src/compiler/nir/nir_clone.c

   1 /*
   2  * Copyright © 2015 Red Hat
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "nir.h"
  25 #include "nir_control_flow.h"
  26
  27 /* Secret Decoder Ring:
  28  *   clone_foo():
  29  *        Allocate and clone a foo.
  30  *   __clone_foo():
  31  *        Clone body of foo (ie. parent class, embedded struct, etc)
  32  */
  33
  34 typedef struct {
  35    /* True if we are cloning an entire shader. */
  36    bool global_clone;
  37
  38    /* If true allows the clone operation to fall back to the original pointer
  39     * if no clone pointer is found in the remap table.  This allows us to
  40     * clone a loop body without having to add srcs from outside the loop to
  41     * the remap table. This is useful for loop unrolling.
  42     */
  43    bool allow_remap_fallback;
  44
  45    /* maps orig ptr -> cloned ptr: */
  46    struct hash_table *remap_table;
  47
  48    /* List of phi sources. */
  49    struct list_head phi_srcs;
  50
  51    /* new shader object, used as memctx for just about everything else: */
  52    nir_shader *ns;
  53 } clone_state;
  54
  55 static void
  56 init_clone_state(clone_state *state, struct hash_table *remap_table,
  57                  bool global, bool allow_remap_fallback)
  58 {
  59    state->global_clone = global;
  60    state->allow_remap_fallback = allow_remap_fallback;
  61
  62    if (remap_table) {
  63       state->remap_table = remap_table;
  64    } else {
  65       state->remap_table = _mesa_pointer_hash_table_create(NULL);
  66    }
  67
  68    list_inithead(&state->phi_srcs);
  69 }
  70
  71 static void
  72 free_clone_state(clone_state *state)
  73 {
  74    _mesa_hash_table_destroy(state->remap_table, NULL);
  75 }
  76
  77 static inline void *
  78 _lookup_ptr(clone_state *state, const void *ptr, bool global)
  79 {
  80    struct hash_entry *entry;
  81
  82    if (!ptr)
  83       return NULL;
  84
  85    if (!state->global_clone && global)
  86       return (void *)ptr;
  87
  88    if (unlikely(!state->remap_table)) {
  89       assert(state->allow_remap_fallback);
  90       return (void *)ptr;
  91    }
  92
  93    entry = _mesa_hash_table_search(state->remap_table, ptr);
  94    if (!entry) {
  95       assert(state->allow_remap_fallback);
  96       return (void *)ptr;
  97    }
  98
  99    return entry->data;
 100 }
 101
 102 static void
 103 add_remap(clone_state *state, void *nptr, const void *ptr)
 104 {
 105    _mesa_hash_table_insert(state->remap_table, ptr, nptr);
 106 }
 107
 108 static void *
 109 remap_local(clone_state *state, const void *ptr)
 110 {
 111    return _lookup_ptr(state, ptr, false);
 112 }
 113
 114 static void *
 115 remap_global(clone_state *state, const void *ptr)
 116 {
 117    return _lookup_ptr(state, ptr, true);
 118 }
 119
 120 static nir_register *
 121 remap_reg(clone_state *state, const nir_register *reg)
 122 {
 123    return _lookup_ptr(state, reg, false);
 124 }
 125
 126 static nir_variable *
 127 remap_var(clone_state *state, const nir_variable *var)
 128 {
 129    return _lookup_ptr(state, var, nir_variable_is_global(var));
 130 }
 131
 132 nir_constant *
 133 nir_constant_clone(const nir_constant *c, nir_variable *nvar)
 134 {
 135    nir_constant *nc = ralloc(nvar, nir_constant);
 136
 137    memcpy(nc->values, c->values, sizeof(nc->values));
 138    nc->num_elements = c->num_elements;
 139    nc->elements = ralloc_array(nvar, nir_constant *, c->num_elements);
 140    for (unsigned i = 0; i < c->num_elements; i++) {
 141       nc->elements[i] = nir_constant_clone(c->elements[i], nvar);
 142    }
 143
 144    return nc;
 145 }
 146
 147 /* NOTE: for cloning nir_variables, bypass nir_variable_create to avoid
 148  * having to deal with locals and globals separately:
 149  */
 150 nir_variable *
 151 nir_variable_clone(const nir_variable *var, nir_shader *shader)
 152 {
 153    nir_variable *nvar = rzalloc(shader, nir_variable);
 154
 155    nvar->type = var->type;
 156    nvar->name = ralloc_strdup(nvar, var->name);
 157    nvar->data = var->data;
 158    nvar->num_state_slots = var->num_state_slots;
 159    if (var->num_state_slots) {
 160       nvar->state_slots = ralloc_array(nvar, nir_state_slot, var->num_state_slots);
 161       memcpy(nvar->state_slots, var->state_slots,
 162              var->num_state_slots * sizeof(nir_state_slot));
 163    }
 164    if (var->constant_initializer) {
 165       nvar->constant_initializer =
 166          nir_constant_clone(var->constant_initializer, nvar);
 167    }
 168    nvar->interface_type = var->interface_type;
 169
 170    nvar->num_members = var->num_members;
 171    if (var->num_members) {
 172       nvar->members = ralloc_array(nvar, struct nir_variable_data,
 173                                    var->num_members);
 174       memcpy(nvar->members, var->members,
 175              var->num_members * sizeof(*var->members));
 176    }
 177
 178    return nvar;
 179 }
 180
 181 static nir_variable *
 182 clone_variable(clone_state *state, const nir_variable *var)
 183 {
 184    nir_variable *nvar = nir_variable_clone(var, state->ns);
 185    add_remap(state, nvar, var);
 186
 187    return nvar;
 188 }
 189
 190 /* clone list of nir_variable: */
 191 static void
 192 clone_var_list(clone_state *state, struct exec_list *dst,
 193                const struct exec_list *list)
 194 {
 195    exec_list_make_empty(dst);
 196    foreach_list_typed(nir_variable, var, node, list) {
 197       nir_variable *nvar = clone_variable(state, var);
 198       exec_list_push_tail(dst, &nvar->node);
 199    }
 200 }
 201
 202 /* NOTE: for cloning nir_registers, bypass nir_global/local_reg_create()
 203  * to avoid having to deal with locals and globals separately:
 204  */
 205 static nir_register *
 206 clone_register(clone_state *state, const nir_register *reg)
 207 {
 208    nir_register *nreg = rzalloc(state->ns, nir_register);
 209    add_remap(state, nreg, reg);
 210
 211    nreg->num_components = reg->num_components;
 212    nreg->bit_size = reg->bit_size;
 213    nreg->num_array_elems = reg->num_array_elems;
 214    nreg->index = reg->index;
 215    nreg->name = ralloc_strdup(nreg, reg->name);
 216
 217    /* reconstructing uses/defs/if_uses handled by nir_instr_insert() */
 218    list_inithead(&nreg->uses);
 219    list_inithead(&nreg->defs);
 220    list_inithead(&nreg->if_uses);
 221
 222    return nreg;
 223 }
 224
 225 /* clone list of nir_register: */
 226 static void
 227 clone_reg_list(clone_state *state, struct exec_list *dst,
 228                const struct exec_list *list)
 229 {
 230    exec_list_make_empty(dst);
 231    foreach_list_typed(nir_register, reg, node, list) {
 232       nir_register *nreg = clone_register(state, reg);
 233       exec_list_push_tail(dst, &nreg->node);
 234    }
 235 }
 236
 237 static void
 238 __clone_src(clone_state *state, void *ninstr_or_if,
 239             nir_src *nsrc, const nir_src *src)
 240 {
 241    nsrc->is_ssa = src->is_ssa;
 242    if (src->is_ssa) {
 243       nsrc->ssa = remap_local(state, src->ssa);
 244    } else {
 245       nsrc->reg.reg = remap_reg(state, src->reg.reg);
 246       if (src->reg.indirect) {
 247          nsrc->reg.indirect = ralloc(ninstr_or_if, nir_src);
 248          __clone_src(state, ninstr_or_if, nsrc->reg.indirect, src->reg.indirect);
 249       }
 250       nsrc->reg.base_offset = src->reg.base_offset;
 251    }
 252 }
 253
 254 static void
 255 __clone_dst(clone_state *state, nir_instr *ninstr,
 256             nir_dest *ndst, const nir_dest *dst)
 257 {
 258    ndst->is_ssa = dst->is_ssa;
 259    if (dst->is_ssa) {
 260       nir_ssa_dest_init(ninstr, ndst, dst->ssa.num_components,
 261                         dst->ssa.bit_size, dst->ssa.name);
 262       if (likely(state->remap_table))
 263          add_remap(state, &ndst->ssa, &dst->ssa);
 264    } else {
 265       ndst->reg.reg = remap_reg(state, dst->reg.reg);
 266       if (dst->reg.indirect) {
 267          ndst->reg.indirect = ralloc(ninstr, nir_src);
 268          __clone_src(state, ninstr, ndst->reg.indirect, dst->reg.indirect);
 269       }
 270       ndst->reg.base_offset = dst->reg.base_offset;
 271    }
 272 }
 273
 274 static nir_alu_instr *
 275 clone_alu(clone_state *state, const nir_alu_instr *alu)
 276 {
 277    nir_alu_instr *nalu = nir_alu_instr_create(state->ns, alu->op);
 278    nalu->exact = alu->exact;
 279    nalu->no_signed_wrap = alu->no_signed_wrap;
 280    nalu->no_unsigned_wrap = alu->no_unsigned_wrap;
 281
 282    __clone_dst(state, &nalu->instr, &nalu->dest.dest, &alu->dest.dest);
 283    nalu->dest.saturate = alu->dest.saturate;
 284    nalu->dest.write_mask = alu->dest.write_mask;
 285
 286    for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
 287       __clone_src(state, &nalu->instr, &nalu->src[i].src, &alu->src[i].src);
 288       nalu->src[i].negate = alu->src[i].negate;
 289       nalu->src[i].abs = alu->src[i].abs;
 290       memcpy(nalu->src[i].swizzle, alu->src[i].swizzle,
 291              sizeof(nalu->src[i].swizzle));
 292    }
 293
 294    return nalu;
 295 }
 296
 297 nir_alu_instr *
 298 nir_alu_instr_clone(nir_shader *shader, const nir_alu_instr *orig)
 299 {
 300    clone_state state = {
 301       .allow_remap_fallback = true,
 302       .ns = shader,
 303    };
 304    return clone_alu(&state, orig);
 305 }
 306
 307 static nir_deref_instr *
 308 clone_deref_instr(clone_state *state, const nir_deref_instr *deref)
 309 {
 310    nir_deref_instr *nderef =
 311       nir_deref_instr_create(state->ns, deref->deref_type);
 312
 313    __clone_dst(state, &nderef->instr, &nderef->dest, &deref->dest);
 314
 315    nderef->mode = deref->mode;
 316    nderef->type = deref->type;
 317
 318    if (deref->deref_type == nir_deref_type_var) {
 319       nderef->var = remap_var(state, deref->var);
 320       return nderef;
 321    }
 322
 323    __clone_src(state, &nderef->instr, &nderef->parent, &deref->parent);
 324
 325    switch (deref->deref_type) {
 326    case nir_deref_type_struct:
 327       nderef->strct.index = deref->strct.index;
 328       break;
 329
 330    case nir_deref_type_array:
 331    case nir_deref_type_ptr_as_array:
 332       __clone_src(state, &nderef->instr,
 333                   &nderef->arr.index, &deref->arr.index);
 334       break;
 335
 336    case nir_deref_type_array_wildcard:
 337       /* Nothing to do */
 338       break;
 339
 340    case nir_deref_type_cast:
 341       nderef->cast.ptr_stride = deref->cast.ptr_stride;
 342       break;
 343
 344    default:
 345       unreachable("Invalid instruction deref type");
 346    }
 347
 348    return nderef;
 349 }
 350
 351 static nir_intrinsic_instr *
 352 clone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr)
 353 {
 354    nir_intrinsic_instr *nitr =
 355       nir_intrinsic_instr_create(state->ns, itr->intrinsic);
 356
 357    unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs;
 358
 359    if (nir_intrinsic_infos[itr->intrinsic].has_dest)
 360       __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest);
 361
 362    nitr->num_components = itr->num_components;
 363    memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index));
 364
 365    for (unsigned i = 0; i < num_srcs; i++)
 366       __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]);
 367
 368    return nitr;
 369 }
 370
 371 static nir_load_const_instr *
 372 clone_load_const(clone_state *state, const nir_load_const_instr *lc)
 373 {
 374    nir_load_const_instr *nlc =
 375       nir_load_const_instr_create(state->ns, lc->def.num_components,
 376                                   lc->def.bit_size);
 377
 378    memcpy(&nlc->value, &lc->value, sizeof(*nlc->value) * lc->def.num_components);
 379
 380    add_remap(state, &nlc->def, &lc->def);
 381
 382    return nlc;
 383 }
 384
 385 static nir_ssa_undef_instr *
 386 clone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa)
 387 {
 388    nir_ssa_undef_instr *nsa =
 389       nir_ssa_undef_instr_create(state->ns, sa->def.num_components,
 390                                  sa->def.bit_size);
 391
 392    add_remap(state, &nsa->def, &sa->def);
 393
 394    return nsa;
 395 }
 396
 397 static nir_tex_instr *
 398 clone_tex(clone_state *state, const nir_tex_instr *tex)
 399 {
 400    nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs);
 401
 402    ntex->sampler_dim = tex->sampler_dim;
 403    ntex->dest_type = tex->dest_type;
 404    ntex->op = tex->op;
 405    __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest);
 406    for (unsigned i = 0; i < ntex->num_srcs; i++) {
 407       ntex->src[i].src_type = tex->src[i].src_type;
 408       __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src);
 409    }
 410    ntex->coord_components = tex->coord_components;
 411    ntex->is_array = tex->is_array;
 412    ntex->is_shadow = tex->is_shadow;
 413    ntex->is_new_style_shadow = tex->is_new_style_shadow;
 414    ntex->component = tex->component;
 415    memcpy(ntex->tg4_offsets, tex->tg4_offsets, sizeof(tex->tg4_offsets));
 416
 417    ntex->texture_index = tex->texture_index;
 418    ntex->sampler_index = tex->sampler_index;
 419
 420    ntex->texture_non_uniform = tex->texture_non_uniform;
 421    ntex->sampler_non_uniform = tex->sampler_non_uniform;
 422
 423    return ntex;
 424 }
 425
 426 static nir_phi_instr *
 427 clone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk)
 428 {
 429    nir_phi_instr *nphi = nir_phi_instr_create(state->ns);
 430
 431    __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest);
 432
 433    /* Cloning a phi node is a bit different from other instructions.  The
 434     * sources of phi instructions are the only time where we can use an SSA
 435     * def before it is defined.  In order to handle this, we just copy over
 436     * the sources from the old phi instruction directly and then fix them up
 437     * in a second pass once all the instrutions in the function have been
 438     * properly cloned.
 439     *
 440     * In order to ensure that the copied sources (which are the same as the
 441     * old phi instruction's sources for now) don't get inserted into the old
 442     * shader's use-def lists, we have to add the phi instruction *before* we
 443     * set up its sources.
 444     */
 445    nir_instr_insert_after_block(nblk, &nphi->instr);
 446
 447    foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
 448       nir_phi_src *nsrc = ralloc(nphi, nir_phi_src);
 449
 450       /* Just copy the old source for now. */
 451       memcpy(nsrc, src, sizeof(*src));
 452
 453       /* Since we're not letting nir_insert_instr handle use/def stuff for us,
 454        * we have to set the parent_instr manually.  It doesn't really matter
 455        * when we do it, so we might as well do it here.
 456        */
 457       nsrc->src.parent_instr = &nphi->instr;
 458
 459       /* Stash it in the list of phi sources.  We'll walk this list and fix up
 460        * sources at the very end of clone_function_impl.
 461        */
 462       list_add(&nsrc->src.use_link, &state->phi_srcs);
 463
 464       exec_list_push_tail(&nphi->srcs, &nsrc->node);
 465    }
 466
 467    return nphi;
 468 }
 469
 470 static nir_jump_instr *
 471 clone_jump(clone_state *state, const nir_jump_instr *jmp)
 472 {
 473    nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type);
 474
 475    return njmp;
 476 }
 477
 478 static nir_call_instr *
 479 clone_call(clone_state *state, const nir_call_instr *call)
 480 {
 481    nir_function *ncallee = remap_global(state, call->callee);
 482    nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee);
 483
 484    for (unsigned i = 0; i < ncall->num_params; i++)
 485       __clone_src(state, ncall, &ncall->params[i], &call->params[i]);
 486
 487    return ncall;
 488 }
 489
 490 static nir_instr *
 491 clone_instr(clone_state *state, const nir_instr *instr)
 492 {
 493    switch (instr->type) {
 494    case nir_instr_type_alu:
 495       return &clone_alu(state, nir_instr_as_alu(instr))->instr;
 496    case nir_instr_type_deref:
 497       return &clone_deref_instr(state, nir_instr_as_deref(instr))->instr;
 498    case nir_instr_type_intrinsic:
 499       return &clone_intrinsic(state, nir_instr_as_intrinsic(instr))->instr;
 500    case nir_instr_type_load_const:
 501       return &clone_load_const(state, nir_instr_as_load_const(instr))->instr;
 502    case nir_instr_type_ssa_undef:
 503       return &clone_ssa_undef(state, nir_instr_as_ssa_undef(instr))->instr;
 504    case nir_instr_type_tex:
 505       return &clone_tex(state, nir_instr_as_tex(instr))->instr;
 506    case nir_instr_type_phi:
 507       unreachable("Cannot clone phis with clone_instr");
 508    case nir_instr_type_jump:
 509       return &clone_jump(state, nir_instr_as_jump(instr))->instr;
 510    case nir_instr_type_call:
 511       return &clone_call(state, nir_instr_as_call(instr))->instr;
 512    case nir_instr_type_parallel_copy:
 513       unreachable("Cannot clone parallel copies");
 514    default:
 515       unreachable("bad instr type");
 516       return NULL;
 517    }
 518 }
 519
 520 nir_instr *
 521 nir_instr_clone(nir_shader *shader, const nir_instr *orig)
 522 {
 523    clone_state state = {
 524       .allow_remap_fallback = true,
 525       .ns = shader,
 526    };
 527    return clone_instr(&state, orig);
 528 }
 529
 530 static nir_block *
 531 clone_block(clone_state *state, struct exec_list *cf_list, const nir_block *blk)
 532 {
 533    /* Don't actually create a new block.  Just use the one from the tail of
 534     * the list.  NIR guarantees that the tail of the list is a block and that
 535     * no two blocks are side-by-side in the IR;  It should be empty.
 536     */
 537    nir_block *nblk =
 538       exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node);
 539    assert(nblk->cf_node.type == nir_cf_node_block);
 540    assert(exec_list_is_empty(&nblk->instr_list));
 541
 542    /* We need this for phi sources */
 543    add_remap(state, nblk, blk);
 544
 545    nir_foreach_instr(instr, blk) {
 546       if (instr->type == nir_instr_type_phi) {
 547          /* Phi instructions are a bit of a special case when cloning because
 548           * we don't want inserting the instruction to automatically handle
 549           * use/defs for us.  Instead, we need to wait until all the
 550           * blocks/instructions are in so that we can set their sources up.
 551           */
 552          clone_phi(state, nir_instr_as_phi(instr), nblk);
 553       } else {
 554          nir_instr *ninstr = clone_instr(state, instr);
 555          nir_instr_insert_after_block(nblk, ninstr);
 556       }
 557    }
 558
 559    return nblk;
 560 }
 561
 562 static void
 563 clone_cf_list(clone_state *state, struct exec_list *dst,
 564               const struct exec_list *list);
 565
 566 static nir_if *
 567 clone_if(clone_state *state, struct exec_list *cf_list, const nir_if *i)
 568 {
 569    nir_if *ni = nir_if_create(state->ns);
 570    ni->control = i->control;
 571
 572    __clone_src(state, ni, &ni->condition, &i->condition);
 573
 574    nir_cf_node_insert_end(cf_list, &ni->cf_node);
 575
 576    clone_cf_list(state, &ni->then_list, &i->then_list);
 577    clone_cf_list(state, &ni->else_list, &i->else_list);
 578
 579    return ni;
 580 }
 581
 582 static nir_loop *
 583 clone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop)
 584 {
 585    nir_loop *nloop = nir_loop_create(state->ns);
 586    nloop->control = loop->control;
 587    nloop->partially_unrolled = loop->partially_unrolled;
 588
 589    nir_cf_node_insert_end(cf_list, &nloop->cf_node);
 590
 591    clone_cf_list(state, &nloop->body, &loop->body);
 592
 593    return nloop;
 594 }
 595
 596 /* clone list of nir_cf_node: */
 597 static void
 598 clone_cf_list(clone_state *state, struct exec_list *dst,
 599               const struct exec_list *list)
 600 {
 601    foreach_list_typed(nir_cf_node, cf, node, list) {
 602       switch (cf->type) {
 603       case nir_cf_node_block:
 604          clone_block(state, dst, nir_cf_node_as_block(cf));
 605          break;
 606       case nir_cf_node_if:
 607          clone_if(state, dst, nir_cf_node_as_if(cf));
 608          break;
 609       case nir_cf_node_loop:
 610          clone_loop(state, dst, nir_cf_node_as_loop(cf));
 611          break;
 612       default:
 613          unreachable("bad cf type");
 614       }
 615    }
 616 }
 617
 618 /* After we've cloned almost everything, we have to walk the list of phi
 619  * sources and fix them up.  Thanks to loops, the block and SSA value for a
 620  * phi source may not be defined when we first encounter it.  Instead, we
 621  * add it to the phi_srcs list and we fix it up here.
 622  */
 623 static void
 624 fixup_phi_srcs(clone_state *state)
 625 {
 626    list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) {
 627       src->pred = remap_local(state, src->pred);
 628
 629       /* Remove from this list */
 630       list_del(&src->src.use_link);
 631
 632       if (src->src.is_ssa) {
 633          src->src.ssa = remap_local(state, src->src.ssa);
 634          list_addtail(&src->src.use_link, &src->src.ssa->uses);
 635       } else {
 636          src->src.reg.reg = remap_reg(state, src->src.reg.reg);
 637          list_addtail(&src->src.use_link, &src->src.reg.reg->uses);
 638       }
 639    }
 640    assert(list_is_empty(&state->phi_srcs));
 641 }
 642
 643 void
 644 nir_cf_list_clone(nir_cf_list *dst, nir_cf_list *src, nir_cf_node *parent,
 645                   struct hash_table *remap_table)
 646 {
 647    exec_list_make_empty(&dst->list);
 648    dst->impl = src->impl;
 649
 650    if (exec_list_is_empty(&src->list))
 651       return;
 652
 653    clone_state state;
 654    init_clone_state(&state, remap_table, false, true);
 655
 656    /* We use the same shader */
 657    state.ns = src->impl->function->shader;
 658
 659    /* The control-flow code assumes that the list of cf_nodes always starts
 660     * and ends with a block.  We start by adding an empty block.
 661     */
 662    nir_block *nblk = nir_block_create(state.ns);
 663    nblk->cf_node.parent = parent;
 664    exec_list_push_tail(&dst->list, &nblk->cf_node.node);
 665
 666    clone_cf_list(&state, &dst->list, &src->list);
 667
 668    fixup_phi_srcs(&state);
 669
 670    if (!remap_table)
 671       free_clone_state(&state);
 672 }
 673
 674 static nir_function_impl *
 675 clone_function_impl(clone_state *state, const nir_function_impl *fi)
 676 {
 677    nir_function_impl *nfi = nir_function_impl_create_bare(state->ns);
 678
 679    clone_var_list(state, &nfi->locals, &fi->locals);
 680    clone_reg_list(state, &nfi->registers, &fi->registers);
 681    nfi->reg_alloc = fi->reg_alloc;
 682
 683    assert(list_is_empty(&state->phi_srcs));
 684
 685    clone_cf_list(state, &nfi->body, &fi->body);
 686
 687    fixup_phi_srcs(state);
 688
 689    /* All metadata is invalidated in the cloning process */
 690    nfi->valid_metadata = 0;
 691
 692    return nfi;
 693 }
 694
 695 nir_function_impl *
 696 nir_function_impl_clone(nir_shader *shader, const nir_function_impl *fi)
 697 {
 698    clone_state state;
 699    init_clone_state(&state, NULL, false, false);
 700
 701    state.ns = shader;
 702
 703    nir_function_impl *nfi = clone_function_impl(&state, fi);
 704
 705    free_clone_state(&state);
 706
 707    return nfi;
 708 }
 709
 710 static nir_function *
 711 clone_function(clone_state *state, const nir_function *fxn, nir_shader *ns)
 712 {
 713    assert(ns == state->ns);
 714    nir_function *nfxn = nir_function_create(ns, fxn->name);
 715
 716    /* Needed for call instructions */
 717    add_remap(state, nfxn, fxn);
 718
 719    nfxn->num_params = fxn->num_params;
 720    if (fxn->num_params) {
 721            nfxn->params = ralloc_array(state->ns, nir_parameter, fxn->num_params);
 722            memcpy(nfxn->params, fxn->params, sizeof(nir_parameter) * fxn->num_params);
 723    }
 724    nfxn->is_entrypoint = fxn->is_entrypoint;
 725
 726    /* At first glance, it looks like we should clone the function_impl here.
 727     * However, call instructions need to be able to reference at least the
 728     * function and those will get processed as we clone the function_impls.
 729     * We stop here and do function_impls as a second pass.
 730     */
 731
 732    return nfxn;
 733 }
 734
 735 nir_shader *
 736 nir_shader_clone(void *mem_ctx, const nir_shader *s)
 737 {
 738    clone_state state;
 739    init_clone_state(&state, NULL, true, false);
 740
 741    nir_shader *ns = nir_shader_create(mem_ctx, s->info.stage, s->options, NULL);
 742    state.ns = ns;
 743
 744    clone_var_list(&state, &ns->variables, &s->variables);
 745
 746    /* Go through and clone functions */
 747    foreach_list_typed(nir_function, fxn, node, &s->functions)
 748       clone_function(&state, fxn, ns);
 749
 750    /* Only after all functions are cloned can we clone the actual function
 751     * implementations.  This is because nir_call_instrs need to reference the
 752     * functions of other functions and we don't know what order the functions
 753     * will have in the list.
 754     */
 755    nir_foreach_function(fxn, s) {
 756       nir_function *nfxn = remap_global(&state, fxn);
 757       nfxn->impl = clone_function_impl(&state, fxn->impl);
 758       nfxn->impl->function = nfxn;
 759    }
 760
 761    ns->info = s->info;
 762    ns->info.name = ralloc_strdup(ns, ns->info.name);
 763    if (ns->info.label)
 764       ns->info.label = ralloc_strdup(ns, ns->info.label);
 765
 766    ns->num_inputs = s->num_inputs;
 767    ns->num_uniforms = s->num_uniforms;
 768    ns->num_outputs = s->num_outputs;
 769    ns->shared_size = s->shared_size;
 770    ns->scratch_size = s->scratch_size;
 771
 772    ns->constant_data_size = s->constant_data_size;
 773    if (s->constant_data_size > 0) {
 774       ns->constant_data = ralloc_size(ns, s->constant_data_size);
 775       memcpy(ns->constant_data, s->constant_data, s->constant_data_size);
 776    }
 777
 778    free_clone_state(&state);
 779
 780    return ns;
 781 }
 782
 783 /** Overwrites dst and replaces its contents with src
 784  *
 785  * Everything ralloc parented to dst and src itself (but not its children)
 786  * will be freed.
 787  *
 788  * This should only be used by test code which needs to swap out shaders with
 789  * a cloned or deserialized version.
 790  */
 791 void
 792 nir_shader_replace(nir_shader *dst, nir_shader *src)
 793 {
 794    /* Delete all of dest's ralloc children */
 795    void *dead_ctx = ralloc_context(NULL);
 796    ralloc_adopt(dead_ctx, dst);
 797    ralloc_free(dead_ctx);
 798
 799    /* Re-parent all of src's ralloc children to dst */
 800    ralloc_adopt(dst, src);
 801
 802    memcpy(dst, src, sizeof(*dst));
 803
 804    /* We have to move all the linked lists over separately because we need the
 805     * pointers in the list elements to point to the lists in dst and not src.
 806     */
 807    exec_list_move_nodes_to(&src->variables, &dst->variables);
 808
 809    /* Now move the functions over.  This takes a tiny bit more work */
 810    exec_list_move_nodes_to(&src->functions, &dst->functions);
 811    nir_foreach_function(function, dst)
 812       function->shader = dst;
 813
 814    ralloc_free(src);
 815 }