src/glsl/nir/nir.c

   1 /*
   2  * Copyright © 2014 Intel Corporation
   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  * Authors:
  24  *    Connor Abbott (cwabbott0@gmail.com)
  25  *
  26  */
  27
  28 #include "nir.h"
  29 #include <assert.h>
  30
  31 nir_shader *
  32 nir_shader_create(void *mem_ctx)
  33 {
  34    nir_shader *shader = ralloc(mem_ctx, nir_shader);
  35
  36    shader->uniforms = _mesa_hash_table_create(shader, _mesa_key_hash_string,
  37                                               _mesa_key_string_equal);
  38    shader->inputs = _mesa_hash_table_create(shader, _mesa_key_hash_string,
  39                                             _mesa_key_string_equal);
  40    shader->outputs = _mesa_hash_table_create(shader, _mesa_key_hash_string,
  41                                              _mesa_key_string_equal);
  42
  43    shader->num_user_structures = 0;
  44    shader->user_structures = NULL;
  45
  46    exec_list_make_empty(&shader->functions);
  47    exec_list_make_empty(&shader->registers);
  48    exec_list_make_empty(&shader->globals);
  49    exec_list_make_empty(&shader->system_values);
  50    shader->reg_alloc = 0;
  51
  52    shader->num_inputs = 0;
  53    shader->num_outputs = 0;
  54    shader->num_uniforms = 0;
  55
  56    return shader;
  57 }
  58
  59 static nir_register *
  60 reg_create(void *mem_ctx, struct exec_list *list)
  61 {
  62    nir_register *reg = ralloc(mem_ctx, nir_register);
  63
  64    reg->uses = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
  65                                 _mesa_key_pointer_equal);
  66    reg->defs = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
  67                                 _mesa_key_pointer_equal);
  68    reg->if_uses = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
  69                                    _mesa_key_pointer_equal);
  70
  71    reg->num_components = 0;
  72    reg->num_array_elems = 0;
  73    reg->is_packed = false;
  74    reg->name = NULL;
  75
  76    exec_list_push_tail(list, &reg->node);
  77
  78    return reg;
  79 }
  80
  81 nir_register *
  82 nir_global_reg_create(nir_shader *shader)
  83 {
  84    nir_register *reg = reg_create(shader, &shader->registers);
  85    reg->index = shader->reg_alloc++;
  86    reg->is_global = true;
  87
  88    return reg;
  89 }
  90
  91 nir_register *
  92 nir_local_reg_create(nir_function_impl *impl)
  93 {
  94    nir_register *reg = reg_create(ralloc_parent(impl), &impl->registers);
  95    reg->index = impl->reg_alloc++;
  96    reg->is_global = false;
  97
  98    return reg;
  99 }
 100
 101 void
 102 nir_reg_remove(nir_register *reg)
 103 {
 104    exec_node_remove(&reg->node);
 105 }
 106
 107 nir_function *
 108 nir_function_create(nir_shader *shader, const char *name)
 109 {
 110    nir_function *func = ralloc(shader, nir_function);
 111
 112    exec_list_push_tail(&shader->functions, &func->node);
 113    exec_list_make_empty(&func->overload_list);
 114    func->name = name;
 115    func->shader = shader;
 116
 117    return func;
 118 }
 119
 120 nir_function_overload *
 121 nir_function_overload_create(nir_function *func)
 122 {
 123    void *mem_ctx = ralloc_parent(func);
 124
 125    nir_function_overload *overload = ralloc(mem_ctx, nir_function_overload);
 126
 127    overload->num_params = 0;
 128    overload->params = NULL;
 129    overload->return_type = glsl_void_type();
 130    overload->impl = NULL;
 131
 132    exec_list_push_tail(&func->overload_list, &overload->node);
 133    overload->function = func;
 134
 135    return overload;
 136 }
 137
 138 void nir_src_copy(nir_src *dest, const nir_src *src, void *mem_ctx)
 139 {
 140    dest->is_ssa = src->is_ssa;
 141    if (src->is_ssa) {
 142       dest->ssa = src->ssa;
 143    } else {
 144       dest->reg.base_offset = src->reg.base_offset;
 145       dest->reg.reg = src->reg.reg;
 146       if (src->reg.indirect) {
 147          dest->reg.indirect = ralloc(mem_ctx, nir_src);
 148          nir_src_copy(dest->reg.indirect, src->reg.indirect, mem_ctx);
 149       } else {
 150          dest->reg.indirect = NULL;
 151       }
 152    }
 153 }
 154
 155 void nir_dest_copy(nir_dest *dest, const nir_dest *src, void *mem_ctx)
 156 {
 157    dest->is_ssa = src->is_ssa;
 158    if (src->is_ssa) {
 159       dest->ssa = src->ssa;
 160    } else {
 161       dest->reg.base_offset = src->reg.base_offset;
 162       dest->reg.reg = src->reg.reg;
 163       if (src->reg.indirect) {
 164          dest->reg.indirect = ralloc(mem_ctx, nir_src);
 165          nir_src_copy(dest->reg.indirect, src->reg.indirect, mem_ctx);
 166       } else {
 167          dest->reg.indirect = NULL;
 168       }
 169    }
 170 }
 171
 172 void
 173 nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src, void *mem_ctx)
 174 {
 175    nir_src_copy(&dest->src, &src->src, mem_ctx);
 176    dest->abs = src->abs;
 177    dest->negate = src->negate;
 178    for (unsigned i = 0; i < 4; i++)
 179       dest->swizzle[i] = src->swizzle[i];
 180 }
 181
 182 void
 183 nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src, void *mem_ctx)
 184 {
 185    nir_dest_copy(&dest->dest, &src->dest, mem_ctx);
 186    dest->write_mask = src->write_mask;
 187    dest->saturate = src->saturate;
 188 }
 189
 190 static inline void
 191 block_add_pred(nir_block *block, nir_block *pred)
 192 {
 193    _mesa_set_add(block->predecessors, pred);
 194 }
 195
 196 static void
 197 cf_init(nir_cf_node *node, nir_cf_node_type type)
 198 {
 199    exec_node_init(&node->node);
 200    node->parent = NULL;
 201    node->type = type;
 202 }
 203
 204 static void
 205 link_blocks(nir_block *pred, nir_block *succ1, nir_block *succ2)
 206 {
 207    pred->successors[0] = succ1;
 208    block_add_pred(succ1, pred);
 209
 210    pred->successors[1] = succ2;
 211    if (succ2 != NULL)
 212       block_add_pred(succ2, pred);
 213 }
 214
 215 static void
 216 unlink_blocks(nir_block *pred, nir_block *succ)
 217 {
 218    if (pred->successors[0] == succ) {
 219       pred->successors[0] = pred->successors[1];
 220       pred->successors[1] = NULL;
 221    } else {
 222       assert(pred->successors[1] == succ);
 223       pred->successors[1] = NULL;
 224    }
 225
 226    struct set_entry *entry = _mesa_set_search(succ->predecessors, pred);
 227
 228    assert(entry);
 229
 230    _mesa_set_remove(succ->predecessors, entry);
 231 }
 232
 233 static void
 234 unlink_block_successors(nir_block *block)
 235 {
 236    if (block->successors[0] != NULL)
 237       unlink_blocks(block, block->successors[0]);
 238    if (block->successors[1] != NULL)
 239       unlink_blocks(block, block->successors[1]);
 240 }
 241
 242
 243 nir_function_impl *
 244 nir_function_impl_create(nir_function_overload *overload)
 245 {
 246    assert(overload->impl == NULL);
 247
 248    void *mem_ctx = ralloc_parent(overload);
 249
 250    nir_function_impl *impl = ralloc(mem_ctx, nir_function_impl);
 251
 252    overload->impl = impl;
 253    impl->overload = overload;
 254
 255    cf_init(&impl->cf_node, nir_cf_node_function);
 256
 257    exec_list_make_empty(&impl->body);
 258    exec_list_make_empty(&impl->registers);
 259    exec_list_make_empty(&impl->locals);
 260    impl->num_params = 0;
 261    impl->params = NULL;
 262    impl->return_var = NULL;
 263    impl->reg_alloc = 0;
 264    impl->ssa_alloc = 0;
 265    impl->valid_metadata = nir_metadata_none;
 266
 267    /* create start & end blocks */
 268    nir_block *start_block = nir_block_create(mem_ctx);
 269    nir_block *end_block = nir_block_create(mem_ctx);
 270    start_block->cf_node.parent = &impl->cf_node;
 271    end_block->cf_node.parent = &impl->cf_node;
 272    impl->start_block = start_block;
 273    impl->end_block = end_block;
 274
 275    exec_list_push_tail(&impl->body, &start_block->cf_node.node);
 276
 277    start_block->successors[0] = end_block;
 278    block_add_pred(end_block, start_block);
 279
 280    return impl;
 281 }
 282
 283 nir_block *
 284 nir_block_create(void *mem_ctx)
 285 {
 286    nir_block *block = ralloc(mem_ctx, nir_block);
 287
 288    cf_init(&block->cf_node, nir_cf_node_block);
 289
 290    block->successors[0] = block->successors[1] = NULL;
 291    block->predecessors = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
 292                                           _mesa_key_pointer_equal);
 293    block->imm_dom = NULL;
 294    block->dom_frontier = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
 295                                           _mesa_key_pointer_equal);
 296
 297    exec_list_make_empty(&block->instr_list);
 298
 299    return block;
 300 }
 301
 302 static inline void
 303 src_init(nir_src *src)
 304 {
 305    src->is_ssa = false;
 306    src->reg.reg = NULL;
 307    src->reg.indirect = NULL;
 308    src->reg.base_offset = 0;
 309 }
 310
 311 nir_if *
 312 nir_if_create(void *mem_ctx)
 313 {
 314    nir_if *if_stmt = ralloc(mem_ctx, nir_if);
 315
 316    cf_init(&if_stmt->cf_node, nir_cf_node_if);
 317    src_init(&if_stmt->condition);
 318
 319    nir_block *then = nir_block_create(mem_ctx);
 320    exec_list_make_empty(&if_stmt->then_list);
 321    exec_list_push_tail(&if_stmt->then_list, &then->cf_node.node);
 322    then->cf_node.parent = &if_stmt->cf_node;
 323
 324    nir_block *else_stmt = nir_block_create(mem_ctx);
 325    exec_list_make_empty(&if_stmt->else_list);
 326    exec_list_push_tail(&if_stmt->else_list, &else_stmt->cf_node.node);
 327    else_stmt->cf_node.parent = &if_stmt->cf_node;
 328
 329    return if_stmt;
 330 }
 331
 332 nir_loop *
 333 nir_loop_create(void *mem_ctx)
 334 {
 335    nir_loop *loop = ralloc(mem_ctx, nir_loop);
 336
 337    cf_init(&loop->cf_node, nir_cf_node_loop);
 338
 339    nir_block *body = nir_block_create(mem_ctx);
 340    exec_list_make_empty(&loop->body);
 341    exec_list_push_tail(&loop->body, &body->cf_node.node);
 342    body->cf_node.parent = &loop->cf_node;
 343
 344    body->successors[0] = body;
 345    block_add_pred(body, body);
 346
 347    return loop;
 348 }
 349
 350 static void
 351 instr_init(nir_instr *instr, nir_instr_type type)
 352 {
 353    instr->type = type;
 354    instr->block = NULL;
 355    exec_node_init(&instr->node);
 356 }
 357
 358 static void
 359 dest_init(nir_dest *dest)
 360 {
 361    dest->is_ssa = false;
 362    dest->reg.reg = NULL;
 363    dest->reg.indirect = NULL;
 364    dest->reg.base_offset = 0;
 365 }
 366
 367 static void
 368 alu_dest_init(nir_alu_dest *dest)
 369 {
 370    dest_init(&dest->dest);
 371    dest->saturate = false;
 372    dest->write_mask = 0xf;
 373 }
 374
 375 static void
 376 alu_src_init(nir_alu_src *src)
 377 {
 378    src_init(&src->src);
 379    src->abs = src->negate = false;
 380    src->swizzle[0] = 0;
 381    src->swizzle[1] = 1;
 382    src->swizzle[2] = 2;
 383    src->swizzle[3] = 3;
 384 }
 385
 386 nir_alu_instr *
 387 nir_alu_instr_create(void *mem_ctx, nir_op op)
 388 {
 389    unsigned num_srcs = nir_op_infos[op].num_inputs;
 390    nir_alu_instr *instr =
 391       ralloc_size(mem_ctx,
 392                   sizeof(nir_alu_instr) + num_srcs * sizeof(nir_alu_src));
 393
 394    instr_init(&instr->instr, nir_instr_type_alu);
 395    instr->op = op;
 396    alu_dest_init(&instr->dest);
 397    for (unsigned i = 0; i < num_srcs; i++)
 398       alu_src_init(&instr->src[i]);
 399
 400    return instr;
 401 }
 402
 403 nir_jump_instr *
 404 nir_jump_instr_create(void *mem_ctx, nir_jump_type type)
 405 {
 406    nir_jump_instr *instr = ralloc(mem_ctx, nir_jump_instr);
 407    instr_init(&instr->instr, nir_instr_type_jump);
 408    instr->type = type;
 409    return instr;
 410 }
 411
 412 nir_load_const_instr *
 413 nir_load_const_instr_create(void *mem_ctx, unsigned num_components)
 414 {
 415    nir_load_const_instr *instr = ralloc(mem_ctx, nir_load_const_instr);
 416    instr_init(&instr->instr, nir_instr_type_load_const);
 417
 418    nir_ssa_def_init(&instr->instr, &instr->def, num_components, NULL);
 419
 420    return instr;
 421 }
 422
 423 nir_intrinsic_instr *
 424 nir_intrinsic_instr_create(void *mem_ctx, nir_intrinsic_op op)
 425 {
 426    unsigned num_srcs = nir_intrinsic_infos[op].num_srcs;
 427    nir_intrinsic_instr *instr =
 428       ralloc_size(mem_ctx,
 429                   sizeof(nir_intrinsic_instr) + num_srcs * sizeof(nir_src));
 430
 431    instr_init(&instr->instr, nir_instr_type_intrinsic);
 432    instr->intrinsic = op;
 433
 434    if (nir_intrinsic_infos[op].has_dest)
 435       dest_init(&instr->dest);
 436
 437    for (unsigned i = 0; i < num_srcs; i++)
 438       src_init(&instr->src[i]);
 439
 440    return instr;
 441 }
 442
 443 nir_call_instr *
 444 nir_call_instr_create(void *mem_ctx, nir_function_overload *callee)
 445 {
 446    nir_call_instr *instr = ralloc(mem_ctx, nir_call_instr);
 447    instr_init(&instr->instr, nir_instr_type_call);
 448
 449    instr->callee = callee;
 450    instr->num_params = callee->num_params;
 451    instr->params = ralloc_array(mem_ctx, nir_deref_var *, instr->num_params);
 452    instr->return_deref = NULL;
 453
 454    return instr;
 455 }
 456
 457 nir_tex_instr *
 458 nir_tex_instr_create(void *mem_ctx, unsigned num_srcs)
 459 {
 460    nir_tex_instr *instr = ralloc(mem_ctx, nir_tex_instr);
 461    instr_init(&instr->instr, nir_instr_type_tex);
 462
 463    dest_init(&instr->dest);
 464
 465    instr->num_srcs = num_srcs;
 466    instr->src = ralloc_array(mem_ctx, nir_tex_src, num_srcs);
 467    for (unsigned i = 0; i < num_srcs; i++)
 468       src_init(&instr->src[i].src);
 469
 470    instr->sampler_index = 0;
 471    instr->sampler_array_size = 0;
 472    instr->sampler = NULL;
 473
 474    return instr;
 475 }
 476
 477 nir_phi_instr *
 478 nir_phi_instr_create(void *mem_ctx)
 479 {
 480    nir_phi_instr *instr = ralloc(mem_ctx, nir_phi_instr);
 481    instr_init(&instr->instr, nir_instr_type_phi);
 482
 483    dest_init(&instr->dest);
 484    exec_list_make_empty(&instr->srcs);
 485    return instr;
 486 }
 487
 488 nir_parallel_copy_instr *
 489 nir_parallel_copy_instr_create(void *mem_ctx)
 490 {
 491    nir_parallel_copy_instr *instr = ralloc(mem_ctx, nir_parallel_copy_instr);
 492    instr_init(&instr->instr, nir_instr_type_parallel_copy);
 493
 494    exec_list_make_empty(&instr->entries);
 495
 496    return instr;
 497 }
 498
 499 nir_ssa_undef_instr *
 500 nir_ssa_undef_instr_create(void *mem_ctx, unsigned num_components)
 501 {
 502    nir_ssa_undef_instr *instr = ralloc(mem_ctx, nir_ssa_undef_instr);
 503    instr_init(&instr->instr, nir_instr_type_ssa_undef);
 504
 505    nir_ssa_def_init(&instr->instr, &instr->def, num_components, NULL);
 506
 507    return instr;
 508 }
 509
 510 nir_deref_var *
 511 nir_deref_var_create(void *mem_ctx, nir_variable *var)
 512 {
 513    nir_deref_var *deref = ralloc(mem_ctx, nir_deref_var);
 514    deref->deref.deref_type = nir_deref_type_var;
 515    deref->deref.child = NULL;
 516    deref->deref.type = var->type;
 517    deref->var = var;
 518    return deref;
 519 }
 520
 521 nir_deref_array *
 522 nir_deref_array_create(void *mem_ctx)
 523 {
 524    nir_deref_array *deref = ralloc(mem_ctx, nir_deref_array);
 525    deref->deref.deref_type = nir_deref_type_array;
 526    deref->deref.child = NULL;
 527    deref->deref_array_type = nir_deref_array_type_direct;
 528    src_init(&deref->indirect);
 529    deref->base_offset = 0;
 530    return deref;
 531 }
 532
 533 nir_deref_struct *
 534 nir_deref_struct_create(void *mem_ctx, unsigned field_index)
 535 {
 536    nir_deref_struct *deref = ralloc(mem_ctx, nir_deref_struct);
 537    deref->deref.deref_type = nir_deref_type_struct;
 538    deref->deref.child = NULL;
 539    deref->index = field_index;
 540    return deref;
 541 }
 542
 543 static nir_deref_var *
 544 copy_deref_var(void *mem_ctx, nir_deref_var *deref)
 545 {
 546    nir_deref_var *ret = nir_deref_var_create(mem_ctx, deref->var);
 547    ret->deref.type = deref->deref.type;
 548    if (deref->deref.child)
 549       ret->deref.child = nir_copy_deref(mem_ctx, deref->deref.child);
 550    return ret;
 551 }
 552
 553 static nir_deref_array *
 554 copy_deref_array(void *mem_ctx, nir_deref_array *deref)
 555 {
 556    nir_deref_array *ret = nir_deref_array_create(mem_ctx);
 557    ret->base_offset = deref->base_offset;
 558    ret->deref_array_type = deref->deref_array_type;
 559    if (deref->deref_array_type == nir_deref_array_type_indirect) {
 560       nir_src_copy(&ret->indirect, &deref->indirect, mem_ctx);
 561    }
 562    ret->deref.type = deref->deref.type;
 563    if (deref->deref.child)
 564       ret->deref.child = nir_copy_deref(mem_ctx, deref->deref.child);
 565    return ret;
 566 }
 567
 568 static nir_deref_struct *
 569 copy_deref_struct(void *mem_ctx, nir_deref_struct *deref)
 570 {
 571    nir_deref_struct *ret = nir_deref_struct_create(mem_ctx, deref->index);
 572    ret->deref.type = deref->deref.type;
 573    if (deref->deref.child)
 574       ret->deref.child = nir_copy_deref(mem_ctx, deref->deref.child);
 575    return ret;
 576 }
 577
 578 nir_deref *
 579 nir_copy_deref(void *mem_ctx, nir_deref *deref)
 580 {
 581    switch (deref->deref_type) {
 582    case nir_deref_type_var:
 583       return &copy_deref_var(mem_ctx, nir_deref_as_var(deref))->deref;
 584    case nir_deref_type_array:
 585       return &copy_deref_array(mem_ctx, nir_deref_as_array(deref))->deref;
 586    case nir_deref_type_struct:
 587       return &copy_deref_struct(mem_ctx, nir_deref_as_struct(deref))->deref;
 588    default:
 589       unreachable("Invalid dereference type");
 590    }
 591
 592    return NULL;
 593 }
 594
 595
 596 /**
 597  * \name Control flow modification
 598  *
 599  * These functions modify the control flow tree while keeping the control flow
 600  * graph up-to-date. The invariants respected are:
 601  * 1. Each then statement, else statement, or loop body must have at least one
 602  *    control flow node.
 603  * 2. Each if-statement and loop must have one basic block before it and one
 604  *    after.
 605  * 3. Two basic blocks cannot be directly next to each other.
 606  * 4. If a basic block has a jump instruction, there must be only one and it
 607  *    must be at the end of the block.
 608  * 5. The CFG must always be connected - this means that we must insert a fake
 609  *    CFG edge for loops with no break statement.
 610  *
 611  * The purpose of the second one is so that we have places to insert code during
 612  * GCM, as well as eliminating the possibility of critical edges.
 613  */
 614 /*@{*/
 615
 616 static void
 617 link_non_block_to_block(nir_cf_node *node, nir_block *block)
 618 {
 619    if (node->type == nir_cf_node_if) {
 620       /*
 621        * We're trying to link an if to a block after it; this just means linking
 622        * the last block of the then and else branches.
 623        */
 624
 625       nir_if *if_stmt = nir_cf_node_as_if(node);
 626
 627       nir_cf_node *last_then = nir_if_last_then_node(if_stmt);
 628       assert(last_then->type == nir_cf_node_block);
 629       nir_block *last_then_block = nir_cf_node_as_block(last_then);
 630
 631       nir_cf_node *last_else = nir_if_last_else_node(if_stmt);
 632       assert(last_else->type == nir_cf_node_block);
 633       nir_block *last_else_block = nir_cf_node_as_block(last_else);
 634
 635       if (exec_list_is_empty(&last_then_block->instr_list) ||
 636           nir_block_last_instr(last_then_block)->type != nir_instr_type_jump) {
 637          unlink_block_successors(last_then_block);
 638          link_blocks(last_then_block, block, NULL);
 639       }
 640
 641       if (exec_list_is_empty(&last_else_block->instr_list) ||
 642           nir_block_last_instr(last_else_block)->type != nir_instr_type_jump) {
 643          unlink_block_successors(last_else_block);
 644          link_blocks(last_else_block, block, NULL);
 645       }
 646    } else {
 647       assert(node->type == nir_cf_node_loop);
 648
 649       /*
 650        * We can only get to this codepath if we're inserting a new loop, or
 651        * at least a loop with no break statements; we can't insert break
 652        * statements into a loop when we haven't inserted it into the CFG
 653        * because we wouldn't know which block comes after the loop
 654        * and therefore, which block should be the successor of the block with
 655        * the break). Therefore, we need to insert a fake edge (see invariant
 656        * #5).
 657        */
 658
 659       nir_loop *loop = nir_cf_node_as_loop(node);
 660
 661       nir_cf_node *last = nir_loop_last_cf_node(loop);
 662       assert(last->type == nir_cf_node_block);
 663       nir_block *last_block =  nir_cf_node_as_block(last);
 664
 665       last_block->successors[1] = block;
 666       block_add_pred(block, last_block);
 667    }
 668 }
 669
 670 static void
 671 link_block_to_non_block(nir_block *block, nir_cf_node *node)
 672 {
 673    if (node->type == nir_cf_node_if) {
 674       /*
 675        * We're trying to link a block to an if after it; this just means linking
 676        * the block to the first block of the then and else branches.
 677        */
 678
 679       nir_if *if_stmt = nir_cf_node_as_if(node);
 680
 681       nir_cf_node *first_then = nir_if_first_then_node(if_stmt);
 682       assert(first_then->type == nir_cf_node_block);
 683       nir_block *first_then_block = nir_cf_node_as_block(first_then);
 684
 685       nir_cf_node *first_else = nir_if_first_else_node(if_stmt);
 686       assert(first_else->type == nir_cf_node_block);
 687       nir_block *first_else_block = nir_cf_node_as_block(first_else);
 688
 689       unlink_block_successors(block);
 690       link_blocks(block, first_then_block, first_else_block);
 691    } else {
 692       /*
 693        * For similar reasons as the corresponding case in
 694        * link_non_block_to_block(), don't worry about if the loop header has
 695        * any predecessors that need to be unlinked.
 696        */
 697
 698       assert(node->type == nir_cf_node_loop);
 699
 700       nir_loop *loop = nir_cf_node_as_loop(node);
 701
 702       nir_cf_node *loop_header = nir_loop_first_cf_node(loop);
 703       assert(loop_header->type == nir_cf_node_block);
 704       nir_block *loop_header_block = nir_cf_node_as_block(loop_header);
 705
 706       unlink_block_successors(block);
 707       link_blocks(block, loop_header_block, NULL);
 708    }
 709
 710 }
 711
 712 /**
 713  * Takes a basic block and inserts a new empty basic block before it, making its
 714  * predecessors point to the new block. This essentially splits the block into
 715  * an empty header and a body so that another non-block CF node can be inserted
 716  * between the two. Note that this does *not* link the two basic blocks, so
 717  * some kind of cleanup *must* be performed after this call.
 718  */
 719
 720 static nir_block *
 721 split_block_beginning(nir_block *block)
 722 {
 723    nir_block *new_block = nir_block_create(ralloc_parent(block));
 724    new_block->cf_node.parent = block->cf_node.parent;
 725    exec_node_insert_node_before(&block->cf_node.node, &new_block->cf_node.node);
 726
 727    struct set_entry *entry;
 728    set_foreach(block->predecessors, entry) {
 729       nir_block *pred = (nir_block *) entry->key;
 730
 731       unlink_blocks(pred, block);
 732       link_blocks(pred, new_block, NULL);
 733    }
 734
 735    return new_block;
 736 }
 737
 738 static void
 739 rewrite_phi_preds(nir_block *block, nir_block *old_pred, nir_block *new_pred)
 740 {
 741    nir_foreach_instr_safe(block, instr) {
 742       if (instr->type != nir_instr_type_phi)
 743          break;
 744
 745       nir_phi_instr *phi = nir_instr_as_phi(instr);
 746       nir_foreach_phi_src(phi, src) {
 747          if (src->pred == old_pred) {
 748             src->pred = new_pred;
 749             break;
 750          }
 751       }
 752    }
 753 }
 754
 755 /**
 756  * Moves the successors of source to the successors of dest, leaving both
 757  * successors of source NULL.
 758  */
 759
 760 static void
 761 move_successors(nir_block *source, nir_block *dest)
 762 {
 763    nir_block *succ1 = source->successors[0];
 764    nir_block *succ2 = source->successors[1];
 765
 766    if (succ1) {
 767       unlink_blocks(source, succ1);
 768       rewrite_phi_preds(succ1, source, dest);
 769    }
 770
 771    if (succ2) {
 772       unlink_blocks(source, succ2);
 773       rewrite_phi_preds(succ2, source, dest);
 774    }
 775
 776    unlink_block_successors(dest);
 777    link_blocks(dest, succ1, succ2);
 778 }
 779
 780 static nir_block *
 781 split_block_end(nir_block *block)
 782 {
 783    nir_block *new_block = nir_block_create(ralloc_parent(block));
 784    new_block->cf_node.parent = block->cf_node.parent;
 785    exec_node_insert_after(&block->cf_node.node, &new_block->cf_node.node);
 786
 787    move_successors(block, new_block);
 788
 789    return new_block;
 790 }
 791
 792 /**
 793  * Inserts a non-basic block between two basic blocks and links them together.
 794  */
 795
 796 static void
 797 insert_non_block(nir_block *before, nir_cf_node *node, nir_block *after)
 798 {
 799    node->parent = before->cf_node.parent;
 800    exec_node_insert_after(&before->cf_node.node, &node->node);
 801    link_block_to_non_block(before, node);
 802    link_non_block_to_block(node, after);
 803 }
 804
 805 /**
 806  * Inserts a non-basic block before a basic block.
 807  */
 808
 809 static void
 810 insert_non_block_before_block(nir_cf_node *node, nir_block *block)
 811 {
 812    /* split off the beginning of block into new_block */
 813    nir_block *new_block = split_block_beginning(block);
 814
 815    /* insert our node in between new_block and block */
 816    insert_non_block(new_block, node, block);
 817 }
 818
 819 static void
 820 insert_non_block_after_block(nir_block *block, nir_cf_node *node)
 821 {
 822    /* split off the end of block into new_block */
 823    nir_block *new_block = split_block_end(block);
 824
 825    /* insert our node in between block and new_block */
 826    insert_non_block(block, node, new_block);
 827 }
 828
 829 /* walk up the control flow tree to find the innermost enclosed loop */
 830 static nir_loop *
 831 nearest_loop(nir_cf_node *node)
 832 {
 833    while (node->type != nir_cf_node_loop) {
 834       node = node->parent;
 835    }
 836
 837    return nir_cf_node_as_loop(node);
 838 }
 839
 840 nir_function_impl *
 841 nir_cf_node_get_function(nir_cf_node *node)
 842 {
 843    while (node->type != nir_cf_node_function) {
 844       node = node->parent;
 845    }
 846
 847    return nir_cf_node_as_function(node);
 848 }
 849
 850 /*
 851  * update the CFG after a jump instruction has been added to the end of a block
 852  */
 853
 854 static void
 855 handle_jump(nir_block *block)
 856 {
 857    nir_instr *instr = nir_block_last_instr(block);
 858    nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
 859
 860    unlink_block_successors(block);
 861
 862    nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
 863    nir_metadata_preserve(impl, nir_metadata_none);
 864
 865    if (jump_instr->type == nir_jump_break ||
 866        jump_instr->type == nir_jump_continue) {
 867       nir_loop *loop = nearest_loop(&block->cf_node);
 868
 869       if (jump_instr->type == nir_jump_continue) {
 870          nir_cf_node *first_node = nir_loop_first_cf_node(loop);
 871          assert(first_node->type == nir_cf_node_block);
 872          nir_block *first_block = nir_cf_node_as_block(first_node);
 873          link_blocks(block, first_block, NULL);
 874       } else {
 875          nir_cf_node *after = nir_cf_node_next(&loop->cf_node);
 876          assert(after->type == nir_cf_node_block);
 877          nir_block *after_block = nir_cf_node_as_block(after);
 878          link_blocks(block, after_block, NULL);
 879
 880          /* If we inserted a fake link, remove it */
 881          nir_cf_node *last = nir_loop_last_cf_node(loop);
 882          assert(last->type == nir_cf_node_block);
 883          nir_block *last_block =  nir_cf_node_as_block(last);
 884          if (last_block->successors[1] != NULL)
 885             unlink_blocks(last_block, after_block);
 886       }
 887    } else {
 888       assert(jump_instr->type == nir_jump_return);
 889       link_blocks(block, impl->end_block, NULL);
 890    }
 891 }
 892
 893 static void
 894 handle_remove_jump(nir_block *block, nir_jump_type type)
 895 {
 896    unlink_block_successors(block);
 897
 898    if (exec_node_is_tail_sentinel(block->cf_node.node.next)) {
 899       nir_cf_node *parent = block->cf_node.parent;
 900       if (parent->type == nir_cf_node_if) {
 901          nir_cf_node *next = nir_cf_node_next(parent);
 902          assert(next->type == nir_cf_node_block);
 903          nir_block *next_block = nir_cf_node_as_block(next);
 904
 905          link_blocks(block, next_block, NULL);
 906       } else {
 907          assert(parent->type == nir_cf_node_loop);
 908          nir_loop *loop = nir_cf_node_as_loop(parent);
 909
 910          nir_cf_node *head = nir_loop_first_cf_node(loop);
 911          assert(head->type == nir_cf_node_block);
 912          nir_block *head_block = nir_cf_node_as_block(head);
 913
 914          link_blocks(block, head_block, NULL);
 915       }
 916    } else {
 917       nir_cf_node *next = nir_cf_node_next(&block->cf_node);
 918       if (next->type == nir_cf_node_if) {
 919          nir_if *next_if = nir_cf_node_as_if(next);
 920
 921          nir_cf_node *first_then = nir_if_first_then_node(next_if);
 922          assert(first_then->type == nir_cf_node_block);
 923          nir_block *first_then_block = nir_cf_node_as_block(first_then);
 924
 925          nir_cf_node *first_else = nir_if_first_else_node(next_if);
 926          assert(first_else->type == nir_cf_node_block);
 927          nir_block *first_else_block = nir_cf_node_as_block(first_else);
 928
 929          link_blocks(block, first_then_block, first_else_block);
 930       } else {
 931          assert(next->type == nir_cf_node_loop);
 932          nir_loop *next_loop = nir_cf_node_as_loop(next);
 933
 934          nir_cf_node *first = nir_loop_first_cf_node(next_loop);
 935          assert(first->type == nir_cf_node_block);
 936          nir_block *first_block = nir_cf_node_as_block(first);
 937
 938          link_blocks(block, first_block, NULL);
 939       }
 940    }
 941
 942    if (type == nir_jump_break) {
 943       nir_loop *loop = nearest_loop(&block->cf_node);
 944
 945       nir_cf_node *next = nir_cf_node_next(&loop->cf_node);
 946       assert(next->type == nir_cf_node_block);
 947       nir_block *next_block = nir_cf_node_as_block(next);
 948
 949       if (next_block->predecessors->entries == 0) {
 950          /* insert fake link */
 951          nir_cf_node *last = nir_loop_last_cf_node(loop);
 952          assert(last->type == nir_cf_node_block);
 953          nir_block *last_block = nir_cf_node_as_block(last);
 954
 955          last_block->successors[1] = next_block;
 956          block_add_pred(next_block, last_block);
 957       }
 958    }
 959
 960    nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
 961    nir_metadata_preserve(impl, nir_metadata_none);
 962 }
 963
 964 /**
 965  * Inserts a basic block before another by merging the instructions.
 966  *
 967  * @param block the target of the insertion
 968  * @param before the block to be inserted - must not have been inserted before
 969  * @param has_jump whether \before has a jump instruction at the end
 970  */
 971
 972 static void
 973 insert_block_before_block(nir_block *block, nir_block *before, bool has_jump)
 974 {
 975    assert(!has_jump || exec_list_is_empty(&block->instr_list));
 976
 977    foreach_list_typed(nir_instr, instr, node, &before->instr_list) {
 978       instr->block = block;
 979    }
 980
 981    exec_list_prepend(&block->instr_list, &before->instr_list);
 982
 983    if (has_jump)
 984       handle_jump(block);
 985 }
 986
 987 /**
 988  * Inserts a basic block after another by merging the instructions.
 989  *
 990  * @param block the target of the insertion
 991  * @param after the block to be inserted - must not have been inserted before
 992  * @param has_jump whether \after has a jump instruction at the end
 993  */
 994
 995 static void
 996 insert_block_after_block(nir_block *block, nir_block *after, bool has_jump)
 997 {
 998    foreach_list_typed(nir_instr, instr, node, &after->instr_list) {
 999       instr->block = block;
1000    }
1001
1002    exec_list_append(&block->instr_list, &after->instr_list);
1003
1004    if (has_jump)
1005       handle_jump(block);
1006 }
1007
1008 static void
1009 update_if_uses(nir_cf_node *node)
1010 {
1011    if (node->type != nir_cf_node_if)
1012       return;
1013
1014    nir_if *if_stmt = nir_cf_node_as_if(node);
1015
1016    struct set *if_uses_set = if_stmt->condition.is_ssa ?
1017                              if_stmt->condition.ssa->if_uses :
1018                              if_stmt->condition.reg.reg->uses;
1019
1020    _mesa_set_add(if_uses_set, if_stmt);
1021 }
1022
1023 void
1024 nir_cf_node_insert_after(nir_cf_node *node, nir_cf_node *after)
1025 {
1026    update_if_uses(after);
1027
1028    if (after->type == nir_cf_node_block) {
1029       /*
1030        * either node or the one after it must be a basic block, by invariant #2;
1031        * in either case, just merge the blocks together.
1032        */
1033       nir_block *after_block = nir_cf_node_as_block(after);
1034
1035       bool has_jump = !exec_list_is_empty(&after_block->instr_list) &&
1036          nir_block_last_instr(after_block)->type == nir_instr_type_jump;
1037
1038       if (node->type == nir_cf_node_block) {
1039          insert_block_after_block(nir_cf_node_as_block(node), after_block,
1040                                   has_jump);
1041       } else {
1042          nir_cf_node *next = nir_cf_node_next(node);
1043          assert(next->type == nir_cf_node_block);
1044          nir_block *next_block = nir_cf_node_as_block(next);
1045
1046          insert_block_before_block(next_block, after_block, has_jump);
1047       }
1048    } else {
1049       if (node->type == nir_cf_node_block) {
1050          insert_non_block_after_block(nir_cf_node_as_block(node), after);
1051       } else {
1052          /*
1053           * We have to insert a non-basic block after a non-basic block. Since
1054           * every non-basic block has a basic block after it, this is equivalent
1055           * to inserting a non-basic block before a basic block.
1056           */
1057
1058          nir_cf_node *next = nir_cf_node_next(node);
1059          assert(next->type == nir_cf_node_block);
1060          nir_block *next_block = nir_cf_node_as_block(next);
1061
1062          insert_non_block_before_block(after, next_block);
1063       }
1064    }
1065
1066    nir_function_impl *impl = nir_cf_node_get_function(node);
1067    nir_metadata_preserve(impl, nir_metadata_none);
1068 }
1069
1070 void
1071 nir_cf_node_insert_before(nir_cf_node *node, nir_cf_node *before)
1072 {
1073    update_if_uses(before);
1074
1075    if (before->type == nir_cf_node_block) {
1076       nir_block *before_block = nir_cf_node_as_block(before);
1077
1078       bool has_jump = !exec_list_is_empty(&before_block->instr_list) &&
1079          nir_block_last_instr(before_block)->type == nir_instr_type_jump;
1080
1081       if (node->type == nir_cf_node_block) {
1082          insert_block_before_block(nir_cf_node_as_block(node), before_block,
1083                                    has_jump);
1084       } else {
1085          nir_cf_node *prev = nir_cf_node_prev(node);
1086          assert(prev->type == nir_cf_node_block);
1087          nir_block *prev_block = nir_cf_node_as_block(prev);
1088
1089          insert_block_after_block(prev_block, before_block, has_jump);
1090       }
1091    } else {
1092       if (node->type == nir_cf_node_block) {
1093          insert_non_block_before_block(before, nir_cf_node_as_block(node));
1094       } else {
1095          /*
1096           * We have to insert a non-basic block before a non-basic block. This
1097           * is equivalent to inserting a non-basic block after a basic block.
1098           */
1099
1100          nir_cf_node *prev_node = nir_cf_node_prev(node);
1101          assert(prev_node->type == nir_cf_node_block);
1102          nir_block *prev_block = nir_cf_node_as_block(prev_node);
1103
1104          insert_non_block_after_block(prev_block, before);
1105       }
1106    }
1107
1108    nir_function_impl *impl = nir_cf_node_get_function(node);
1109    nir_metadata_preserve(impl, nir_metadata_none);
1110 }
1111
1112 void
1113 nir_cf_node_insert_begin(struct exec_list *list, nir_cf_node *node)
1114 {
1115    nir_cf_node *begin = exec_node_data(nir_cf_node, list->head, node);
1116    nir_cf_node_insert_before(begin, node);
1117 }
1118
1119 void
1120 nir_cf_node_insert_end(struct exec_list *list, nir_cf_node *node)
1121 {
1122    nir_cf_node *end = exec_node_data(nir_cf_node, list->tail_pred, node);
1123    nir_cf_node_insert_after(end, node);
1124 }
1125
1126 /**
1127  * Stitch two basic blocks together into one. The aggregate must have the same
1128  * predecessors as the first and the same successors as the second.
1129  */
1130
1131 static void
1132 stitch_blocks(nir_block *before, nir_block *after)
1133 {
1134    /*
1135     * We move after into before, so we have to deal with up to 2 successors vs.
1136     * possibly a large number of predecessors.
1137     *
1138     * TODO: special case when before is empty and after isn't?
1139     */
1140
1141    move_successors(after, before);
1142
1143    foreach_list_typed(nir_instr, instr, node, &after->instr_list) {
1144       instr->block = before;
1145    }
1146
1147    exec_list_append(&before->instr_list, &after->instr_list);
1148    exec_node_remove(&after->cf_node.node);
1149 }
1150
1151 void
1152 nir_cf_node_remove(nir_cf_node *node)
1153 {
1154    nir_function_impl *impl = nir_cf_node_get_function(node);
1155    nir_metadata_preserve(impl, nir_metadata_none);
1156
1157    if (node->type == nir_cf_node_block) {
1158       /*
1159        * Basic blocks can't really be removed by themselves, since they act as
1160        * padding between the non-basic blocks. So all we do here is empty the
1161        * block of instructions.
1162        *
1163        * TODO: could we assert here?
1164        */
1165       exec_list_make_empty(&nir_cf_node_as_block(node)->instr_list);
1166    } else {
1167       nir_cf_node *before = nir_cf_node_prev(node);
1168       assert(before->type == nir_cf_node_block);
1169       nir_block *before_block = nir_cf_node_as_block(before);
1170
1171       nir_cf_node *after = nir_cf_node_next(node);
1172       assert(after->type == nir_cf_node_block);
1173       nir_block *after_block = nir_cf_node_as_block(after);
1174
1175       exec_node_remove(&node->node);
1176       stitch_blocks(before_block, after_block);
1177    }
1178 }
1179
1180 static bool
1181 add_use_cb(nir_src *src, void *state)
1182 {
1183    nir_instr *instr = (nir_instr *) state;
1184
1185    struct set *uses_set = src->is_ssa ? src->ssa->uses : src->reg.reg->uses;
1186
1187    _mesa_set_add(uses_set, instr);
1188
1189    return true;
1190 }
1191
1192 static bool
1193 add_ssa_def_cb(nir_ssa_def *def, void *state)
1194 {
1195    nir_instr *instr = (nir_instr *) state;
1196
1197    if (instr->block && def->index == UINT_MAX) {
1198       nir_function_impl *impl =
1199          nir_cf_node_get_function(&instr->block->cf_node);
1200
1201       def->index = impl->ssa_alloc++;
1202    }
1203
1204    return true;
1205 }
1206
1207 static bool
1208 add_reg_def_cb(nir_dest *dest, void *state)
1209 {
1210    nir_instr *instr = (nir_instr *) state;
1211
1212    if (!dest->is_ssa)
1213       _mesa_set_add(dest->reg.reg->defs, instr);
1214
1215    return true;
1216 }
1217
1218 static void
1219 add_defs_uses(nir_instr *instr)
1220 {
1221    nir_foreach_src(instr, add_use_cb, instr);
1222    nir_foreach_dest(instr, add_reg_def_cb, instr);
1223    nir_foreach_ssa_def(instr, add_ssa_def_cb, instr);
1224 }
1225
1226 void
1227 nir_instr_insert_before(nir_instr *instr, nir_instr *before)
1228 {
1229    assert(before->type != nir_instr_type_jump);
1230    before->block = instr->block;
1231    add_defs_uses(before);
1232    exec_node_insert_node_before(&instr->node, &before->node);
1233 }
1234
1235 void
1236 nir_instr_insert_after(nir_instr *instr, nir_instr *after)
1237 {
1238    if (after->type == nir_instr_type_jump) {
1239       assert(instr == nir_block_last_instr(instr->block));
1240       assert(instr->type != nir_instr_type_jump);
1241    }
1242
1243    after->block = instr->block;
1244    add_defs_uses(after);
1245    exec_node_insert_after(&instr->node, &after->node);
1246
1247    if (after->type == nir_instr_type_jump)
1248       handle_jump(after->block);
1249 }
1250
1251 void
1252 nir_instr_insert_before_block(nir_block *block, nir_instr *before)
1253 {
1254    if (before->type == nir_instr_type_jump)
1255       assert(exec_list_is_empty(&block->instr_list));
1256
1257    before->block = block;
1258    add_defs_uses(before);
1259    exec_list_push_head(&block->instr_list, &before->node);
1260
1261    if (before->type == nir_instr_type_jump)
1262       handle_jump(block);
1263 }
1264
1265 void
1266 nir_instr_insert_after_block(nir_block *block, nir_instr *after)
1267 {
1268    if (after->type == nir_instr_type_jump) {
1269       assert(exec_list_is_empty(&block->instr_list) ||
1270              nir_block_last_instr(block)->type != nir_instr_type_jump);
1271    }
1272
1273    after->block = block;
1274    add_defs_uses(after);
1275    exec_list_push_tail(&block->instr_list, &after->node);
1276
1277    if (after->type == nir_instr_type_jump)
1278       handle_jump(block);
1279 }
1280
1281 void
1282 nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before)
1283 {
1284    if (node->type == nir_cf_node_block) {
1285       nir_instr_insert_before_block(nir_cf_node_as_block(node), before);
1286    } else {
1287       nir_cf_node *prev = nir_cf_node_prev(node);
1288       assert(prev->type == nir_cf_node_block);
1289       nir_block *prev_block = nir_cf_node_as_block(prev);
1290
1291       nir_instr_insert_before_block(prev_block, before);
1292    }
1293 }
1294
1295 void
1296 nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after)
1297 {
1298    if (node->type == nir_cf_node_block) {
1299       nir_instr_insert_after_block(nir_cf_node_as_block(node), after);
1300    } else {
1301       nir_cf_node *next = nir_cf_node_next(node);
1302       assert(next->type == nir_cf_node_block);
1303       nir_block *next_block = nir_cf_node_as_block(next);
1304
1305       nir_instr_insert_before_block(next_block, after);
1306    }
1307 }
1308
1309 void
1310 nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before)
1311 {
1312    nir_cf_node *first_node = exec_node_data(nir_cf_node,
1313                                             exec_list_get_head(list), node);
1314    nir_instr_insert_before_cf(first_node, before);
1315 }
1316
1317 void
1318 nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after)
1319 {
1320    nir_cf_node *last_node = exec_node_data(nir_cf_node,
1321                                            exec_list_get_tail(list), node);
1322    nir_instr_insert_after_cf(last_node, after);
1323 }
1324
1325 static bool
1326 remove_use_cb(nir_src *src, void *state)
1327 {
1328    nir_instr *instr = (nir_instr *) state;
1329
1330    struct set *uses_set = src->is_ssa ? src->ssa->uses : src->reg.reg->uses;
1331
1332    struct set_entry *entry = _mesa_set_search(uses_set, instr);
1333    if (entry)
1334       _mesa_set_remove(uses_set, entry);
1335
1336    return true;
1337 }
1338
1339 static bool
1340 remove_def_cb(nir_dest *dest, void *state)
1341 {
1342    nir_instr *instr = (nir_instr *) state;
1343
1344    if (dest->is_ssa)
1345       return true;
1346
1347    nir_register *reg = dest->reg.reg;
1348
1349    struct set_entry *entry = _mesa_set_search(reg->defs, instr);
1350    if (entry)
1351       _mesa_set_remove(reg->defs, entry);
1352
1353    return true;
1354 }
1355
1356 static void
1357 remove_defs_uses(nir_instr *instr)
1358 {
1359    nir_foreach_dest(instr, remove_def_cb, instr);
1360    nir_foreach_src(instr, remove_use_cb, instr);
1361 }
1362
1363 void nir_instr_remove(nir_instr *instr)
1364 {
1365    remove_defs_uses(instr);
1366    exec_node_remove(&instr->node);
1367
1368    if (instr->type == nir_instr_type_jump) {
1369       nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
1370       handle_remove_jump(instr->block, jump_instr->type);
1371    }
1372 }
1373
1374 /*@}*/
1375
1376 void
1377 nir_index_local_regs(nir_function_impl *impl)
1378 {
1379    unsigned index = 0;
1380    foreach_list_typed(nir_register, reg, node, &impl->registers) {
1381       reg->index = index++;
1382    }
1383    impl->reg_alloc = index;
1384 }
1385
1386 void
1387 nir_index_global_regs(nir_shader *shader)
1388 {
1389    unsigned index = 0;
1390    foreach_list_typed(nir_register, reg, node, &shader->registers) {
1391       reg->index = index++;
1392    }
1393    shader->reg_alloc = index;
1394 }
1395
1396 static bool
1397 visit_alu_dest(nir_alu_instr *instr, nir_foreach_dest_cb cb, void *state)
1398 {
1399    return cb(&instr->dest.dest, state);
1400 }
1401
1402 static bool
1403 visit_intrinsic_dest(nir_intrinsic_instr *instr, nir_foreach_dest_cb cb,
1404                      void *state)
1405 {
1406    if (nir_intrinsic_infos[instr->intrinsic].has_dest)
1407       return cb(&instr->dest, state);
1408
1409    return true;
1410 }
1411
1412 static bool
1413 visit_texture_dest(nir_tex_instr *instr, nir_foreach_dest_cb cb,
1414                    void *state)
1415 {
1416    return cb(&instr->dest, state);
1417 }
1418
1419 static bool
1420 visit_phi_dest(nir_phi_instr *instr, nir_foreach_dest_cb cb, void *state)
1421 {
1422    return cb(&instr->dest, state);
1423 }
1424
1425 static bool
1426 visit_parallel_copy_dest(nir_parallel_copy_instr *instr,
1427                          nir_foreach_dest_cb cb, void *state)
1428 {
1429    nir_foreach_parallel_copy_entry(instr, entry) {
1430       if (!cb(&entry->dest, state))
1431          return false;
1432    }
1433
1434    return true;
1435 }
1436
1437 bool
1438 nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state)
1439 {
1440    switch (instr->type) {
1441    case nir_instr_type_alu:
1442       return visit_alu_dest(nir_instr_as_alu(instr), cb, state);
1443    case nir_instr_type_intrinsic:
1444       return visit_intrinsic_dest(nir_instr_as_intrinsic(instr), cb, state);
1445    case nir_instr_type_tex:
1446       return visit_texture_dest(nir_instr_as_tex(instr), cb, state);
1447    case nir_instr_type_phi:
1448       return visit_phi_dest(nir_instr_as_phi(instr), cb, state);
1449    case nir_instr_type_parallel_copy:
1450       return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr),
1451                                       cb, state);
1452
1453    case nir_instr_type_load_const:
1454    case nir_instr_type_ssa_undef:
1455    case nir_instr_type_call:
1456    case nir_instr_type_jump:
1457       break;
1458
1459    default:
1460       unreachable("Invalid instruction type");
1461       break;
1462    }
1463
1464    return true;
1465 }
1466
1467 struct foreach_ssa_def_state {
1468    nir_foreach_ssa_def_cb cb;
1469    void *client_state;
1470 };
1471
1472 static inline bool
1473 nir_ssa_def_visitor(nir_dest *dest, void *void_state)
1474 {
1475    struct foreach_ssa_def_state *state = void_state;
1476
1477    if (dest->is_ssa)
1478       return state->cb(&dest->ssa, state->client_state);
1479    else
1480       return true;
1481 }
1482
1483 bool
1484 nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb, void *state)
1485 {
1486    switch (instr->type) {
1487    case nir_instr_type_alu:
1488    case nir_instr_type_tex:
1489    case nir_instr_type_intrinsic:
1490    case nir_instr_type_phi:
1491    case nir_instr_type_parallel_copy: {
1492       struct foreach_ssa_def_state foreach_state = {cb, state};
1493       return nir_foreach_dest(instr, nir_ssa_def_visitor, &foreach_state);
1494    }
1495
1496    case nir_instr_type_load_const:
1497       return cb(&nir_instr_as_load_const(instr)->def, state);
1498    case nir_instr_type_ssa_undef:
1499       return cb(&nir_instr_as_ssa_undef(instr)->def, state);
1500    case nir_instr_type_call:
1501    case nir_instr_type_jump:
1502       return true;
1503    default:
1504       unreachable("Invalid instruction type");
1505    }
1506 }
1507
1508 static bool
1509 visit_src(nir_src *src, nir_foreach_src_cb cb, void *state)
1510 {
1511    if (!cb(src, state))
1512       return false;
1513    if (!src->is_ssa && src->reg.indirect)
1514       return cb(src->reg.indirect, state);
1515    return true;
1516 }
1517
1518 static bool
1519 visit_deref_array_src(nir_deref_array *deref, nir_foreach_src_cb cb,
1520                       void *state)
1521 {
1522    if (deref->deref_array_type == nir_deref_array_type_indirect)
1523       return visit_src(&deref->indirect, cb, state);
1524    return true;
1525 }
1526
1527 static bool
1528 visit_deref_src(nir_deref_var *deref, nir_foreach_src_cb cb, void *state)
1529 {
1530    nir_deref *cur = &deref->deref;
1531    while (cur != NULL) {
1532       if (cur->deref_type == nir_deref_type_array)
1533          if (!visit_deref_array_src(nir_deref_as_array(cur), cb, state))
1534             return false;
1535
1536       cur = cur->child;
1537    }
1538
1539    return true;
1540 }
1541
1542 static bool
1543 visit_alu_src(nir_alu_instr *instr, nir_foreach_src_cb cb, void *state)
1544 {
1545    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1546       if (!visit_src(&instr->src[i].src, cb, state))
1547          return false;
1548
1549    return true;
1550 }
1551
1552 static bool
1553 visit_tex_src(nir_tex_instr *instr, nir_foreach_src_cb cb, void *state)
1554 {
1555    for (unsigned i = 0; i < instr->num_srcs; i++)
1556       if (!visit_src(&instr->src[i].src, cb, state))
1557          return false;
1558
1559    if (instr->sampler != NULL)
1560       if (!visit_deref_src(instr->sampler, cb, state))
1561          return false;
1562
1563    return true;
1564 }
1565
1566 static bool
1567 visit_intrinsic_src(nir_intrinsic_instr *instr, nir_foreach_src_cb cb,
1568                     void *state)
1569 {
1570    unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
1571    for (unsigned i = 0; i < num_srcs; i++)
1572       if (!visit_src(&instr->src[i], cb, state))
1573          return false;
1574
1575    unsigned num_vars =
1576       nir_intrinsic_infos[instr->intrinsic].num_variables;
1577    for (unsigned i = 0; i < num_vars; i++)
1578       if (!visit_deref_src(instr->variables[i], cb, state))
1579          return false;
1580
1581    return true;
1582 }
1583
1584 static bool
1585 visit_call_src(nir_call_instr *instr, nir_foreach_src_cb cb, void *state)
1586 {
1587    return true;
1588 }
1589
1590 static bool
1591 visit_load_const_src(nir_load_const_instr *instr, nir_foreach_src_cb cb,
1592                      void *state)
1593 {
1594    return true;
1595 }
1596
1597 static bool
1598 visit_phi_src(nir_phi_instr *instr, nir_foreach_src_cb cb, void *state)
1599 {
1600    nir_foreach_phi_src(instr, src) {
1601       if (!visit_src(&src->src, cb, state))
1602          return false;
1603    }
1604
1605    return true;
1606 }
1607
1608 static bool
1609 visit_parallel_copy_src(nir_parallel_copy_instr *instr,
1610                         nir_foreach_src_cb cb, void *state)
1611 {
1612    nir_foreach_parallel_copy_entry(instr, entry) {
1613       if (!visit_src(&entry->src, cb, state))
1614          return false;
1615    }
1616
1617    return true;
1618 }
1619
1620 typedef struct {
1621    void *state;
1622    nir_foreach_src_cb cb;
1623 } visit_dest_indirect_state;
1624
1625 static bool
1626 visit_dest_indirect(nir_dest *dest, void *_state)
1627 {
1628    visit_dest_indirect_state *state = (visit_dest_indirect_state *) _state;
1629
1630    if (!dest->is_ssa && dest->reg.indirect)
1631       return state->cb(dest->reg.indirect, state->state);
1632
1633    return true;
1634 }
1635
1636 bool
1637 nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state)
1638 {
1639    switch (instr->type) {
1640    case nir_instr_type_alu:
1641       if (!visit_alu_src(nir_instr_as_alu(instr), cb, state))
1642          return false;
1643       break;
1644    case nir_instr_type_intrinsic:
1645       if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr), cb, state))
1646          return false;
1647       break;
1648    case nir_instr_type_tex:
1649       if (!visit_tex_src(nir_instr_as_tex(instr), cb, state))
1650          return false;
1651       break;
1652    case nir_instr_type_call:
1653       if (!visit_call_src(nir_instr_as_call(instr), cb, state))
1654          return false;
1655       break;
1656    case nir_instr_type_load_const:
1657       if (!visit_load_const_src(nir_instr_as_load_const(instr), cb, state))
1658          return false;
1659       break;
1660    case nir_instr_type_phi:
1661       if (!visit_phi_src(nir_instr_as_phi(instr), cb, state))
1662          return false;
1663       break;
1664    case nir_instr_type_parallel_copy:
1665       if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr),
1666                                    cb, state))
1667          return false;
1668       break;
1669    case nir_instr_type_jump:
1670    case nir_instr_type_ssa_undef:
1671       return true;
1672
1673    default:
1674       unreachable("Invalid instruction type");
1675       break;
1676    }
1677
1678    visit_dest_indirect_state dest_state;
1679    dest_state.state = state;
1680    dest_state.cb = cb;
1681    return nir_foreach_dest(instr, visit_dest_indirect, &dest_state);
1682 }
1683
1684 nir_const_value *
1685 nir_src_as_const_value(nir_src src)
1686 {
1687    if (!src.is_ssa)
1688       return NULL;
1689
1690    if (src.ssa->parent_instr->type != nir_instr_type_load_const)
1691       return NULL;
1692
1693    nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
1694
1695    return &load->value;
1696 }
1697
1698 bool
1699 nir_srcs_equal(nir_src src1, nir_src src2)
1700 {
1701    if (src1.is_ssa) {
1702       if (src2.is_ssa) {
1703          return src1.ssa == src2.ssa;
1704       } else {
1705          return false;
1706       }
1707    } else {
1708       if (src2.is_ssa) {
1709          return false;
1710       } else {
1711          if ((src1.reg.indirect == NULL) != (src2.reg.indirect == NULL))
1712             return false;
1713
1714          if (src1.reg.indirect) {
1715             if (!nir_srcs_equal(*src1.reg.indirect, *src2.reg.indirect))
1716                return false;
1717          }
1718
1719          return src1.reg.reg == src2.reg.reg &&
1720                 src1.reg.base_offset == src2.reg.base_offset;
1721       }
1722    }
1723 }
1724
1725 static bool
1726 src_does_not_use_def(nir_src *src, void *void_def)
1727 {
1728    nir_ssa_def *def = void_def;
1729
1730    if (src->is_ssa) {
1731       return src->ssa != def;
1732    } else {
1733       return true;
1734    }
1735 }
1736
1737 static bool
1738 src_does_not_use_reg(nir_src *src, void *void_reg)
1739 {
1740    nir_register *reg = void_reg;
1741
1742    if (src->is_ssa) {
1743       return true;
1744    } else {
1745       return src->reg.reg != reg;
1746    }
1747 }
1748
1749 void
1750 nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src)
1751 {
1752    if (src->is_ssa) {
1753       nir_ssa_def *old_ssa = src->ssa;
1754       *src = new_src;
1755       if (old_ssa && nir_foreach_src(instr, src_does_not_use_def, old_ssa)) {
1756          struct set_entry *entry = _mesa_set_search(old_ssa->uses, instr);
1757          assert(entry);
1758          _mesa_set_remove(old_ssa->uses, entry);
1759       }
1760    } else {
1761       if (src->reg.indirect)
1762          nir_instr_rewrite_src(instr, src->reg.indirect, new_src);
1763
1764       nir_register *old_reg = src->reg.reg;
1765       *src = new_src;
1766       if (old_reg && nir_foreach_src(instr, src_does_not_use_reg, old_reg)) {
1767          struct set_entry *entry = _mesa_set_search(old_reg->uses, instr);
1768          assert(entry);
1769          _mesa_set_remove(old_reg->uses, entry);
1770       }
1771    }
1772
1773    if (new_src.is_ssa) {
1774       if (new_src.ssa)
1775          _mesa_set_add(new_src.ssa->uses, instr);
1776    } else {
1777       if (new_src.reg.reg)
1778          _mesa_set_add(new_src.reg.reg->uses, instr);
1779    }
1780 }
1781
1782 void
1783 nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def,
1784                  unsigned num_components, const char *name)
1785 {
1786    void *mem_ctx = ralloc_parent(instr);
1787
1788    def->name = name;
1789    def->parent_instr = instr;
1790    def->uses = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
1791                                 _mesa_key_pointer_equal);
1792    def->if_uses = _mesa_set_create(mem_ctx, _mesa_hash_pointer,
1793                                    _mesa_key_pointer_equal);
1794    def->num_components = num_components;
1795
1796    if (instr->block) {
1797       nir_function_impl *impl =
1798          nir_cf_node_get_function(&instr->block->cf_node);
1799
1800       def->index = impl->ssa_alloc++;
1801    } else {
1802       def->index = UINT_MAX;
1803    }
1804 }
1805
1806 void
1807 nir_ssa_dest_init(nir_instr *instr, nir_dest *dest,
1808                  unsigned num_components, const char *name)
1809 {
1810    dest->is_ssa = true;
1811    nir_ssa_def_init(instr, &dest->ssa, num_components, name);
1812 }
1813
1814 struct ssa_def_rewrite_state {
1815    void *mem_ctx;
1816    nir_ssa_def *old;
1817    nir_src new_src;
1818 };
1819
1820 static bool
1821 ssa_def_rewrite_uses_src(nir_src *src, void *void_state)
1822 {
1823    struct ssa_def_rewrite_state *state = void_state;
1824
1825    if (src->is_ssa && src->ssa == state->old)
1826       nir_src_copy(src, &state->new_src, state->mem_ctx);
1827
1828    return true;
1829 }
1830
1831 void
1832 nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src, void *mem_ctx)
1833 {
1834    struct ssa_def_rewrite_state state;
1835    state.mem_ctx = mem_ctx;
1836    state.old = def;
1837    state.new_src = new_src;
1838
1839    assert(!new_src.is_ssa || def != new_src.ssa);
1840
1841    struct set *new_uses, *new_if_uses;
1842    if (new_src.is_ssa) {
1843       new_uses = new_src.ssa->uses;
1844       new_if_uses = new_src.ssa->if_uses;
1845    } else {
1846       new_uses = new_src.reg.reg->uses;
1847       new_if_uses = new_src.reg.reg->if_uses;
1848    }
1849
1850    struct set_entry *entry;
1851    set_foreach(def->uses, entry) {
1852       nir_instr *instr = (nir_instr *)entry->key;
1853
1854       _mesa_set_remove(def->uses, entry);
1855       nir_foreach_src(instr, ssa_def_rewrite_uses_src, &state);
1856       _mesa_set_add(new_uses, instr);
1857    }
1858
1859    set_foreach(def->if_uses, entry) {
1860       nir_if *if_use = (nir_if *)entry->key;
1861
1862       _mesa_set_remove(def->if_uses, entry);
1863       nir_src_copy(&if_use->condition, &new_src, mem_ctx);
1864       _mesa_set_add(new_if_uses, if_use);
1865    }
1866 }
1867
1868
1869 static bool foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb,
1870                             bool reverse, void *state);
1871
1872 static inline bool
1873 foreach_if(nir_if *if_stmt, nir_foreach_block_cb cb, bool reverse, void *state)
1874 {
1875    if (reverse) {
1876       foreach_list_typed_safe_reverse(nir_cf_node, node, node,
1877                                       &if_stmt->else_list) {
1878          if (!foreach_cf_node(node, cb, reverse, state))
1879             return false;
1880       }
1881
1882       foreach_list_typed_safe_reverse(nir_cf_node, node, node,
1883                                       &if_stmt->then_list) {
1884          if (!foreach_cf_node(node, cb, reverse, state))
1885             return false;
1886       }
1887    } else {
1888       foreach_list_typed_safe(nir_cf_node, node, node, &if_stmt->then_list) {
1889          if (!foreach_cf_node(node, cb, reverse, state))
1890             return false;
1891       }
1892
1893       foreach_list_typed_safe(nir_cf_node, node, node, &if_stmt->else_list) {
1894          if (!foreach_cf_node(node, cb, reverse, state))
1895             return false;
1896       }
1897    }
1898
1899    return true;
1900 }
1901
1902 static inline bool
1903 foreach_loop(nir_loop *loop, nir_foreach_block_cb cb, bool reverse, void *state)
1904 {
1905    if (reverse) {
1906       foreach_list_typed_safe_reverse(nir_cf_node, node, node, &loop->body) {
1907          if (!foreach_cf_node(node, cb, reverse, state))
1908             return false;
1909       }
1910    } else {
1911       foreach_list_typed_safe(nir_cf_node, node, node, &loop->body) {
1912          if (!foreach_cf_node(node, cb, reverse, state))
1913             return false;
1914       }
1915    }
1916
1917    return true;
1918 }
1919
1920 static bool
1921 foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb,
1922                 bool reverse, void *state)
1923 {
1924    switch (node->type) {
1925    case nir_cf_node_block:
1926       return cb(nir_cf_node_as_block(node), state);
1927    case nir_cf_node_if:
1928       return foreach_if(nir_cf_node_as_if(node), cb, reverse, state);
1929    case nir_cf_node_loop:
1930       return foreach_loop(nir_cf_node_as_loop(node), cb, reverse, state);
1931       break;
1932
1933    default:
1934       unreachable("Invalid CFG node type");
1935       break;
1936    }
1937
1938    return false;
1939 }
1940
1941 bool
1942 nir_foreach_block(nir_function_impl *impl, nir_foreach_block_cb cb, void *state)
1943 {
1944    foreach_list_typed_safe(nir_cf_node, node, node, &impl->body) {
1945       if (!foreach_cf_node(node, cb, false, state))
1946          return false;
1947    }
1948
1949    return cb(impl->end_block, state);
1950 }
1951
1952 bool
1953 nir_foreach_block_reverse(nir_function_impl *impl, nir_foreach_block_cb cb,
1954                           void *state)
1955 {
1956    if (!cb(impl->end_block, state))
1957       return false;
1958
1959    foreach_list_typed_safe_reverse(nir_cf_node, node, node, &impl->body) {
1960       if (!foreach_cf_node(node, cb, true, state))
1961          return false;
1962    }
1963
1964    return true;
1965 }
1966
1967 nir_if *
1968 nir_block_get_following_if(nir_block *block)
1969 {
1970    if (exec_node_is_tail_sentinel(&block->cf_node.node))
1971       return NULL;
1972
1973    if (nir_cf_node_is_last(&block->cf_node))
1974       return NULL;
1975
1976    nir_cf_node *next_node = nir_cf_node_next(&block->cf_node);
1977
1978    if (next_node->type != nir_cf_node_if)
1979       return NULL;
1980
1981    return nir_cf_node_as_if(next_node);
1982 }
1983
1984 static bool
1985 index_block(nir_block *block, void *state)
1986 {
1987    unsigned *index = (unsigned *) state;
1988    block->index = (*index)++;
1989    return true;
1990 }
1991
1992 void
1993 nir_index_blocks(nir_function_impl *impl)
1994 {
1995    unsigned index = 0;
1996
1997    if (impl->valid_metadata & nir_metadata_block_index)
1998       return;
1999
2000    nir_foreach_block(impl, index_block, &index);
2001
2002    impl->num_blocks = index;
2003 }
2004
2005 static void
2006 index_ssa_def(nir_ssa_def *def, unsigned *index)
2007 {
2008    def->index = (*index)++;
2009 }
2010
2011 static bool
2012 index_ssa_def_cb(nir_dest *dest, void *state)
2013 {
2014    unsigned *index = (unsigned *) state;
2015    if (dest->is_ssa)
2016       index_ssa_def(&dest->ssa, index);
2017    return true;
2018 }
2019
2020 static void
2021 index_ssa_undef(nir_ssa_undef_instr *instr, unsigned *index)
2022 {
2023    index_ssa_def(&instr->def, index);
2024 }
2025
2026 static bool
2027 index_ssa_block(nir_block *block, void *state)
2028 {
2029    unsigned *index = (unsigned *) state;
2030
2031    nir_foreach_instr(block, instr) {
2032       if (instr->type == nir_instr_type_ssa_undef)
2033          index_ssa_undef(nir_instr_as_ssa_undef(instr), index);
2034       else
2035          nir_foreach_dest(instr, index_ssa_def_cb, state);
2036    }
2037
2038    return true;
2039 }
2040
2041 void
2042 nir_index_ssa_defs(nir_function_impl *impl)
2043 {
2044    unsigned index = 0;
2045    nir_foreach_block(impl, index_ssa_block, &index);
2046    impl->ssa_alloc = index;
2047 }