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_private.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    /* maps orig ptr -> cloned ptr: */
  39    struct hash_table *remap_table;
  40
  41    /* List of phi sources. */
  42    struct list_head phi_srcs;
  43
  44    /* new shader object, used as memctx for just about everything else: */
  45    nir_shader *ns;
  46 } clone_state;
  47
  48 static void
  49 init_clone_state(clone_state *state, bool global)
  50 {
  51    state->global_clone = global;
  52    state->remap_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
  53                                                 _mesa_key_pointer_equal);
  54    list_inithead(&state->phi_srcs);
  55 }
  56
  57 static void
  58 free_clone_state(clone_state *state)
  59 {
  60    _mesa_hash_table_destroy(state->remap_table, NULL);
  61 }
  62
  63 static inline void *
  64 _lookup_ptr(clone_state *state, const void *ptr, bool global)
  65 {
  66    struct hash_entry *entry;
  67
  68    if (!ptr)
  69       return NULL;
  70
  71    if (!state->global_clone && global)
  72       return (void *)ptr;
  73
  74    entry = _mesa_hash_table_search(state->remap_table, ptr);
  75    assert(entry && "Failed to find pointer!");
  76    if (!entry)
  77       return NULL;
  78
  79    return entry->data;
  80 }
  81
  82 static void
  83 add_remap(clone_state *state, void *nptr, const void *ptr)
  84 {
  85    _mesa_hash_table_insert(state->remap_table, ptr, nptr);
  86 }
  87
  88 static void *
  89 remap_local(clone_state *state, const void *ptr)
  90 {
  91    return _lookup_ptr(state, ptr, false);
  92 }
  93
  94 static void *
  95 remap_global(clone_state *state, const void *ptr)
  96 {
  97    return _lookup_ptr(state, ptr, true);
  98 }
  99
 100 static nir_register *
 101 remap_reg(clone_state *state, const nir_register *reg)
 102 {
 103    return _lookup_ptr(state, reg, reg->is_global);
 104 }
 105
 106 static nir_variable *
 107 remap_var(clone_state *state, const nir_variable *var)
 108 {
 109    return _lookup_ptr(state, var, nir_variable_is_global(var));
 110 }
 111
 112 nir_constant *
 113 nir_constant_clone(const nir_constant *c, nir_variable *nvar)
 114 {
 115    nir_constant *nc = ralloc(nvar, nir_constant);
 116
 117    nc->value = c->value;
 118    nc->num_elements = c->num_elements;
 119    nc->elements = ralloc_array(nvar, nir_constant *, c->num_elements);
 120    for (unsigned i = 0; i < c->num_elements; i++) {
 121       nc->elements[i] = nir_constant_clone(c->elements[i], nvar);
 122    }
 123
 124    return nc;
 125 }
 126
 127 /* NOTE: for cloning nir_variable's, bypass nir_variable_create to avoid
 128  * having to deal with locals and globals separately:
 129  */
 130 static nir_variable *
 131 clone_variable(clone_state *state, const nir_variable *var)
 132 {
 133    nir_variable *nvar = rzalloc(state->ns, nir_variable);
 134    add_remap(state, nvar, var);
 135
 136    nvar->type = var->type;
 137    nvar->name = ralloc_strdup(nvar, var->name);
 138    nvar->data = var->data;
 139    nvar->num_state_slots = var->num_state_slots;
 140    nvar->state_slots = ralloc_array(nvar, nir_state_slot, var->num_state_slots);
 141    memcpy(nvar->state_slots, var->state_slots,
 142           var->num_state_slots * sizeof(nir_state_slot));
 143    if (var->constant_initializer) {
 144       nvar->constant_initializer =
 145          nir_constant_clone(var->constant_initializer, nvar);
 146    }
 147    nvar->interface_type = var->interface_type;
 148
 149    return nvar;
 150 }
 151
 152 /* clone list of nir_variable: */
 153 static void
 154 clone_var_list(clone_state *state, struct exec_list *dst,
 155                const struct exec_list *list)
 156 {
 157    exec_list_make_empty(dst);
 158    foreach_list_typed(nir_variable, var, node, list) {
 159       nir_variable *nvar = clone_variable(state, var);
 160       exec_list_push_tail(dst, &nvar->node);
 161    }
 162 }
 163
 164 /* NOTE: for cloning nir_register's, bypass nir_global/local_reg_create()
 165  * to avoid having to deal with locals and globals separately:
 166  */
 167 static nir_register *
 168 clone_register(clone_state *state, const nir_register *reg)
 169 {
 170    nir_register *nreg = rzalloc(state->ns, nir_register);
 171    add_remap(state, nreg, reg);
 172
 173    nreg->num_components = reg->num_components;
 174    nreg->num_array_elems = reg->num_array_elems;
 175    nreg->index = reg->index;
 176    nreg->name = ralloc_strdup(nreg, reg->name);
 177    nreg->is_global = reg->is_global;
 178    nreg->is_packed = reg->is_packed;
 179
 180    /* reconstructing uses/defs/if_uses handled by nir_instr_insert() */
 181    list_inithead(&nreg->uses);
 182    list_inithead(&nreg->defs);
 183    list_inithead(&nreg->if_uses);
 184
 185    return nreg;
 186 }
 187
 188 /* clone list of nir_register: */
 189 static void
 190 clone_reg_list(clone_state *state, struct exec_list *dst,
 191                const struct exec_list *list)
 192 {
 193    exec_list_make_empty(dst);
 194    foreach_list_typed(nir_register, reg, node, list) {
 195       nir_register *nreg = clone_register(state, reg);
 196       exec_list_push_tail(dst, &nreg->node);
 197    }
 198 }
 199
 200 static void
 201 __clone_src(clone_state *state, void *ninstr_or_if,
 202             nir_src *nsrc, const nir_src *src)
 203 {
 204    nsrc->is_ssa = src->is_ssa;
 205    if (src->is_ssa) {
 206       nsrc->ssa = remap_local(state, src->ssa);
 207    } else {
 208       nsrc->reg.reg = remap_reg(state, src->reg.reg);
 209       if (src->reg.indirect) {
 210          nsrc->reg.indirect = ralloc(ninstr_or_if, nir_src);
 211          __clone_src(state, ninstr_or_if, nsrc->reg.indirect, src->reg.indirect);
 212       }
 213       nsrc->reg.base_offset = src->reg.base_offset;
 214    }
 215 }
 216
 217 static void
 218 __clone_dst(clone_state *state, nir_instr *ninstr,
 219             nir_dest *ndst, const nir_dest *dst)
 220 {
 221    ndst->is_ssa = dst->is_ssa;
 222    if (dst->is_ssa) {
 223       nir_ssa_dest_init(ninstr, ndst, dst->ssa.num_components, dst->ssa.name);
 224       add_remap(state, &ndst->ssa, &dst->ssa);
 225    } else {
 226       ndst->reg.reg = remap_reg(state, dst->reg.reg);
 227       if (dst->reg.indirect) {
 228          ndst->reg.indirect = ralloc(ninstr, nir_src);
 229          __clone_src(state, ninstr, ndst->reg.indirect, dst->reg.indirect);
 230       }
 231       ndst->reg.base_offset = dst->reg.base_offset;
 232    }
 233 }
 234
 235 static nir_deref *clone_deref(clone_state *state, const nir_deref *deref,
 236                               nir_instr *ninstr, nir_deref *parent);
 237
 238 static nir_deref_var *
 239 clone_deref_var(clone_state *state, const nir_deref_var *dvar,
 240                 nir_instr *ninstr)
 241 {
 242    nir_variable *nvar = remap_var(state, dvar->var);
 243    nir_deref_var *ndvar = nir_deref_var_create(ninstr, nvar);
 244
 245    if (dvar->deref.child)
 246       ndvar->deref.child = clone_deref(state, dvar->deref.child,
 247                                        ninstr, &ndvar->deref);
 248
 249    return ndvar;
 250 }
 251
 252 static nir_deref_array *
 253 clone_deref_array(clone_state *state, const nir_deref_array *darr,
 254                   nir_instr *ninstr, nir_deref *parent)
 255 {
 256    nir_deref_array *ndarr = nir_deref_array_create(parent);
 257
 258    ndarr->deref.type = darr->deref.type;
 259    if (darr->deref.child)
 260       ndarr->deref.child = clone_deref(state, darr->deref.child,
 261                                        ninstr, &ndarr->deref);
 262
 263    ndarr->deref_array_type = darr->deref_array_type;
 264    ndarr->base_offset = darr->base_offset;
 265    if (ndarr->deref_array_type == nir_deref_array_type_indirect)
 266       __clone_src(state, ninstr, &ndarr->indirect, &darr->indirect);
 267
 268    return ndarr;
 269 }
 270
 271 static nir_deref_struct *
 272 clone_deref_struct(clone_state *state, const nir_deref_struct *dstr,
 273                    nir_instr *ninstr, nir_deref *parent)
 274 {
 275    nir_deref_struct *ndstr = nir_deref_struct_create(parent, dstr->index);
 276
 277    ndstr->deref.type = dstr->deref.type;
 278    if (dstr->deref.child)
 279       ndstr->deref.child = clone_deref(state, dstr->deref.child,
 280                                        ninstr, &ndstr->deref);
 281
 282    return ndstr;
 283 }
 284
 285 static nir_deref *
 286 clone_deref(clone_state *state, const nir_deref *dref,
 287             nir_instr *ninstr, nir_deref *parent)
 288 {
 289    switch (dref->deref_type) {
 290    case nir_deref_type_array:
 291       return &clone_deref_array(state, nir_deref_as_array(dref),
 292                                 ninstr, parent)->deref;
 293    case nir_deref_type_struct:
 294       return &clone_deref_struct(state, nir_deref_as_struct(dref),
 295                                  ninstr, parent)->deref;
 296    default:
 297       unreachable("bad deref type");
 298       return NULL;
 299    }
 300 }
 301
 302 static nir_alu_instr *
 303 clone_alu(clone_state *state, const nir_alu_instr *alu)
 304 {
 305    nir_alu_instr *nalu = nir_alu_instr_create(state->ns, alu->op);
 306
 307    __clone_dst(state, &nalu->instr, &nalu->dest.dest, &alu->dest.dest);
 308    nalu->dest.saturate = alu->dest.saturate;
 309    nalu->dest.write_mask = alu->dest.write_mask;
 310
 311    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 312       __clone_src(state, &nalu->instr, &nalu->src[i].src, &alu->src[i].src);
 313       nalu->src[i].negate = alu->src[i].negate;
 314       nalu->src[i].abs = alu->src[i].abs;
 315       memcpy(nalu->src[i].swizzle, alu->src[i].swizzle,
 316              sizeof(nalu->src[i].swizzle));
 317    }
 318
 319    return nalu;
 320 }
 321
 322 static nir_intrinsic_instr *
 323 clone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr)
 324 {
 325    nir_intrinsic_instr *nitr =
 326       nir_intrinsic_instr_create(state->ns, itr->intrinsic);
 327
 328    unsigned num_variables = nir_intrinsic_infos[itr->intrinsic].num_variables;
 329    unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs;
 330
 331    if (nir_intrinsic_infos[itr->intrinsic].has_dest)
 332       __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest);
 333
 334    nitr->num_components = itr->num_components;
 335    memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index));
 336
 337    for (unsigned i = 0; i < num_variables; i++) {
 338       nitr->variables[i] = clone_deref_var(state, itr->variables[i],
 339                                            &nitr->instr);
 340    }
 341
 342    for (unsigned i = 0; i < num_srcs; i++)
 343       __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]);
 344
 345    return nitr;
 346 }
 347
 348 static nir_load_const_instr *
 349 clone_load_const(clone_state *state, const nir_load_const_instr *lc)
 350 {
 351    nir_load_const_instr *nlc =
 352       nir_load_const_instr_create(state->ns, lc->def.num_components);
 353
 354    memcpy(&nlc->value, &lc->value, sizeof(nlc->value));
 355
 356    add_remap(state, &nlc->def, &lc->def);
 357
 358    return nlc;
 359 }
 360
 361 static nir_ssa_undef_instr *
 362 clone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa)
 363 {
 364    nir_ssa_undef_instr *nsa =
 365       nir_ssa_undef_instr_create(state->ns, sa->def.num_components);
 366
 367    add_remap(state, &nsa->def, &sa->def);
 368
 369    return nsa;
 370 }
 371
 372 static nir_tex_instr *
 373 clone_tex(clone_state *state, const nir_tex_instr *tex)
 374 {
 375    nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs);
 376
 377    ntex->sampler_dim = tex->sampler_dim;
 378    ntex->dest_type = tex->dest_type;
 379    ntex->op = tex->op;
 380    __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest);
 381    for (unsigned i = 0; i < ntex->num_srcs; i++) {
 382       ntex->src[i].src_type = tex->src[i].src_type;
 383       __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src);
 384    }
 385    ntex->coord_components = tex->coord_components;
 386    ntex->is_array = tex->is_array;
 387    ntex->is_shadow = tex->is_shadow;
 388    ntex->is_new_style_shadow = tex->is_new_style_shadow;
 389    ntex->component = tex->component;
 390
 391    ntex->texture_index = tex->texture_index;
 392    if (tex->texture)
 393       ntex->texture = clone_deref_var(state, tex->texture, &ntex->instr);
 394    ntex->texture_array_size = tex->texture_array_size;
 395
 396    ntex->sampler_index = tex->sampler_index;
 397    if (tex->sampler)
 398       ntex->sampler = clone_deref_var(state, tex->sampler, &ntex->instr);
 399
 400    return ntex;
 401 }
 402
 403 static nir_phi_instr *
 404 clone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk)
 405 {
 406    nir_phi_instr *nphi = nir_phi_instr_create(state->ns);
 407
 408    __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest);
 409
 410    /* Cloning a phi node is a bit different from other instructions.  The
 411     * sources of phi instructions are the only time where we can use an SSA
 412     * def before it is defined.  In order to handle this, we just copy over
 413     * the sources from the old phi instruction directly and then fix them up
 414     * in a second pass once all the instrutions in the function have been
 415     * properly cloned.
 416     *
 417     * In order to ensure that the copied sources (which are the same as the
 418     * old phi instruction's sources for now) don't get inserted into the old
 419     * shader's use-def lists, we have to add the phi instruction *before* we
 420     * set up its sources.
 421     */
 422    nir_instr_insert_after_block(nblk, &nphi->instr);
 423
 424    foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 425       nir_phi_src *nsrc = ralloc(nphi, nir_phi_src);
 426
 427       /* Just copy the old source for now. */
 428       memcpy(nsrc, src, sizeof(*src));
 429
 430       /* Since we're not letting nir_insert_instr handle use/def stuff for us,
 431        * we have to set the parent_instr manually.  It doesn't really matter
 432        * when we do it, so we might as well do it here.
 433        */
 434       nsrc->src.parent_instr = &nphi->instr;
 435
 436       /* Stash it in the list of phi sources.  We'll walk this list and fix up
 437        * sources at the very end of clone_function_impl.
 438        */
 439       list_add(&nsrc->src.use_link, &state->phi_srcs);
 440
 441       exec_list_push_tail(&nphi->srcs, &nsrc->node);
 442    }
 443
 444    return nphi;
 445 }
 446
 447 static nir_jump_instr *
 448 clone_jump(clone_state *state, const nir_jump_instr *jmp)
 449 {
 450    nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type);
 451
 452    return njmp;
 453 }
 454
 455 static nir_call_instr *
 456 clone_call(clone_state *state, const nir_call_instr *call)
 457 {
 458    nir_function *ncallee = remap_global(state, call->callee);
 459    nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee);
 460
 461    for (unsigned i = 0; i < ncall->num_params; i++)
 462       ncall->params[i] = clone_deref_var(state, call->params[i], &ncall->instr);
 463
 464    ncall->return_deref = clone_deref_var(state, call->return_deref,
 465                                          &ncall->instr);
 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_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(blk, instr) {
 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
 539    __clone_src(state, ni, &ni->condition, &i->condition);
 540
 541    nir_cf_node_insert_end(cf_list, &ni->cf_node);
 542
 543    clone_cf_list(state, &ni->then_list, &i->then_list);
 544    clone_cf_list(state, &ni->else_list, &i->else_list);
 545
 546    return ni;
 547 }
 548
 549 static nir_loop *
 550 clone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop)
 551 {
 552    nir_loop *nloop = nir_loop_create(state->ns);
 553
 554    nir_cf_node_insert_end(cf_list, &nloop->cf_node);
 555
 556    clone_cf_list(state, &nloop->body, &loop->body);
 557
 558    return nloop;
 559 }
 560
 561 /* clone list of nir_cf_node: */
 562 static void
 563 clone_cf_list(clone_state *state, struct exec_list *dst,
 564               const struct exec_list *list)
 565 {
 566    foreach_list_typed(nir_cf_node, cf, node, list) {
 567       switch (cf->type) {
 568       case nir_cf_node_block:
 569          clone_block(state, dst, nir_cf_node_as_block(cf));
 570          break;
 571       case nir_cf_node_if:
 572          clone_if(state, dst, nir_cf_node_as_if(cf));
 573          break;
 574       case nir_cf_node_loop:
 575          clone_loop(state, dst, nir_cf_node_as_loop(cf));
 576          break;
 577       default:
 578          unreachable("bad cf type");
 579       }
 580    }
 581 }
 582
 583 static nir_function_impl *
 584 clone_function_impl(clone_state *state, const nir_function_impl *fi)
 585 {
 586    nir_function_impl *nfi = nir_function_impl_create_bare(state->ns);
 587
 588    clone_var_list(state, &nfi->locals, &fi->locals);
 589    clone_reg_list(state, &nfi->registers, &fi->registers);
 590    nfi->reg_alloc = fi->reg_alloc;
 591
 592    nfi->num_params = fi->num_params;
 593    nfi->params = ralloc_array(state->ns, nir_variable *, fi->num_params);
 594    for (unsigned i = 0; i < fi->num_params; i++) {
 595       nfi->params[i] = clone_variable(state, fi->params[i]);
 596    }
 597    if (fi->return_var)
 598       nfi->return_var = clone_variable(state, fi->return_var);
 599
 600    assert(list_empty(&state->phi_srcs));
 601
 602    clone_cf_list(state, &nfi->body, &fi->body);
 603
 604    /* After we've cloned almost everything, we have to walk the list of phi
 605     * sources and fix them up.  Thanks to loops, the block and SSA value for a
 606     * phi source may not be defined when we first encounter it.  Instead, we
 607     * add it to the phi_srcs list and we fix it up here.
 608     */
 609    list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) {
 610       src->pred = remap_local(state, src->pred);
 611       assert(src->src.is_ssa);
 612       src->src.ssa = remap_local(state, src->src.ssa);
 613
 614       /* Remove from this list and place in the uses of the SSA def */
 615       list_del(&src->src.use_link);
 616       list_addtail(&src->src.use_link, &src->src.ssa->uses);
 617    }
 618    assert(list_empty(&state->phi_srcs));
 619
 620    /* All metadata is invalidated in the cloning process */
 621    nfi->valid_metadata = 0;
 622
 623    return nfi;
 624 }
 625
 626 nir_function_impl *
 627 nir_function_impl_clone(const nir_function_impl *fi)
 628 {
 629    clone_state state;
 630    init_clone_state(&state, false);
 631
 632    /* We use the same shader */
 633    state.ns = fi->function->shader;
 634
 635    nir_function_impl *nfi = clone_function_impl(&state, fi);
 636
 637    free_clone_state(&state);
 638
 639    return nfi;
 640 }
 641
 642 static nir_function *
 643 clone_function(clone_state *state, const nir_function *fxn, nir_shader *ns)
 644 {
 645    assert(ns == state->ns);
 646    nir_function *nfxn = nir_function_create(ns, fxn->name);
 647
 648    /* Needed for call instructions */
 649    add_remap(state, nfxn, fxn);
 650
 651    nfxn->num_params = fxn->num_params;
 652    nfxn->params = ralloc_array(state->ns, nir_parameter, fxn->num_params);
 653    memcpy(nfxn->params, fxn->params, sizeof(nir_parameter) * fxn->num_params);
 654
 655    nfxn->return_type = fxn->return_type;
 656
 657    /* At first glance, it looks like we should clone the function_impl here.
 658     * However, call instructions need to be able to reference at least the
 659     * function and those will get processed as we clone the function_impl's.
 660     * We stop here and do function_impls as a second pass.
 661     */
 662
 663    return nfxn;
 664 }
 665
 666 nir_shader *
 667 nir_shader_clone(void *mem_ctx, const nir_shader *s)
 668 {
 669    clone_state state;
 670    init_clone_state(&state, true);
 671
 672    nir_shader *ns = nir_shader_create(mem_ctx, s->stage, s->options);
 673    state.ns = ns;
 674
 675    clone_var_list(&state, &ns->uniforms, &s->uniforms);
 676    clone_var_list(&state, &ns->inputs,   &s->inputs);
 677    clone_var_list(&state, &ns->outputs,  &s->outputs);
 678    clone_var_list(&state, &ns->shared,   &s->shared);
 679    clone_var_list(&state, &ns->globals,  &s->globals);
 680    clone_var_list(&state, &ns->system_values, &s->system_values);
 681
 682    /* Go through and clone functions */
 683    foreach_list_typed(nir_function, fxn, node, &s->functions)
 684       clone_function(&state, fxn, ns);
 685
 686    /* Only after all functions are cloned can we clone the actual function
 687     * implementations.  This is because nir_call_instr's need to reference the
 688     * functions of other functions and we don't know what order the functions
 689     * will have in the list.
 690     */
 691    nir_foreach_function(s, fxn) {
 692       nir_function *nfxn = remap_global(&state, fxn);
 693       nfxn->impl = clone_function_impl(&state, fxn->impl);
 694       nfxn->impl->function = nfxn;
 695    }
 696
 697    clone_reg_list(&state, &ns->registers, &s->registers);
 698    ns->reg_alloc = s->reg_alloc;
 699
 700    ns->info = s->info;
 701    ns->info.name = ralloc_strdup(ns, ns->info.name);
 702    if (ns->info.label)
 703       ns->info.label = ralloc_strdup(ns, ns->info.label);
 704
 705    ns->num_inputs = s->num_inputs;
 706    ns->num_uniforms = s->num_uniforms;
 707    ns->num_outputs = s->num_outputs;
 708    ns->num_shared = s->num_shared;
 709
 710    free_clone_state(&state);
 711
 712    return ns;
 713 }