       __clone_src(state, ncall, &ncall->params[i], &call->params[i]);
 464
 465    return ncall;
 466 }
 467
 468 static nir_instr *
 469 clone_instr(clone_state *state, const nir_instr *instr)
 470 {
 471    switch (instr->type) {
 472    case nir_instr_type_alu:
 473       return &clone_alu(state, nir_instr_as_alu(instr))->instr;
 474    case nir_instr_type_deref:
 475       return &clone_deref_instr(state, nir_instr_as_deref(instr))->instr;
 476    case nir_instr_type_intrinsic:
 477       return &clone_intrinsic(state, nir_instr_as_intrinsic(instr))->instr;
 478    case nir_instr_type_load_const:
 479       return &clone_load_const(state, nir_instr_as_load_const(instr))->instr;
 480    case nir_instr_type_ssa_undef:
 481       return &clone_ssa_undef(state, nir_instr_as_ssa_undef(instr))->instr;
 482    case nir_instr_type_tex:
 483       return &clone_tex(state, nir_instr_as_tex(instr))->instr;
 484    case nir_instr_type_phi:
 485       unreachable("Cannot clone phis with clone_instr");
 486    case nir_instr_type_jump:
 487       return &clone_jump(state, nir_instr_as_jump(instr))->instr;
 488    case nir_instr_type_call:
 489       return &clone_call(state, nir_instr_as_call(instr))->instr;
 490    case nir_instr_type_parallel_copy:
 491       unreachable("Cannot clone parallel copies");
 492    default:
 493       unreachable("bad instr type");
 494       return NULL;
 495    }
 496 }
 497
 498 static nir_block *
 499 clone_block(clone_state *state, struct exec_list *cf_list, const nir_block *blk)
 500 {
 501    /* Don't actually create a new block.  Just use the one from the tail of
 502     * the list.  NIR guarantees that the tail of the list is a block and that
 503     * no two blocks are side-by-side in the IR;  It should be empty.
 504     */
 505    nir_block *nblk =
 506       exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node);
 507    assert(nblk->cf_node.type == nir_cf_node_block);
 508    assert(exec_list_is_empty(&nblk->instr_list));
 509
 510    /* We need this for phi sources */
 511    add_remap(state, nblk, blk);
 512
 513    nir_foreach_instr(instr, blk) {
 514       if (instr->type == nir_instr_type_phi) {
 515          /* Phi instructions are a bit of a special case when cloning because
 516           * we don't want inserting the instruction to automatically handle
 517           * use/defs for us.  Instead, we need to wait until all the
 518           * blocks/instructions are in so that we can set their sources up.
 519           */
 520          clone_phi(state, nir_instr_as_phi(instr), nblk);
 521       } else {
 522          nir_instr *ninstr = clone_instr(state, instr);
 523          nir_instr_insert_after_block(nblk, ninstr);
 524       }
 525    }
 526
 527    return nblk;
 528 }
 529
 530 static void
 531 clone_cf_list(clone_state *state, struct exec_list *dst,
 532               const struct exec_list *list);
 533
 534 static nir_if *
 535 clone_if(clone_state *state, struct exec_list *cf_list, const nir_if *i)
 536 {
 537    nir_if *ni = nir_if_create(state->ns);
 538    ni->control = i->control;
 539
 540    __clone_src(state, ni, &ni->condition, &i->condition);
 541
 542    nir_cf_node_insert_end(cf_list, &ni->cf_node);
 543
 544    clone_cf_list(state, &ni->then_list, &i->then_list);
 545    clone_cf_list(state, &ni->else_list, &i->else_list);
 546
 547    return ni;
 548 }
 549
 550 static nir_loop *
 551 clone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop)
 552 {
 553    nir_loop *nloop = nir_loop_create(state->ns);
 554    nloop->control = loop->control;
 555    nloop->partially_unrolled = loop->partially_unrolled;
 556
 557    nir_cf_node_insert_end(cf_list, &nloop->cf_node);
 558
 559    clone_cf_list(state, &nloop->body, &loop->body);
 560
 561    return nloop;
 562 }
 563
 564 /* clone list of nir_cf_node: */
 565 static void
 566 clone_cf_list(clone_state *state, struct exec_list *dst,
 567               const struct exec_list *list)
 568 {
 569    foreach_list_typed(nir_cf_node, cf, node, list) {
 570       switch (cf->type) {
 571       case nir_cf_node_block:
 572          clone_block(state, dst, nir_cf_node_as_block(cf));
 573          break;
 574       case nir_cf_node_if:
 575          clone_if(state, dst, nir_cf_node_as_if(cf));
 576          break;
 577       case nir_cf_node_loop:
 578          clone_loop(state, dst, nir_cf_node_as_loop(cf));
 579          break;
 580       default:
 581          unreachable("bad cf type");
 582       }
 583    }
 584 }
 585
 586 /* After we've cloned almost everything, we have to walk the list of phi
 587  * sources and fix them up.  Thanks to loops, the block and SSA value for a
 588  * phi source may not be defined when we first encounter it.  Instead, we
 589  * add it to the phi_srcs list and we fix it up here.
 590  */
 591 static void
 592 fixup_phi_srcs(clone_state *state)
 593 {
 594    list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) {
 595       src->pred = remap_local(state, src->pred);
 596
 597       /* Remove from this list */
 598       list_del(&src->src.use_link);
 599
 600       if (src->src.is_ssa) {
 601          src->src.ssa = remap_local(state, src->src.ssa);
 602          list_addtail(&src->src.use_link, &src->src.ssa->uses);
 603       } else {
 604          src->src.reg.reg = remap_reg(state, src->src.reg.reg);
 605          list_addtail(&src->src.use_link, &src->src.reg.reg->uses);
 606       }
 607    }
 608    assert(list_empty(&state->phi_srcs));
 609 }
 610
 611 void
 612 nir_cf_list_clone(nir_cf_list *dst, nir_cf_list *src, nir_cf_node *parent,
 613                   struct hash_table *remap_table)
 614 {
 615    exec_list_make_empty(&dst->list);
 616    dst->impl = src->impl;
 617
 618    if (exec_list_is_empty(&src->list))
 619       return;
 620
 621    clone_state state;
 622    init_clone_state(&state, remap_table, false, true);
 623
 624    /* We use the same shader */
 625    state.ns = src->impl->function->shader;
 626
 627    /* The control-flow code assumes that the list of cf_nodes always starts
 628     * and ends with a block.  We start by adding an empty block.
 629     */
 630    nir_block *nblk = nir_block_create(state.ns);
 631    nblk->cf_node.parent = parent;
 632    exec_list_push_tail(&dst->list, &nblk->cf_node.node);
 633
 634    clone_cf_list(&state, &dst->list, &src->list);
 635
 636    fixup_phi_srcs(&state);
 637 }
 638
 639 static nir_function_impl *
 640 clone_function_impl(clone_state *state, const nir_function_impl *fi)
 641 {
 642    nir_function_impl *nfi = nir_function_impl_create_bare(state->ns);
 643
 644    clone_var_list(state, &nfi->locals, &fi->locals);
 645    clone_reg_list(state, &nfi->registers, &fi->registers);
 646    nfi->reg_alloc = fi->reg_alloc;
 647
 648    assert(list_empty(&state->phi_srcs));
 649
 650    clone_cf_list(state, &nfi->body, &fi->body);
 651
 652    fixup_phi_srcs(state);
 653
 654    /* All metadata is invalidated in the cloning process */
 655    nfi->valid_metadata = 0;
 656
 657    return nfi;
 658 }
 659
 660 nir_function_impl *
 661 nir_function_impl_clone(nir_shader *shader, const nir_function_impl *fi)
 662 {
 663    clone_state state;
 664    init_clone_state(&state, NULL, false, false);
 665
 666    state.ns = 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    nfxn->is_entrypoint = fxn->is_entrypoint;
 688
 689    /* At first glance, it looks like we should clone the function_impl here.
 690     * However, call instructions need to be able to reference at least the
 691     * function and those will get processed as we clone the function_impls.
 692     * We stop here and do function_impls as a second pass.
 693     */
 694
 695    return nfxn;
 696 }
 697
 698 nir_shader *
 699 nir_shader_clone(void *mem_ctx, const nir_shader *s)
 700 {
 701    clone_state state;
 702    init_clone_state(&state, NULL, true, false);
 703
 704    nir_shader *ns = nir_shader_create(mem_ctx, s->info.stage, s->options, NULL);
 705    state.ns = ns;
 706
 707    clone_var_list(&state, &ns->uniforms, &s->uniforms);
 708    clone_var_list(&state, &ns->inputs,   &s->inputs);
 709    clone_var_list(&state, &ns->outputs,  &s->outputs);
 710    clone_var_list(&state, &ns->shared,   &s->shared);
 711    clone_var_list(&state, &ns->globals,  &s->globals);
 712    clone_var_list(&state, &ns->system_values, &s->system_values);
 713
 714    /* Go through and clone functions */
 715    foreach_list_typed(nir_function, fxn, node, &s->functions)
 716       clone_function(&state, fxn, ns);
 717
 718    /* Only after all functions are cloned can we clone the actual function
 719     * implementations.  This is because nir_call_instrs need to reference the
 720     * functions of other functions and we don't know what order the functions
 721     * will have in the list.
 722     */
 723    nir_foreach_function(fxn, s) {
 724       nir_function *nfxn = remap_global(&state, fxn);
 725       nfxn->impl = clone_function_impl(&state, fxn->impl);
 726       nfxn->impl->function = nfxn;
 727    }
 728
 729    ns->info = s->info;
 730    ns->info.name = ralloc_strdup(ns, ns->info.name);
 731    if (ns->info.label)
 732       ns->info.label = ralloc_strdup(ns, ns->info.label);
 733
 734    ns->num_inputs = s->num_inputs;
 735    ns->num_uniforms = s->num_uniforms;
 736    ns->num_outputs = s->num_outputs;
 737    ns->num_shared = s->num_shared;
 738    ns->scratch_size = s->scratch_size;
 739
 740    ns->constant_data_size = s->constant_data_size;
 741    if (s->constant_data_size > 0) {
 742       ns->constant_data = ralloc_size(ns, s->constant_data_size);
 743       memcpy(ns->constant_data, s->constant_data, s->constant_data_size);
 744    }
 745
 746    free_clone_state(&state);
 747
 748    return ns;
 749 }