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 nir_variable *
 131 nir_variable_clone(const nir_variable *var, nir_shader *shader)
 132 {
 133    nir_variable *nvar = rzalloc(shader, nir_variable);
 134
 135    nvar->type = var->type;
 136    nvar->name = ralloc_strdup(nvar, var->name);
 137    nvar->data = var->data;
 138    nvar->num_state_slots = var->num_state_slots;
 139    nvar->state_slots = ralloc_array(nvar, nir_state_slot, var->num_state_slots);
 140    memcpy(nvar->state_slots, var->state_slots,
 141           var->num_state_slots * sizeof(nir_state_slot));
 142    if (var->constant_initializer) {
 143       nvar->constant_initializer =
 144          nir_constant_clone(var->constant_initializer, nvar);
 145    }
 146    nvar->interface_type = var->interface_type;
 147
 148    return nvar;
 149 }
 150
 151 static nir_variable *
 152 clone_variable(clone_state *state, const nir_variable *var)
 153 {
 154    nir_variable *nvar = nir_variable_clone(var, state->ns);
 155    add_remap(state, nvar, var);
 156
 157    return nvar;
 158 }
 159
 160 /* clone list of nir_variable: */
 161 static void
 162 clone_var_list(clone_state *state, struct exec_list *dst,
 163                const struct exec_list *list)
 164 {
 165    exec_list_make_empty(dst);
 166    foreach_list_typed(nir_variable, var, node, list) {
 167       nir_variable *nvar = clone_variable(state, var);
 168       exec_list_push_tail(dst, &nvar->node);
 169    }
 170 }
 171
 172 /* NOTE: for cloning nir_register's, bypass nir_global/local_reg_create()
 173  * to avoid having to deal with locals and globals separately:
 174  */
 175 static nir_register *
 176 clone_register(clone_state *state, const nir_register *reg)
 177 {
 178    nir_register *nreg = rzalloc(state->ns, nir_register);
 179    add_remap(state, nreg, reg);
 180
 181    nreg->num_components = reg->num_components;
 182    nreg->num_array_elems = reg->num_array_elems;
 183    nreg->index = reg->index;
 184    nreg->name = ralloc_strdup(nreg, reg->name);
 185    nreg->is_global = reg->is_global;
 186    nreg->is_packed = reg->is_packed;
 187
 188    /* reconstructing uses/defs/if_uses handled by nir_instr_insert() */
 189    list_inithead(&nreg->uses);
 190    list_inithead(&nreg->defs);
 191    list_inithead(&nreg->if_uses);
 192
 193    return nreg;
 194 }
 195
 196 /* clone list of nir_register: */
 197 static void
 198 clone_reg_list(clone_state *state, struct exec_list *dst,
 199                const struct exec_list *list)
 200 {
 201    exec_list_make_empty(dst);
 202    foreach_list_typed(nir_register, reg, node, list) {
 203       nir_register *nreg = clone_register(state, reg);
 204       exec_list_push_tail(dst, &nreg->node);
 205    }
 206 }
 207
 208 static void
 209 __clone_src(clone_state *state, void *ninstr_or_if,
 210             nir_src *nsrc, const nir_src *src)
 211 {
 212    nsrc->is_ssa = src->is_ssa;
 213    if (src->is_ssa) {
 214       nsrc->ssa = remap_local(state, src->ssa);
 215    } else {
 216       nsrc->reg.reg = remap_reg(state, src->reg.reg);
 217       if (src->reg.indirect) {
 218          nsrc->reg.indirect = ralloc(ninstr_or_if, nir_src);
 219          __clone_src(state, ninstr_or_if, nsrc->reg.indirect, src->reg.indirect);
 220       }
 221       nsrc->reg.base_offset = src->reg.base_offset;
 222    }
 223 }
 224
 225 static void
 226 __clone_dst(clone_state *state, nir_instr *ninstr,
 227             nir_dest *ndst, const nir_dest *dst)
 228 {
 229    ndst->is_ssa = dst->is_ssa;
 230    if (dst->is_ssa) {
 231       nir_ssa_dest_init(ninstr, ndst, dst->ssa.num_components,
 232                         dst->ssa.bit_size, dst->ssa.name);
 233       add_remap(state, &ndst->ssa, &dst->ssa);
 234    } else {
 235       ndst->reg.reg = remap_reg(state, dst->reg.reg);
 236       if (dst->reg.indirect) {
 237          ndst->reg.indirect = ralloc(ninstr, nir_src);
 238          __clone_src(state, ninstr, ndst->reg.indirect, dst->reg.indirect);
 239       }
 240       ndst->reg.base_offset = dst->reg.base_offset;
 241    }
 242 }
 243
 244 static nir_deref *clone_deref(clone_state *state, const nir_deref *deref,
 245                               nir_instr *ninstr, nir_deref *parent);
 246
 247 static nir_deref_var *
 248 clone_deref_var(clone_state *state, const nir_deref_var *dvar,
 249                 nir_instr *ninstr)
 250 {
 251    nir_variable *nvar = remap_var(state, dvar->var);
 252    nir_deref_var *ndvar = nir_deref_var_create(ninstr, nvar);
 253
 254    if (dvar->deref.child)
 255       ndvar->deref.child = clone_deref(state, dvar->deref.child,
 256                                        ninstr, &ndvar->deref);
 257
 258    return ndvar;
 259 }
 260
 261 static nir_deref_array *
 262 clone_deref_array(clone_state *state, const nir_deref_array *darr,
 263                   nir_instr *ninstr, nir_deref *parent)
 264 {
 265    nir_deref_array *ndarr = nir_deref_array_create(parent);
 266
 267    ndarr->deref.type = darr->deref.type;
 268    if (darr->deref.child)
 269       ndarr->deref.child = clone_deref(state, darr->deref.child,
 270                                        ninstr, &ndarr->deref);
 271
 272    ndarr->deref_array_type = darr->deref_array_type;
 273    ndarr->base_offset = darr->base_offset;
 274    if (ndarr->deref_array_type == nir_deref_array_type_indirect)
 275       __clone_src(state, ninstr, &ndarr->indirect, &darr->indirect);
 276
 277    return ndarr;
 278 }
 279
 280 static nir_deref_struct *
 281 clone_deref_struct(clone_state *state, const nir_deref_struct *dstr,
 282                    nir_instr *ninstr, nir_deref *parent)
 283 {
 284    nir_deref_struct *ndstr = nir_deref_struct_create(parent, dstr->index);
 285
 286    ndstr->deref.type = dstr->deref.type;
 287    if (dstr->deref.child)
 288       ndstr->deref.child = clone_deref(state, dstr->deref.child,
 289                                        ninstr, &ndstr->deref);
 290
 291    return ndstr;
 292 }
 293
 294 static nir_deref *
 295 clone_deref(clone_state *state, const nir_deref *dref,
 296             nir_instr *ninstr, nir_deref *parent)
 297 {
 298    switch (dref->deref_type) {
 299    case nir_deref_type_array:
 300       return &clone_deref_array(state, nir_deref_as_array(dref),
 301                                 ninstr, parent)->deref;
 302    case nir_deref_type_struct:
 303       return &clone_deref_struct(state, nir_deref_as_struct(dref),
 304                                  ninstr, parent)->deref;
 305    default:
 306       unreachable("bad deref type");
 307       return NULL;
 308    }
 309 }
 310
 311 static nir_alu_instr *
 312 clone_alu(clone_state *state, const nir_alu_instr *alu)
 313 {
 314    nir_alu_instr *nalu = nir_alu_instr_create(state->ns, alu->op);
 315    nalu->exact = alu->exact;
 316
 317    __clone_dst(state, &nalu->instr, &nalu->dest.dest, &alu->dest.dest);
 318    nalu->dest.saturate = alu->dest.saturate;
 319    nalu->dest.write_mask = alu->dest.write_mask;
 320
 321    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 322       __clone_src(state, &nalu->instr, &nalu->src[i].src, &alu->src[i].src);
 323       nalu->src[i].negate = alu->src[i].negate;
 324       nalu->src[i].abs = alu->src[i].abs;
 325       memcpy(nalu->src[i].swizzle, alu->src[i].swizzle,
 326              sizeof(nalu->src[i].swizzle));
 327    }
 328
 329    return nalu;
 330 }
 331
 332 static nir_intrinsic_instr *
 333 clone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr)
 334 {
 335    nir_intrinsic_instr *nitr =
 336       nir_intrinsic_instr_create(state->ns, itr->intrinsic);
 337
 338    unsigned num_variables = nir_intrinsic_infos[itr->intrinsic].num_variables;
 339    unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs;
 340
 341    if (nir_intrinsic_infos[itr->intrinsic].has_dest)
 342       __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest);
 343
 344    nitr->num_components = itr->num_components;
 345    memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index));
 346
 347    for (unsigned i = 0; i < num_variables; i++) {
 348       nitr->variables[i] = clone_deref_var(state, itr->variables[i],
 349                                            &nitr->instr);
 350    }
 351
 352    for (unsigned i = 0; i < num_srcs; i++)
 353       __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]);
 354
 355    return nitr;
 356 }
 357
 358 static nir_load_const_instr *
 359 clone_load_const(clone_state *state, const nir_load_const_instr *lc)
 360 {
 361    nir_load_const_instr *nlc =
 362       nir_load_const_instr_create(state->ns, lc->def.num_components,
 363                                   lc->def.bit_size);
 364
 365    memcpy(&nlc->value, &lc->value, sizeof(nlc->value));
 366
 367    add_remap(state, &nlc->def, &lc->def);
 368
 369    return nlc;
 370 }
 371
 372 static nir_ssa_undef_instr *
 373 clone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa)
 374 {
 375    nir_ssa_undef_instr *nsa =
 376       nir_ssa_undef_instr_create(state->ns, sa->def.num_components,
 377                                  sa->def.bit_size);
 378
 379    add_remap(state, &nsa->def, &sa->def);
 380
 381    return nsa;
 382 }
 383
 384 static nir_tex_instr *
 385 clone_tex(clone_state *state, const nir_tex_instr *tex)
 386 {
 387    nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs);
 388
 389    ntex->sampler_dim = tex->sampler_dim;
 390    ntex->dest_type = tex->dest_type;
 391    ntex->op = tex->op;
 392    __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest);
 393    for (unsigned i = 0; i < ntex->num_srcs; i++) {
 394       ntex->src[i].src_type = tex->src[i].src_type;
 395       __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src);
 396    }
 397    ntex->coord_components = tex->coord_components;
 398    ntex->is_array = tex->is_array;
 399    ntex->is_shadow = tex->is_shadow;
 400    ntex->is_new_style_shadow = tex->is_new_style_shadow;
 401    ntex->component = tex->component;
 402
 403    ntex->texture_index = tex->texture_index;
 404    if (tex->texture)
 405       ntex->texture = clone_deref_var(state, tex->texture, &ntex->instr);
 406    ntex->texture_array_size = tex->texture_array_size;
 407
 408    ntex->sampler_index = tex->sampler_index;
 409    if (tex->sampler)
 410       ntex->sampler = clone_deref_var(state, tex->sampler, &ntex->instr);
 411
 412    return ntex;
 413 }
 414
 415 static nir_phi_instr *
 416 clone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk)
 417 {
 418    nir_phi_instr *nphi = nir_phi_instr_create(state->ns);
 419
 420    __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest);
 421
 422    /* Cloning a phi node is a bit different from other instructions.  The
 423     * sources of phi instructions are the only time where we can use an SSA
 424     * def before it is defined.  In order to handle this, we just copy over
 425     * the sources from the old phi instruction directly and then fix them up
 426     * in a second pass once all the instrutions in the function have been
 427     * properly cloned.
 428     *
 429     * In order to ensure that the copied sources (which are the same as the
 430     * old phi instruction's sources for now) don't get inserted into the old
 431     * shader's use-def lists, we have to add the phi instruction *before* we
 432     * set up its sources.
 433     */
 434    nir_instr_insert_after_block(nblk, &nphi->instr);
 435
 436    foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 437       nir_phi_src *nsrc = ralloc(nphi, nir_phi_src);
 438
 439       /* Just copy the old source for now. */
 440       memcpy(nsrc, src, sizeof(*src));
 441
 442       /* Since we're not letting nir_insert_instr handle use/def stuff for us,
 443        * we have to set the parent_instr manually.  It doesn't really matter
 444        * when we do it, so we might as well do it here.
 445        */
 446       nsrc->src.parent_instr = &nphi->instr;
 447
 448       /* Stash it in the list of phi sources.  We'll walk this list and fix up
 449        * sources at the very end of clone_function_impl.
 450        */
 451       list_add(&nsrc->src.use_link, &state->phi_srcs);
 452
 453       exec_list_push_tail(&nphi->srcs, &nsrc->node);
 454    }
 455
 456    return nphi;
 457 }
 458
 459 static nir_jump_instr *
 460 clone_jump(clone_state *state, const nir_jump_instr *jmp)
 461 {
 462    nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type);
 463
 464    return njmp;
 465 }
 466
 467 static nir_call_instr *
 468 clone_call(clone_state *state, const nir_call_instr *call)
 469 {
 470    nir_function *ncallee = remap_global(state, call->callee);
 471    nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee);
 472
 473    for (unsigned i = 0; i < ncall->num_params; i++)
 474       ncall->params[i] = clone_deref_var(state, call->params[i], &ncall->instr);
 475
 476    ncall->return_deref = clone_deref_var(state, call->return_deref,
 477                                          &ncall->instr);
 478
 479    return ncall;
 480 }
 481
 482 static nir_instr *
 483 clone_instr(clone_state *state, const nir_instr *instr)
 484 {
 485    switch (instr->type) {
 486    case nir_instr_type_alu:
 487       return &clone_alu(state, nir_instr_as_alu(instr))->instr;
 488    case nir_instr_type_intrinsic:
 489       return &clone_intrinsic(state, nir_instr_as_intrinsic(instr))->instr;
 490    case nir_instr_type_load_const:
 491       return &clone_load_const(state, nir_instr_as_load_const(instr))->instr;
 492    case nir_instr_type_ssa_undef:
 493       return &clone_ssa_undef(state, nir_instr_as_ssa_undef(instr))->instr;
 494    case nir_instr_type_tex:
 495       return &clone_tex(state, nir_instr_as_tex(instr))->instr;
 496    case nir_instr_type_phi:
 497       unreachable("Cannot clone phis with clone_instr");
 498    case nir_instr_type_jump:
 499       return &clone_jump(state, nir_instr_as_jump(instr))->instr;
 500    case nir_instr_type_call:
 501       return &clone_call(state, nir_instr_as_call(instr))->instr;
 502    case nir_instr_type_parallel_copy:
 503       unreachable("Cannot clone parallel copies");
 504    default:
 505       unreachable("bad instr type");
 506       return NULL;
 507    }
 508 }
 509
 510 static nir_block *
 511 clone_block(clone_state *state, struct exec_list *cf_list, const nir_block *blk)
 512 {
 513    /* Don't actually create a new block.  Just use the one from the tail of
 514     * the list.  NIR guarantees that the tail of the list is a block and that
 515     * no two blocks are side-by-side in the IR;  It should be empty.
 516     */
 517    nir_block *nblk =
 518       exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node);
 519    assert(nblk->cf_node.type == nir_cf_node_block);
 520    assert(exec_list_is_empty(&nblk->instr_list));
 521
 522    /* We need this for phi sources */
 523    add_remap(state, nblk, blk);
 524
 525    nir_foreach_instr(blk, instr) {
 526       if (instr->type == nir_instr_type_phi) {
 527          /* Phi instructions are a bit of a special case when cloning because
 528           * we don't want inserting the instruction to automatically handle
 529           * use/defs for us.  Instead, we need to wait until all the
 530           * blocks/instructions are in so that we can set their sources up.
 531           */
 532          clone_phi(state, nir_instr_as_phi(instr), nblk);
 533       } else {
 534          nir_instr *ninstr = clone_instr(state, instr);
 535          nir_instr_insert_after_block(nblk, ninstr);
 536       }
 537    }
 538
 539    return nblk;
 540 }
 541
 542 static void
 543 clone_cf_list(clone_state *state, struct exec_list *dst,
 544               const struct exec_list *list);
 545
 546 static nir_if *
 547 clone_if(clone_state *state, struct exec_list *cf_list, const nir_if *i)
 548 {
 549    nir_if *ni = nir_if_create(state->ns);
 550
 551    __clone_src(state, ni, &ni->condition, &i->condition);
 552
 553    nir_cf_node_insert_end(cf_list, &ni->cf_node);
 554
 555    clone_cf_list(state, &ni->then_list, &i->then_list);
 556    clone_cf_list(state, &ni->else_list, &i->else_list);
 557
 558    return ni;
 559 }
 560
 561 static nir_loop *
 562 clone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop)
 563 {
 564    nir_loop *nloop = nir_loop_create(state->ns);
 565
 566    nir_cf_node_insert_end(cf_list, &nloop->cf_node);
 567
 568    clone_cf_list(state, &nloop->body, &loop->body);
 569
 570    return nloop;
 571 }
 572
 573 /* clone list of nir_cf_node: */
 574 static void
 575 clone_cf_list(clone_state *state, struct exec_list *dst,
 576               const struct exec_list *list)
 577 {
 578    foreach_list_typed(nir_cf_node, cf, node, list) {
 579       switch (cf->type) {
 580       case nir_cf_node_block:
 581          clone_block(state, dst, nir_cf_node_as_block(cf));
 582          break;
 583       case nir_cf_node_if:
 584          clone_if(state, dst, nir_cf_node_as_if(cf));
 585          break;
 586       case nir_cf_node_loop:
 587          clone_loop(state, dst, nir_cf_node_as_loop(cf));
 588          break;
 589       default:
 590          unreachable("bad cf type");
 591       }
 592    }
 593 }
 594
 595 static nir_function_impl *
 596 clone_function_impl(clone_state *state, const nir_function_impl *fi)
 597 {
 598    nir_function_impl *nfi = nir_function_impl_create_bare(state->ns);
 599
 600    clone_var_list(state, &nfi->locals, &fi->locals);
 601    clone_reg_list(state, &nfi->registers, &fi->registers);
 602    nfi->reg_alloc = fi->reg_alloc;
 603
 604    nfi->num_params = fi->num_params;
 605    nfi->params = ralloc_array(state->ns, nir_variable *, fi->num_params);
 606    for (unsigned i = 0; i < fi->num_params; i++) {
 607       nfi->params[i] = clone_variable(state, fi->params[i]);
 608    }
 609    if (fi->return_var)
 610       nfi->return_var = clone_variable(state, fi->return_var);
 611
 612    assert(list_empty(&state->phi_srcs));
 613
 614    clone_cf_list(state, &nfi->body, &fi->body);
 615
 616    /* After we've cloned almost everything, we have to walk the list of phi
 617     * sources and fix them up.  Thanks to loops, the block and SSA value for a
 618     * phi source may not be defined when we first encounter it.  Instead, we
 619     * add it to the phi_srcs list and we fix it up here.
 620     */
 621    list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) {
 622       src->pred = remap_local(state, src->pred);
 623       assert(src->src.is_ssa);
 624       src->src.ssa = remap_local(state, src->src.ssa);
 625
 626       /* Remove from this list and place in the uses of the SSA def */
 627       list_del(&src->src.use_link);
 628       list_addtail(&src->src.use_link, &src->src.ssa->uses);
 629    }
 630    assert(list_empty(&state->phi_srcs));
 631
 632    /* All metadata is invalidated in the cloning process */
 633    nfi->valid_metadata = 0;
 634
 635    return nfi;
 636 }
 637
 638 nir_function_impl *
 639 nir_function_impl_clone(const nir_function_impl *fi)
 640 {
 641    clone_state state;
 642    init_clone_state(&state, false);
 643
 644    /* We use the same shader */
 645    state.ns = fi->function->shader;
 646
 647    nir_function_impl *nfi = clone_function_impl(&state, fi);
 648
 649    free_clone_state(&state);
 650
 651    return nfi;
 652 }
 653
 654 static nir_function *
 655 clone_function(clone_state *state, const nir_function *fxn, nir_shader *ns)
 656 {
 657    assert(ns == state->ns);
 658    nir_function *nfxn = nir_function_create(ns, fxn->name);
 659
 660    /* Needed for call instructions */
 661    add_remap(state, nfxn, fxn);
 662
 663    nfxn->num_params = fxn->num_params;
 664    nfxn->params = ralloc_array(state->ns, nir_parameter, fxn->num_params);
 665    memcpy(nfxn->params, fxn->params, sizeof(nir_parameter) * fxn->num_params);
 666
 667    nfxn->return_type = fxn->return_type;
 668
 669    /* At first glance, it looks like we should clone the function_impl here.
 670     * However, call instructions need to be able to reference at least the
 671     * function and those will get processed as we clone the function_impl's.
 672     * We stop here and do function_impls as a second pass.
 673     */
 674
 675    return nfxn;
 676 }
 677
 678 nir_shader *
 679 nir_shader_clone(void *mem_ctx, const nir_shader *s)
 680 {
 681    clone_state state;
 682    init_clone_state(&state, true);
 683
 684    nir_shader *ns = nir_shader_create(mem_ctx, s->stage, s->options);
 685    state.ns = ns;
 686
 687    clone_var_list(&state, &ns->uniforms, &s->uniforms);
 688    clone_var_list(&state, &ns->inputs,   &s->inputs);
 689    clone_var_list(&state, &ns->outputs,  &s->outputs);
 690    clone_var_list(&state, &ns->shared,   &s->shared);
 691    clone_var_list(&state, &ns->globals,  &s->globals);
 692    clone_var_list(&state, &ns->system_values, &s->system_values);
 693
 694    /* Go through and clone functions */
 695    foreach_list_typed(nir_function, fxn, node, &s->functions)
 696       clone_function(&state, fxn, ns);
 697
 698    /* Only after all functions are cloned can we clone the actual function
 699     * implementations.  This is because nir_call_instr's need to reference the
 700     * functions of other functions and we don't know what order the functions
 701     * will have in the list.
 702     */
 703    nir_foreach_function(s, fxn) {
 704       nir_function *nfxn = remap_global(&state, fxn);
 705       nfxn->impl = clone_function_impl(&state, fxn->impl);
 706       nfxn->impl->function = nfxn;
 707    }
 708
 709    clone_reg_list(&state, &ns->registers, &s->registers);
 710    ns->reg_alloc = s->reg_alloc;
 711
 712    ns->info = s->info;
 713    ns->info.name = ralloc_strdup(ns, ns->info.name);
 714    if (ns->info.label)
 715       ns->info.label = ralloc_strdup(ns, ns->info.label);
 716
 717    ns->num_inputs = s->num_inputs;
 718    ns->num_uniforms = s->num_uniforms;
 719    ns->num_outputs = s->num_outputs;
 720    ns->num_shared = s->num_shared;
 721
 722    free_clone_state(&state);
 723
 724    return ns;
 725 }