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       __clone_src(state, &nderef->instr,
 315                   &nderef->arr.index, &deref->arr.index);
 316       break;
 317
 318    case nir_deref_type_array_wildcard:
 319    case nir_deref_type_cast:
 320       /* Nothing to do */
 321       break;
 322
 323    default:
 324       unreachable("Invalid instruction deref type");
 325    }
 326
 327    return nderef;
 328 }
 329
 330 static nir_intrinsic_instr *
 331 clone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr)
 332 {
 333    nir_intrinsic_instr *nitr =
 334       nir_intrinsic_instr_create(state->ns, itr->intrinsic);
 335
 336    unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs;
 337
 338    if (nir_intrinsic_infos[itr->intrinsic].has_dest)
 339       __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest);
 340
 341    nitr->num_components = itr->num_components;
 342    memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index));
 343
 344    for (unsigned i = 0; i < num_srcs; i++)
 345       __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]);
 346
 347    return nitr;
 348 }
 349
 350 static nir_load_const_instr *
 351 clone_load_const(clone_state *state, const nir_load_const_instr *lc)
 352 {
 353    nir_load_const_instr *nlc =
 354       nir_load_const_instr_create(state->ns, lc->def.num_components,
 355                                   lc->def.bit_size);
 356
 357    memcpy(&nlc->value, &lc->value, sizeof(nlc->value));
 358
 359    add_remap(state, &nlc->def, &lc->def);
 360
 361    return nlc;
 362 }
 363
 364 static nir_ssa_undef_instr *
 365 clone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa)
 366 {
 367    nir_ssa_undef_instr *nsa =
 368       nir_ssa_undef_instr_create(state->ns, sa->def.num_components,
 369                                  sa->def.bit_size);
 370
 371    add_remap(state, &nsa->def, &sa->def);
 372
 373    return nsa;
 374 }
 375
 376 static nir_tex_instr *
 377 clone_tex(clone_state *state, const nir_tex_instr *tex)
 378 {
 379    nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs);
 380
 381    ntex->sampler_dim = tex->sampler_dim;
 382    ntex->dest_type = tex->dest_type;
 383    ntex->op = tex->op;
 384    __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest);
 385    for (unsigned i = 0; i < ntex->num_srcs; i++) {
 386       ntex->src[i].src_type = tex->src[i].src_type;
 387       __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src);
 388    }
 389    ntex->coord_components = tex->coord_components;
 390    ntex->is_array = tex->is_array;
 391    ntex->is_shadow = tex->is_shadow;
 392    ntex->is_new_style_shadow = tex->is_new_style_shadow;
 393    ntex->component = tex->component;
 394
 395    ntex->texture_index = tex->texture_index;
 396    ntex->texture_array_size = tex->texture_array_size;
 397    ntex->sampler_index = tex->sampler_index;
 398
 399    return ntex;
 400 }
 401
 402 static nir_phi_instr *
 403 clone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk)
 404 {
 405    nir_phi_instr *nphi = nir_phi_instr_create(state->ns);
 406
 407    __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest);
 408
 409    /* Cloning a phi node is a bit different from other instructions.  The
 410     * sources of phi instructions are the only time where we can use an SSA
 411     * def before it is defined.  In order to handle this, we just copy over
 412     * the sources from the old phi instruction directly and then fix them up
 413     * in a second pass once all the instrutions in the function have been
 414     * properly cloned.
 415     *
 416     * In order to ensure that the copied sources (which are the same as the
 417     * old phi instruction's sources for now) don't get inserted into the old
 418     * shader's use-def lists, we have to add the phi instruction *before* we
 419     * set up its sources.
 420     */
 421    nir_instr_insert_after_block(nblk, &nphi->instr);
 422
 423    foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 424       nir_phi_src *nsrc = ralloc(nphi, nir_phi_src);
 425
 426       /* Just copy the old source for now. */
 427       memcpy(nsrc, src, sizeof(*src));
 428
 429       /* Since we're not letting nir_insert_instr handle use/def stuff for us,
 430        * we have to set the parent_instr manually.  It doesn't really matter
 431        * when we do it, so we might as well do it here.
 432        */
 433       nsrc->src.parent_instr = &nphi->instr;
 434
 435       /* Stash it in the list of phi sources.  We'll walk this list and fix up
 436        * sources at the very end of clone_function_impl.
 437        */
 438       list_add(&nsrc->src.use_link, &state->phi_srcs);
 439
 440       exec_list_push_tail(&nphi->srcs, &nsrc->node);
 441    }
 442
 443    return nphi;
 444 }
 445
 446 static nir_jump_instr *
 447 clone_jump(clone_state *state, const nir_jump_instr *jmp)
 448 {
 449    nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type);
 450
 451    return njmp;
 452 }
 453
 454 static nir_call_instr *
 455 clone_call(clone_state *state, const nir_call_instr *call)
 456 {
 457    nir_function *ncallee = remap_global(state, call->callee);
 458    nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee);
 459
 460    for (unsigned i = 0; i < ncall->num_params; i++)
 461       __clone_src(state, ncall, &ncall->params[i], &call->params[i]);
 462
 463    return ncall;
 464 }
 465
 466 static nir_instr *
 467 clone_instr(clone_state *state, const nir_instr *instr)
 468 {
 469    switch (instr->type) {
 470    case nir_instr_type_alu:
 471       return &clone_alu(state, nir_instr_as_alu(instr))->instr;
 472    case nir_instr_type_deref:
 473       return &clone_deref_instr(state, nir_instr_as_deref(instr))->instr;
 474    case nir_instr_type_intrinsic:
 475       return &clone_intrinsic(state, nir_instr_as_intrinsic(instr))->instr;
 476    case nir_instr_type_load_const:
 477       return &clone_load_const(state, nir_instr_as_load_const(instr))->instr;
 478    case nir_instr_type_ssa_undef:
 479       return &clone_ssa_undef(state, nir_instr_as_ssa_undef(instr))->instr;
 480    case nir_instr_type_tex:
 481       return &clone_tex(state, nir_instr_as_tex(instr))->instr;
 482    case nir_instr_type_phi:
 483       unreachable("Cannot clone phis with clone_instr");
 484    case nir_instr_type_jump:
 485       return &clone_jump(state, nir_instr_as_jump(instr))->instr;
 486    case nir_instr_type_call:
 487       return &clone_call(state, nir_instr_as_call(instr))->instr;
 488    case nir_instr_type_parallel_copy:
 489       unreachable("Cannot clone parallel copies");
 490    default:
 491       unreachable("bad instr type");
 492       return NULL;
 493    }
 494 }
 495
 496 static nir_block *
 497 clone_block(clone_state *state, struct exec_list *cf_list, const nir_block *blk)
 498 {
 499    /* Don't actually create a new block.  Just use the one from the tail of
 500     * the list.  NIR guarantees that the tail of the list is a block and that
 501     * no two blocks are side-by-side in the IR;  It should be empty.
 502     */
 503    nir_block *nblk =
 504       exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node);
 505    assert(nblk->cf_node.type == nir_cf_node_block);
 506    assert(exec_list_is_empty(&nblk->instr_list));
 507
 508    /* We need this for phi sources */
 509    add_remap(state, nblk, blk);
 510
 511    nir_foreach_instr(instr, blk) {
 512       if (instr->type == nir_instr_type_phi) {
 513          /* Phi instructions are a bit of a special case when cloning because
 514           * we don't want inserting the instruction to automatically handle
 515           * use/defs for us.  Instead, we need to wait until all the
 516           * blocks/instructions are in so that we can set their sources up.
 517           */
 518          clone_phi(state, nir_instr_as_phi(instr), nblk);
 519       } else {
 520          nir_instr *ninstr = clone_instr(state, instr);
 521          nir_instr_insert_after_block(nblk, ninstr);
 522       }
 523    }
 524
 525    return nblk;
 526 }
 527
 528 static void
 529 clone_cf_list(clone_state *state, struct exec_list *dst,
 530               const struct exec_list *list);
 531
 532 static nir_if *
 533 clone_if(clone_state *state, struct exec_list *cf_list, const nir_if *i)
 534 {
 535    nir_if *ni = nir_if_create(state->ns);
 536
 537    __clone_src(state, ni, &ni->condition, &i->condition);
 538
 539    nir_cf_node_insert_end(cf_list, &ni->cf_node);
 540
 541    clone_cf_list(state, &ni->then_list, &i->then_list);
 542    clone_cf_list(state, &ni->else_list, &i->else_list);
 543
 544    return ni;
 545 }
 546
 547 static nir_loop *
 548 clone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop)
 549 {
 550    nir_loop *nloop = nir_loop_create(state->ns);
 551
 552    nir_cf_node_insert_end(cf_list, &nloop->cf_node);
 553
 554    clone_cf_list(state, &nloop->body, &loop->body);
 555
 556    return nloop;
 557 }
 558
 559 /* clone list of nir_cf_node: */
 560 static void
 561 clone_cf_list(clone_state *state, struct exec_list *dst,
 562               const struct exec_list *list)
 563 {
 564    foreach_list_typed(nir_cf_node, cf, node, list) {
 565       switch (cf->type) {
 566       case nir_cf_node_block:
 567          clone_block(state, dst, nir_cf_node_as_block(cf));
 568          break;
 569       case nir_cf_node_if:
 570          clone_if(state, dst, nir_cf_node_as_if(cf));
 571          break;
 572       case nir_cf_node_loop:
 573          clone_loop(state, dst, nir_cf_node_as_loop(cf));
 574          break;
 575       default:
 576          unreachable("bad cf type");
 577       }
 578    }
 579 }
 580
 581 /* After we've cloned almost everything, we have to walk the list of phi
 582  * sources and fix them up.  Thanks to loops, the block and SSA value for a
 583  * phi source may not be defined when we first encounter it.  Instead, we
 584  * add it to the phi_srcs list and we fix it up here.
 585  */
 586 static void
 587 fixup_phi_srcs(clone_state *state)
 588 {
 589    list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) {
 590       src->pred = remap_local(state, src->pred);
 591
 592       /* Remove from this list */
 593       list_del(&src->src.use_link);
 594
 595       if (src->src.is_ssa) {
 596          src->src.ssa = remap_local(state, src->src.ssa);
 597          list_addtail(&src->src.use_link, &src->src.ssa->uses);
 598       } else {
 599          src->src.reg.reg = remap_reg(state, src->src.reg.reg);
 600          list_addtail(&src->src.use_link, &src->src.reg.reg->uses);
 601       }
 602    }
 603    assert(list_empty(&state->phi_srcs));
 604 }
 605
 606 void
 607 nir_cf_list_clone(nir_cf_list *dst, nir_cf_list *src, nir_cf_node *parent,
 608                   struct hash_table *remap_table)
 609 {
 610    exec_list_make_empty(&dst->list);
 611    dst->impl = src->impl;
 612
 613    if (exec_list_is_empty(&src->list))
 614       return;
 615
 616    clone_state state;
 617    init_clone_state(&state, remap_table, false, true);
 618
 619    /* We use the same shader */
 620    state.ns = src->impl->function->shader;
 621
 622    /* The control-flow code assumes that the list of cf_nodes always starts
 623     * and ends with a block.  We start by adding an empty block.
 624     */
 625    nir_block *nblk = nir_block_create(state.ns);
 626    nblk->cf_node.parent = parent;
 627    exec_list_push_tail(&dst->list, &nblk->cf_node.node);
 628
 629    clone_cf_list(&state, &dst->list, &src->list);
 630
 631    fixup_phi_srcs(&state);
 632 }
 633
 634 static nir_function_impl *
 635 clone_function_impl(clone_state *state, const nir_function_impl *fi)
 636 {
 637    nir_function_impl *nfi = nir_function_impl_create_bare(state->ns);
 638
 639    clone_var_list(state, &nfi->locals, &fi->locals);
 640    clone_reg_list(state, &nfi->registers, &fi->registers);
 641    nfi->reg_alloc = fi->reg_alloc;
 642
 643    assert(list_empty(&state->phi_srcs));
 644
 645    clone_cf_list(state, &nfi->body, &fi->body);
 646
 647    fixup_phi_srcs(state);
 648
 649    /* All metadata is invalidated in the cloning process */
 650    nfi->valid_metadata = 0;
 651
 652    return nfi;
 653 }
 654
 655 nir_function_impl *
 656 nir_function_impl_clone(const nir_function_impl *fi)
 657 {
 658    clone_state state;
 659    init_clone_state(&state, NULL, false, false);
 660
 661    /* We use the same shader */
 662    state.ns = fi->function->shader;
 663
 664    nir_function_impl *nfi = clone_function_impl(&state, fi);
 665
 666    free_clone_state(&state);
 667
 668    return nfi;
 669 }
 670
 671 static nir_function *
 672 clone_function(clone_state *state, const nir_function *fxn, nir_shader *ns)
 673 {
 674    assert(ns == state->ns);
 675    nir_function *nfxn = nir_function_create(ns, fxn->name);
 676
 677    /* Needed for call instructions */
 678    add_remap(state, nfxn, fxn);
 679
 680    nfxn->num_params = fxn->num_params;
 681    nfxn->params = ralloc_array(state->ns, nir_parameter, fxn->num_params);
 682    memcpy(nfxn->params, fxn->params, sizeof(nir_parameter) * fxn->num_params);
 683
 684    /* At first glance, it looks like we should clone the function_impl here.
 685     * However, call instructions need to be able to reference at least the
 686     * function and those will get processed as we clone the function_impls.
 687     * We stop here and do function_impls as a second pass.
 688     */
 689
 690    return nfxn;
 691 }
 692
 693 nir_shader *
 694 nir_shader_clone(void *mem_ctx, const nir_shader *s)
 695 {
 696    clone_state state;
 697    init_clone_state(&state, NULL, true, false);
 698
 699    nir_shader *ns = nir_shader_create(mem_ctx, s->info.stage, s->options, NULL);
 700    state.ns = ns;
 701
 702    clone_var_list(&state, &ns->uniforms, &s->uniforms);
 703    clone_var_list(&state, &ns->inputs,   &s->inputs);
 704    clone_var_list(&state, &ns->outputs,  &s->outputs);
 705    clone_var_list(&state, &ns->shared,   &s->shared);
 706    clone_var_list(&state, &ns->globals,  &s->globals);
 707    clone_var_list(&state, &ns->system_values, &s->system_values);
 708
 709    /* Go through and clone functions */
 710    foreach_list_typed(nir_function, fxn, node, &s->functions)
 711       clone_function(&state, fxn, ns);
 712
 713    /* Only after all functions are cloned can we clone the actual function
 714     * implementations.  This is because nir_call_instrs need to reference the
 715     * functions of other functions and we don't know what order the functions
 716     * will have in the list.
 717     */
 718    nir_foreach_function(fxn, s) {
 719       nir_function *nfxn = remap_global(&state, fxn);
 720       nfxn->impl = clone_function_impl(&state, fxn->impl);
 721       nfxn->impl->function = nfxn;
 722    }
 723
 724    clone_reg_list(&state, &ns->registers, &s->registers);
 725    ns->reg_alloc = s->reg_alloc;
 726
 727    ns->info = s->info;
 728    ns->info.name = ralloc_strdup(ns, ns->info.name);
 729    if (ns->info.label)
 730       ns->info.label = ralloc_strdup(ns, ns->info.label);
 731
 732    ns->num_inputs = s->num_inputs;
 733    ns->num_uniforms = s->num_uniforms;
 734    ns->num_outputs = s->num_outputs;
 735    ns->num_shared = s->num_shared;
 736
 737    ns->constant_data_size = s->constant_data_size;
 738    if (s->constant_data_size > 0) {
 739       ns->constant_data = ralloc_size(ns, s->constant_data_size);
 740       memcpy(ns->constant_data, s->constant_data, s->constant_data_size);
 741    }
 742
 743    free_clone_state(&state);
 744
 745    return ns;
 746 }