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