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 "nir_control_flow_private.h"
  30 #include <assert.h>
  31
  32 nir_shader *
  33 nir_shader_create(void *mem_ctx,
  34                   gl_shader_stage stage,
  35                   const nir_shader_compiler_options *options)
  36 {
  37    nir_shader *shader = ralloc(mem_ctx, nir_shader);
  38
  39    exec_list_make_empty(&shader->uniforms);
  40    exec_list_make_empty(&shader->inputs);
  41    exec_list_make_empty(&shader->outputs);
  42
  43    shader->options = options;
  44
  45    exec_list_make_empty(&shader->functions);
  46    exec_list_make_empty(&shader->registers);
  47    exec_list_make_empty(&shader->globals);
  48    exec_list_make_empty(&shader->system_values);
  49    shader->reg_alloc = 0;
  50
  51    shader->num_inputs = 0;
  52    shader->num_outputs = 0;
  53    shader->num_uniforms = 0;
  54
  55    shader->stage = stage;
  56
  57    return shader;
  58 }
  59
  60 static nir_register *
  61 reg_create(void *mem_ctx, struct exec_list *list)
  62 {
  63    nir_register *reg = ralloc(mem_ctx, nir_register);
  64
  65    list_inithead(&reg->uses);
  66    list_inithead(&reg->defs);
  67    list_inithead(&reg->if_uses);
  68
  69    reg->num_components = 0;
  70    reg->num_array_elems = 0;
  71    reg->is_packed = false;
  72    reg->name = NULL;
  73
  74    exec_list_push_tail(list, &reg->node);
  75
  76    return reg;
  77 }
  78
  79 nir_register *
  80 nir_global_reg_create(nir_shader *shader)
  81 {
  82    nir_register *reg = reg_create(shader, &shader->registers);
  83    reg->index = shader->reg_alloc++;
  84    reg->is_global = true;
  85
  86    return reg;
  87 }
  88
  89 nir_register *
  90 nir_local_reg_create(nir_function_impl *impl)
  91 {
  92    nir_register *reg = reg_create(ralloc_parent(impl), &impl->registers);
  93    reg->index = impl->reg_alloc++;
  94    reg->is_global = false;
  95
  96    return reg;
  97 }
  98
  99 void
 100 nir_reg_remove(nir_register *reg)
 101 {
 102    exec_node_remove(&reg->node);
 103 }
 104
 105 nir_function *
 106 nir_function_create(nir_shader *shader, const char *name)
 107 {
 108    nir_function *func = ralloc(shader, nir_function);
 109
 110    exec_list_push_tail(&shader->functions, &func->node);
 111    exec_list_make_empty(&func->overload_list);
 112    func->name = ralloc_strdup(func, name);
 113    func->shader = shader;
 114
 115    return func;
 116 }
 117
 118 nir_function_overload *
 119 nir_function_overload_create(nir_function *func)
 120 {
 121    void *mem_ctx = ralloc_parent(func);
 122
 123    nir_function_overload *overload = ralloc(mem_ctx, nir_function_overload);
 124
 125    overload->num_params = 0;
 126    overload->params = NULL;
 127    overload->return_type = glsl_void_type();
 128    overload->impl = NULL;
 129
 130    exec_list_push_tail(&func->overload_list, &overload->node);
 131    overload->function = func;
 132
 133    return overload;
 134 }
 135
 136 void nir_src_copy(nir_src *dest, const nir_src *src, void *mem_ctx)
 137 {
 138    dest->is_ssa = src->is_ssa;
 139    if (src->is_ssa) {
 140       dest->ssa = src->ssa;
 141    } else {
 142       dest->reg.base_offset = src->reg.base_offset;
 143       dest->reg.reg = src->reg.reg;
 144       if (src->reg.indirect) {
 145          dest->reg.indirect = ralloc(mem_ctx, nir_src);
 146          nir_src_copy(dest->reg.indirect, src->reg.indirect, mem_ctx);
 147       } else {
 148          dest->reg.indirect = NULL;
 149       }
 150    }
 151 }
 152
 153 void nir_dest_copy(nir_dest *dest, const nir_dest *src, void *mem_ctx)
 154 {
 155    /* Copying an SSA definition makes no sense whatsoever. */
 156    assert(!src->is_ssa);
 157
 158    dest->is_ssa = false;
 159
 160    dest->reg.base_offset = src->reg.base_offset;
 161    dest->reg.reg = src->reg.reg;
 162    if (src->reg.indirect) {
 163       dest->reg.indirect = ralloc(mem_ctx, nir_src);
 164       nir_src_copy(dest->reg.indirect, src->reg.indirect, mem_ctx);
 165    } else {
 166       dest->reg.indirect = NULL;
 167    }
 168 }
 169
 170 void
 171 nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src, void *mem_ctx)
 172 {
 173    nir_src_copy(&dest->src, &src->src, mem_ctx);
 174    dest->abs = src->abs;
 175    dest->negate = src->negate;
 176    for (unsigned i = 0; i < 4; i++)
 177       dest->swizzle[i] = src->swizzle[i];
 178 }
 179
 180 void
 181 nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src, void *mem_ctx)
 182 {
 183    nir_dest_copy(&dest->dest, &src->dest, mem_ctx);
 184    dest->write_mask = src->write_mask;
 185    dest->saturate = src->saturate;
 186 }
 187
 188
 189 static void
 190 cf_init(nir_cf_node *node, nir_cf_node_type type)
 191 {
 192    exec_node_init(&node->node);
 193    node->parent = NULL;
 194    node->type = type;
 195 }
 196
 197 nir_function_impl *
 198 nir_function_impl_create(nir_function_overload *overload)
 199 {
 200    assert(overload->impl == NULL);
 201
 202    void *mem_ctx = ralloc_parent(overload);
 203
 204    nir_function_impl *impl = ralloc(mem_ctx, nir_function_impl);
 205
 206    overload->impl = impl;
 207    impl->overload = overload;
 208
 209    cf_init(&impl->cf_node, nir_cf_node_function);
 210
 211    exec_list_make_empty(&impl->body);
 212    exec_list_make_empty(&impl->registers);
 213    exec_list_make_empty(&impl->locals);
 214    impl->num_params = 0;
 215    impl->params = NULL;
 216    impl->return_var = NULL;
 217    impl->reg_alloc = 0;
 218    impl->ssa_alloc = 0;
 219    impl->valid_metadata = nir_metadata_none;
 220
 221    /* create start & end blocks */
 222    nir_block *start_block = nir_block_create(mem_ctx);
 223    nir_block *end_block = nir_block_create(mem_ctx);
 224    start_block->cf_node.parent = &impl->cf_node;
 225    end_block->cf_node.parent = &impl->cf_node;
 226    impl->end_block = end_block;
 227
 228    exec_list_push_tail(&impl->body, &start_block->cf_node.node);
 229
 230    start_block->successors[0] = end_block;
 231    _mesa_set_add(end_block->predecessors, start_block);
 232    return impl;
 233 }
 234
 235 nir_block *
 236 nir_block_create(void *mem_ctx)
 237 {
 238    nir_block *block = ralloc(mem_ctx, nir_block);
 239
 240    cf_init(&block->cf_node, nir_cf_node_block);
 241
 242    block->successors[0] = block->successors[1] = NULL;
 243    block->predecessors = _mesa_set_create(block, _mesa_hash_pointer,
 244                                           _mesa_key_pointer_equal);
 245    block->imm_dom = NULL;
 246    block->dom_frontier = _mesa_set_create(block, _mesa_hash_pointer,
 247                                           _mesa_key_pointer_equal);
 248
 249    exec_list_make_empty(&block->instr_list);
 250
 251    return block;
 252 }
 253
 254 static inline void
 255 src_init(nir_src *src)
 256 {
 257    src->is_ssa = false;
 258    src->reg.reg = NULL;
 259    src->reg.indirect = NULL;
 260    src->reg.base_offset = 0;
 261 }
 262
 263 nir_if *
 264 nir_if_create(void *mem_ctx)
 265 {
 266    nir_if *if_stmt = ralloc(mem_ctx, nir_if);
 267
 268    cf_init(&if_stmt->cf_node, nir_cf_node_if);
 269    src_init(&if_stmt->condition);
 270
 271    nir_block *then = nir_block_create(mem_ctx);
 272    exec_list_make_empty(&if_stmt->then_list);
 273    exec_list_push_tail(&if_stmt->then_list, &then->cf_node.node);
 274    then->cf_node.parent = &if_stmt->cf_node;
 275
 276    nir_block *else_stmt = nir_block_create(mem_ctx);
 277    exec_list_make_empty(&if_stmt->else_list);
 278    exec_list_push_tail(&if_stmt->else_list, &else_stmt->cf_node.node);
 279    else_stmt->cf_node.parent = &if_stmt->cf_node;
 280
 281    return if_stmt;
 282 }
 283
 284 nir_loop *
 285 nir_loop_create(void *mem_ctx)
 286 {
 287    nir_loop *loop = ralloc(mem_ctx, nir_loop);
 288
 289    cf_init(&loop->cf_node, nir_cf_node_loop);
 290
 291    nir_block *body = nir_block_create(mem_ctx);
 292    exec_list_make_empty(&loop->body);
 293    exec_list_push_tail(&loop->body, &body->cf_node.node);
 294    body->cf_node.parent = &loop->cf_node;
 295
 296    body->successors[0] = body;
 297    _mesa_set_add(body->predecessors, body);
 298
 299    return loop;
 300 }
 301
 302 static void
 303 instr_init(nir_instr *instr, nir_instr_type type)
 304 {
 305    instr->type = type;
 306    instr->block = NULL;
 307    exec_node_init(&instr->node);
 308 }
 309
 310 static void
 311 dest_init(nir_dest *dest)
 312 {
 313    dest->is_ssa = false;
 314    dest->reg.reg = NULL;
 315    dest->reg.indirect = NULL;
 316    dest->reg.base_offset = 0;
 317 }
 318
 319 static void
 320 alu_dest_init(nir_alu_dest *dest)
 321 {
 322    dest_init(&dest->dest);
 323    dest->saturate = false;
 324    dest->write_mask = 0xf;
 325 }
 326
 327 static void
 328 alu_src_init(nir_alu_src *src)
 329 {
 330    src_init(&src->src);
 331    src->abs = src->negate = false;
 332    src->swizzle[0] = 0;
 333    src->swizzle[1] = 1;
 334    src->swizzle[2] = 2;
 335    src->swizzle[3] = 3;
 336 }
 337
 338 nir_alu_instr *
 339 nir_alu_instr_create(nir_shader *shader, nir_op op)
 340 {
 341    unsigned num_srcs = nir_op_infos[op].num_inputs;
 342    nir_alu_instr *instr =
 343       ralloc_size(shader,
 344                   sizeof(nir_alu_instr) + num_srcs * sizeof(nir_alu_src));
 345
 346    instr_init(&instr->instr, nir_instr_type_alu);
 347    instr->op = op;
 348    alu_dest_init(&instr->dest);
 349    for (unsigned i = 0; i < num_srcs; i++)
 350       alu_src_init(&instr->src[i]);
 351
 352    return instr;
 353 }
 354
 355 nir_jump_instr *
 356 nir_jump_instr_create(nir_shader *shader, nir_jump_type type)
 357 {
 358    nir_jump_instr *instr = ralloc(shader, nir_jump_instr);
 359    instr_init(&instr->instr, nir_instr_type_jump);
 360    instr->type = type;
 361    return instr;
 362 }
 363
 364 nir_load_const_instr *
 365 nir_load_const_instr_create(nir_shader *shader, unsigned num_components)
 366 {
 367    nir_load_const_instr *instr = ralloc(shader, nir_load_const_instr);
 368    instr_init(&instr->instr, nir_instr_type_load_const);
 369
 370    nir_ssa_def_init(&instr->instr, &instr->def, num_components, NULL);
 371
 372    return instr;
 373 }
 374
 375 nir_intrinsic_instr *
 376 nir_intrinsic_instr_create(nir_shader *shader, nir_intrinsic_op op)
 377 {
 378    unsigned num_srcs = nir_intrinsic_infos[op].num_srcs;
 379    nir_intrinsic_instr *instr =
 380       ralloc_size(shader,
 381                   sizeof(nir_intrinsic_instr) + num_srcs * sizeof(nir_src));
 382
 383    instr_init(&instr->instr, nir_instr_type_intrinsic);
 384    instr->intrinsic = op;
 385
 386    if (nir_intrinsic_infos[op].has_dest)
 387       dest_init(&instr->dest);
 388
 389    for (unsigned i = 0; i < num_srcs; i++)
 390       src_init(&instr->src[i]);
 391
 392    return instr;
 393 }
 394
 395 nir_call_instr *
 396 nir_call_instr_create(nir_shader *shader, nir_function_overload *callee)
 397 {
 398    nir_call_instr *instr = ralloc(shader, nir_call_instr);
 399    instr_init(&instr->instr, nir_instr_type_call);
 400
 401    instr->callee = callee;
 402    instr->num_params = callee->num_params;
 403    instr->params = ralloc_array(instr, nir_deref_var *, instr->num_params);
 404    instr->return_deref = NULL;
 405
 406    return instr;
 407 }
 408
 409 nir_tex_instr *
 410 nir_tex_instr_create(nir_shader *shader, unsigned num_srcs)
 411 {
 412    nir_tex_instr *instr = rzalloc(shader, nir_tex_instr);
 413    instr_init(&instr->instr, nir_instr_type_tex);
 414
 415    dest_init(&instr->dest);
 416
 417    instr->num_srcs = num_srcs;
 418    instr->src = ralloc_array(instr, nir_tex_src, num_srcs);
 419    for (unsigned i = 0; i < num_srcs; i++)
 420       src_init(&instr->src[i].src);
 421
 422    instr->sampler_index = 0;
 423    instr->sampler_array_size = 0;
 424    instr->sampler = NULL;
 425
 426    return instr;
 427 }
 428
 429 nir_phi_instr *
 430 nir_phi_instr_create(nir_shader *shader)
 431 {
 432    nir_phi_instr *instr = ralloc(shader, nir_phi_instr);
 433    instr_init(&instr->instr, nir_instr_type_phi);
 434
 435    dest_init(&instr->dest);
 436    exec_list_make_empty(&instr->srcs);
 437    return instr;
 438 }
 439
 440 nir_parallel_copy_instr *
 441 nir_parallel_copy_instr_create(nir_shader *shader)
 442 {
 443    nir_parallel_copy_instr *instr = ralloc(shader, nir_parallel_copy_instr);
 444    instr_init(&instr->instr, nir_instr_type_parallel_copy);
 445
 446    exec_list_make_empty(&instr->entries);
 447
 448    return instr;
 449 }
 450
 451 nir_ssa_undef_instr *
 452 nir_ssa_undef_instr_create(nir_shader *shader, unsigned num_components)
 453 {
 454    nir_ssa_undef_instr *instr = ralloc(shader, nir_ssa_undef_instr);
 455    instr_init(&instr->instr, nir_instr_type_ssa_undef);
 456
 457    nir_ssa_def_init(&instr->instr, &instr->def, num_components, NULL);
 458
 459    return instr;
 460 }
 461
 462 nir_deref_var *
 463 nir_deref_var_create(void *mem_ctx, nir_variable *var)
 464 {
 465    nir_deref_var *deref = ralloc(mem_ctx, nir_deref_var);
 466    deref->deref.deref_type = nir_deref_type_var;
 467    deref->deref.child = NULL;
 468    deref->deref.type = var->type;
 469    deref->var = var;
 470    return deref;
 471 }
 472
 473 nir_deref_array *
 474 nir_deref_array_create(void *mem_ctx)
 475 {
 476    nir_deref_array *deref = ralloc(mem_ctx, nir_deref_array);
 477    deref->deref.deref_type = nir_deref_type_array;
 478    deref->deref.child = NULL;
 479    deref->deref_array_type = nir_deref_array_type_direct;
 480    src_init(&deref->indirect);
 481    deref->base_offset = 0;
 482    return deref;
 483 }
 484
 485 nir_deref_struct *
 486 nir_deref_struct_create(void *mem_ctx, unsigned field_index)
 487 {
 488    nir_deref_struct *deref = ralloc(mem_ctx, nir_deref_struct);
 489    deref->deref.deref_type = nir_deref_type_struct;
 490    deref->deref.child = NULL;
 491    deref->index = field_index;
 492    return deref;
 493 }
 494
 495 static nir_deref_var *
 496 copy_deref_var(void *mem_ctx, nir_deref_var *deref)
 497 {
 498    nir_deref_var *ret = nir_deref_var_create(mem_ctx, deref->var);
 499    ret->deref.type = deref->deref.type;
 500    if (deref->deref.child)
 501       ret->deref.child = nir_copy_deref(ret, deref->deref.child);
 502    return ret;
 503 }
 504
 505 static nir_deref_array *
 506 copy_deref_array(void *mem_ctx, nir_deref_array *deref)
 507 {
 508    nir_deref_array *ret = nir_deref_array_create(mem_ctx);
 509    ret->base_offset = deref->base_offset;
 510    ret->deref_array_type = deref->deref_array_type;
 511    if (deref->deref_array_type == nir_deref_array_type_indirect) {
 512       nir_src_copy(&ret->indirect, &deref->indirect, mem_ctx);
 513    }
 514    ret->deref.type = deref->deref.type;
 515    if (deref->deref.child)
 516       ret->deref.child = nir_copy_deref(ret, deref->deref.child);
 517    return ret;
 518 }
 519
 520 static nir_deref_struct *
 521 copy_deref_struct(void *mem_ctx, nir_deref_struct *deref)
 522 {
 523    nir_deref_struct *ret = nir_deref_struct_create(mem_ctx, deref->index);
 524    ret->deref.type = deref->deref.type;
 525    if (deref->deref.child)
 526       ret->deref.child = nir_copy_deref(ret, deref->deref.child);
 527    return ret;
 528 }
 529
 530 nir_deref *
 531 nir_copy_deref(void *mem_ctx, nir_deref *deref)
 532 {
 533    switch (deref->deref_type) {
 534    case nir_deref_type_var:
 535       return &copy_deref_var(mem_ctx, nir_deref_as_var(deref))->deref;
 536    case nir_deref_type_array:
 537       return &copy_deref_array(mem_ctx, nir_deref_as_array(deref))->deref;
 538    case nir_deref_type_struct:
 539       return &copy_deref_struct(mem_ctx, nir_deref_as_struct(deref))->deref;
 540    default:
 541       unreachable("Invalid dereference type");
 542    }
 543
 544    return NULL;
 545 }
 546
 547 /* Returns a load_const instruction that represents the constant
 548  * initializer for the given deref chain.  The caller is responsible for
 549  * ensuring that there actually is a constant initializer.
 550  */
 551 nir_load_const_instr *
 552 nir_deref_get_const_initializer_load(nir_shader *shader, nir_deref_var *deref)
 553 {
 554    nir_constant *constant = deref->var->constant_initializer;
 555    assert(constant);
 556
 557    const nir_deref *tail = &deref->deref;
 558    unsigned matrix_offset = 0;
 559    while (tail->child) {
 560       switch (tail->child->deref_type) {
 561       case nir_deref_type_array: {
 562          nir_deref_array *arr = nir_deref_as_array(tail->child);
 563          assert(arr->deref_array_type == nir_deref_array_type_direct);
 564          if (glsl_type_is_matrix(tail->type)) {
 565             assert(arr->deref.child == NULL);
 566             matrix_offset = arr->base_offset;
 567          } else {
 568             constant = constant->elements[arr->base_offset];
 569          }
 570          break;
 571       }
 572
 573       case nir_deref_type_struct: {
 574          constant = constant->elements[nir_deref_as_struct(tail->child)->index];
 575          break;
 576       }
 577
 578       default:
 579          unreachable("Invalid deref child type");
 580       }
 581
 582       tail = tail->child;
 583    }
 584
 585    nir_load_const_instr *load =
 586       nir_load_const_instr_create(shader, glsl_get_vector_elements(tail->type));
 587
 588    matrix_offset *= load->def.num_components;
 589    for (unsigned i = 0; i < load->def.num_components; i++) {
 590       switch (glsl_get_base_type(tail->type)) {
 591       case GLSL_TYPE_FLOAT:
 592       case GLSL_TYPE_INT:
 593       case GLSL_TYPE_UINT:
 594          load->value.u[i] = constant->value.u[matrix_offset + i];
 595          break;
 596       case GLSL_TYPE_BOOL:
 597          load->value.u[i] = constant->value.b[matrix_offset + i] ?
 598                              NIR_TRUE : NIR_FALSE;
 599          break;
 600       default:
 601          unreachable("Invalid immediate type");
 602       }
 603    }
 604
 605    return load;
 606 }
 607
 608 nir_function_impl *
 609 nir_cf_node_get_function(nir_cf_node *node)
 610 {
 611    while (node->type != nir_cf_node_function) {
 612       node = node->parent;
 613    }
 614
 615    return nir_cf_node_as_function(node);
 616 }
 617
 618 static bool
 619 add_use_cb(nir_src *src, void *state)
 620 {
 621    nir_instr *instr = state;
 622
 623    src->parent_instr = instr;
 624    list_addtail(&src->use_link,
 625                 src->is_ssa ? &src->ssa->uses : &src->reg.reg->uses);
 626
 627    return true;
 628 }
 629
 630 static bool
 631 add_ssa_def_cb(nir_ssa_def *def, void *state)
 632 {
 633    nir_instr *instr = state;
 634
 635    if (instr->block && def->index == UINT_MAX) {
 636       nir_function_impl *impl =
 637          nir_cf_node_get_function(&instr->block->cf_node);
 638
 639       def->index = impl->ssa_alloc++;
 640    }
 641
 642    return true;
 643 }
 644
 645 static bool
 646 add_reg_def_cb(nir_dest *dest, void *state)
 647 {
 648    nir_instr *instr = state;
 649
 650    if (!dest->is_ssa) {
 651       dest->reg.parent_instr = instr;
 652       list_addtail(&dest->reg.def_link, &dest->reg.reg->defs);
 653    }
 654
 655    return true;
 656 }
 657
 658 static void
 659 add_defs_uses(nir_instr *instr)
 660 {
 661    nir_foreach_src(instr, add_use_cb, instr);
 662    nir_foreach_dest(instr, add_reg_def_cb, instr);
 663    nir_foreach_ssa_def(instr, add_ssa_def_cb, instr);
 664 }
 665
 666 void
 667 nir_instr_insert_before(nir_instr *instr, nir_instr *before)
 668 {
 669    assert(before->type != nir_instr_type_jump);
 670    before->block = instr->block;
 671    add_defs_uses(before);
 672    exec_node_insert_node_before(&instr->node, &before->node);
 673 }
 674
 675 void
 676 nir_instr_insert_after(nir_instr *instr, nir_instr *after)
 677 {
 678    if (after->type == nir_instr_type_jump) {
 679       assert(instr == nir_block_last_instr(instr->block));
 680       assert(instr->type != nir_instr_type_jump);
 681    }
 682
 683    after->block = instr->block;
 684    add_defs_uses(after);
 685    exec_node_insert_after(&instr->node, &after->node);
 686
 687    if (after->type == nir_instr_type_jump)
 688       nir_handle_add_jump(after->block);
 689 }
 690
 691 void
 692 nir_instr_insert_before_block(nir_block *block, nir_instr *before)
 693 {
 694    if (before->type == nir_instr_type_jump)
 695       assert(exec_list_is_empty(&block->instr_list));
 696
 697    before->block = block;
 698    add_defs_uses(before);
 699    exec_list_push_head(&block->instr_list, &before->node);
 700
 701    if (before->type == nir_instr_type_jump)
 702       nir_handle_add_jump(block);
 703 }
 704
 705 void
 706 nir_instr_insert_after_block(nir_block *block, nir_instr *after)
 707 {
 708    if (after->type == nir_instr_type_jump) {
 709       assert(exec_list_is_empty(&block->instr_list) ||
 710              nir_block_last_instr(block)->type != nir_instr_type_jump);
 711    }
 712
 713    after->block = block;
 714    add_defs_uses(after);
 715    exec_list_push_tail(&block->instr_list, &after->node);
 716
 717    if (after->type == nir_instr_type_jump)
 718       nir_handle_add_jump(block);
 719 }
 720
 721 void
 722 nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before)
 723 {
 724    if (node->type == nir_cf_node_block) {
 725       nir_instr_insert_before_block(nir_cf_node_as_block(node), before);
 726    } else {
 727       nir_cf_node *prev = nir_cf_node_prev(node);
 728       assert(prev->type == nir_cf_node_block);
 729       nir_block *prev_block = nir_cf_node_as_block(prev);
 730
 731       nir_instr_insert_before_block(prev_block, before);
 732    }
 733 }
 734
 735 void
 736 nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after)
 737 {
 738    if (node->type == nir_cf_node_block) {
 739       nir_instr_insert_after_block(nir_cf_node_as_block(node), after);
 740    } else {
 741       nir_cf_node *next = nir_cf_node_next(node);
 742       assert(next->type == nir_cf_node_block);
 743       nir_block *next_block = nir_cf_node_as_block(next);
 744
 745       nir_instr_insert_before_block(next_block, after);
 746    }
 747 }
 748
 749 void
 750 nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before)
 751 {
 752    nir_cf_node *first_node = exec_node_data(nir_cf_node,
 753                                             exec_list_get_head(list), node);
 754    nir_instr_insert_before_cf(first_node, before);
 755 }
 756
 757 void
 758 nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after)
 759 {
 760    nir_cf_node *last_node = exec_node_data(nir_cf_node,
 761                                            exec_list_get_tail(list), node);
 762    nir_instr_insert_after_cf(last_node, after);
 763 }
 764
 765 static bool
 766 remove_use_cb(nir_src *src, void *state)
 767 {
 768    list_del(&src->use_link);
 769
 770    return true;
 771 }
 772
 773 static bool
 774 remove_def_cb(nir_dest *dest, void *state)
 775 {
 776    if (!dest->is_ssa)
 777       list_del(&dest->reg.def_link);
 778
 779    return true;
 780 }
 781
 782 static void
 783 remove_defs_uses(nir_instr *instr)
 784 {
 785    nir_foreach_dest(instr, remove_def_cb, instr);
 786    nir_foreach_src(instr, remove_use_cb, instr);
 787 }
 788
 789 void nir_instr_remove(nir_instr *instr)
 790 {
 791    remove_defs_uses(instr);
 792    exec_node_remove(&instr->node);
 793
 794    if (instr->type == nir_instr_type_jump) {
 795       nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
 796       nir_handle_remove_jump(instr->block, jump_instr->type);
 797    }
 798 }
 799
 800 /*@}*/
 801
 802 void
 803 nir_index_local_regs(nir_function_impl *impl)
 804 {
 805    unsigned index = 0;
 806    foreach_list_typed(nir_register, reg, node, &impl->registers) {
 807       reg->index = index++;
 808    }
 809    impl->reg_alloc = index;
 810 }
 811
 812 void
 813 nir_index_global_regs(nir_shader *shader)
 814 {
 815    unsigned index = 0;
 816    foreach_list_typed(nir_register, reg, node, &shader->registers) {
 817       reg->index = index++;
 818    }
 819    shader->reg_alloc = index;
 820 }
 821
 822 static bool
 823 visit_alu_dest(nir_alu_instr *instr, nir_foreach_dest_cb cb, void *state)
 824 {
 825    return cb(&instr->dest.dest, state);
 826 }
 827
 828 static bool
 829 visit_intrinsic_dest(nir_intrinsic_instr *instr, nir_foreach_dest_cb cb,
 830                      void *state)
 831 {
 832    if (nir_intrinsic_infos[instr->intrinsic].has_dest)
 833       return cb(&instr->dest, state);
 834
 835    return true;
 836 }
 837
 838 static bool
 839 visit_texture_dest(nir_tex_instr *instr, nir_foreach_dest_cb cb,
 840                    void *state)
 841 {
 842    return cb(&instr->dest, state);
 843 }
 844
 845 static bool
 846 visit_phi_dest(nir_phi_instr *instr, nir_foreach_dest_cb cb, void *state)
 847 {
 848    return cb(&instr->dest, state);
 849 }
 850
 851 static bool
 852 visit_parallel_copy_dest(nir_parallel_copy_instr *instr,
 853                          nir_foreach_dest_cb cb, void *state)
 854 {
 855    nir_foreach_parallel_copy_entry(instr, entry) {
 856       if (!cb(&entry->dest, state))
 857          return false;
 858    }
 859
 860    return true;
 861 }
 862
 863 bool
 864 nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state)
 865 {
 866    switch (instr->type) {
 867    case nir_instr_type_alu:
 868       return visit_alu_dest(nir_instr_as_alu(instr), cb, state);
 869    case nir_instr_type_intrinsic:
 870       return visit_intrinsic_dest(nir_instr_as_intrinsic(instr), cb, state);
 871    case nir_instr_type_tex:
 872       return visit_texture_dest(nir_instr_as_tex(instr), cb, state);
 873    case nir_instr_type_phi:
 874       return visit_phi_dest(nir_instr_as_phi(instr), cb, state);
 875    case nir_instr_type_parallel_copy:
 876       return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr),
 877                                       cb, state);
 878
 879    case nir_instr_type_load_const:
 880    case nir_instr_type_ssa_undef:
 881    case nir_instr_type_call:
 882    case nir_instr_type_jump:
 883       break;
 884
 885    default:
 886       unreachable("Invalid instruction type");
 887       break;
 888    }
 889
 890    return true;
 891 }
 892
 893 struct foreach_ssa_def_state {
 894    nir_foreach_ssa_def_cb cb;
 895    void *client_state;
 896 };
 897
 898 static inline bool
 899 nir_ssa_def_visitor(nir_dest *dest, void *void_state)
 900 {
 901    struct foreach_ssa_def_state *state = void_state;
 902
 903    if (dest->is_ssa)
 904       return state->cb(&dest->ssa, state->client_state);
 905    else
 906       return true;
 907 }
 908
 909 bool
 910 nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb, void *state)
 911 {
 912    switch (instr->type) {
 913    case nir_instr_type_alu:
 914    case nir_instr_type_tex:
 915    case nir_instr_type_intrinsic:
 916    case nir_instr_type_phi:
 917    case nir_instr_type_parallel_copy: {
 918       struct foreach_ssa_def_state foreach_state = {cb, state};
 919       return nir_foreach_dest(instr, nir_ssa_def_visitor, &foreach_state);
 920    }
 921
 922    case nir_instr_type_load_const:
 923       return cb(&nir_instr_as_load_const(instr)->def, state);
 924    case nir_instr_type_ssa_undef:
 925       return cb(&nir_instr_as_ssa_undef(instr)->def, state);
 926    case nir_instr_type_call:
 927    case nir_instr_type_jump:
 928       return true;
 929    default:
 930       unreachable("Invalid instruction type");
 931    }
 932 }
 933
 934 static bool
 935 visit_src(nir_src *src, nir_foreach_src_cb cb, void *state)
 936 {
 937    if (!cb(src, state))
 938       return false;
 939    if (!src->is_ssa && src->reg.indirect)
 940       return cb(src->reg.indirect, state);
 941    return true;
 942 }
 943
 944 static bool
 945 visit_deref_array_src(nir_deref_array *deref, nir_foreach_src_cb cb,
 946                       void *state)
 947 {
 948    if (deref->deref_array_type == nir_deref_array_type_indirect)
 949       return visit_src(&deref->indirect, cb, state);
 950    return true;
 951 }
 952
 953 static bool
 954 visit_deref_src(nir_deref_var *deref, nir_foreach_src_cb cb, void *state)
 955 {
 956    nir_deref *cur = &deref->deref;
 957    while (cur != NULL) {
 958       if (cur->deref_type == nir_deref_type_array)
 959          if (!visit_deref_array_src(nir_deref_as_array(cur), cb, state))
 960             return false;
 961
 962       cur = cur->child;
 963    }
 964
 965    return true;
 966 }
 967
 968 static bool
 969 visit_alu_src(nir_alu_instr *instr, nir_foreach_src_cb cb, void *state)
 970 {
 971    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
 972       if (!visit_src(&instr->src[i].src, cb, state))
 973          return false;
 974
 975    return true;
 976 }
 977
 978 static bool
 979 visit_tex_src(nir_tex_instr *instr, nir_foreach_src_cb cb, void *state)
 980 {
 981    for (unsigned i = 0; i < instr->num_srcs; i++)
 982       if (!visit_src(&instr->src[i].src, cb, state))
 983          return false;
 984
 985    if (instr->sampler != NULL)
 986       if (!visit_deref_src(instr->sampler, cb, state))
 987          return false;
 988
 989    return true;
 990 }
 991
 992 static bool
 993 visit_intrinsic_src(nir_intrinsic_instr *instr, nir_foreach_src_cb cb,
 994                     void *state)
 995 {
 996    unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
 997    for (unsigned i = 0; i < num_srcs; i++)
 998       if (!visit_src(&instr->src[i], cb, state))
 999          return false;
1000
1001    unsigned num_vars =
1002       nir_intrinsic_infos[instr->intrinsic].num_variables;
1003    for (unsigned i = 0; i < num_vars; i++)
1004       if (!visit_deref_src(instr->variables[i], cb, state))
1005          return false;
1006
1007    return true;
1008 }
1009
1010 static bool
1011 visit_call_src(nir_call_instr *instr, nir_foreach_src_cb cb, void *state)
1012 {
1013    return true;
1014 }
1015
1016 static bool
1017 visit_load_const_src(nir_load_const_instr *instr, nir_foreach_src_cb cb,
1018                      void *state)
1019 {
1020    return true;
1021 }
1022
1023 static bool
1024 visit_phi_src(nir_phi_instr *instr, nir_foreach_src_cb cb, void *state)
1025 {
1026    nir_foreach_phi_src(instr, src) {
1027       if (!visit_src(&src->src, cb, state))
1028          return false;
1029    }
1030
1031    return true;
1032 }
1033
1034 static bool
1035 visit_parallel_copy_src(nir_parallel_copy_instr *instr,
1036                         nir_foreach_src_cb cb, void *state)
1037 {
1038    nir_foreach_parallel_copy_entry(instr, entry) {
1039       if (!visit_src(&entry->src, cb, state))
1040          return false;
1041    }
1042
1043    return true;
1044 }
1045
1046 typedef struct {
1047    void *state;
1048    nir_foreach_src_cb cb;
1049 } visit_dest_indirect_state;
1050
1051 static bool
1052 visit_dest_indirect(nir_dest *dest, void *_state)
1053 {
1054    visit_dest_indirect_state *state = (visit_dest_indirect_state *) _state;
1055
1056    if (!dest->is_ssa && dest->reg.indirect)
1057       return state->cb(dest->reg.indirect, state->state);
1058
1059    return true;
1060 }
1061
1062 bool
1063 nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state)
1064 {
1065    switch (instr->type) {
1066    case nir_instr_type_alu:
1067       if (!visit_alu_src(nir_instr_as_alu(instr), cb, state))
1068          return false;
1069       break;
1070    case nir_instr_type_intrinsic:
1071       if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr), cb, state))
1072          return false;
1073       break;
1074    case nir_instr_type_tex:
1075       if (!visit_tex_src(nir_instr_as_tex(instr), cb, state))
1076          return false;
1077       break;
1078    case nir_instr_type_call:
1079       if (!visit_call_src(nir_instr_as_call(instr), cb, state))
1080          return false;
1081       break;
1082    case nir_instr_type_load_const:
1083       if (!visit_load_const_src(nir_instr_as_load_const(instr), cb, state))
1084          return false;
1085       break;
1086    case nir_instr_type_phi:
1087       if (!visit_phi_src(nir_instr_as_phi(instr), cb, state))
1088          return false;
1089       break;
1090    case nir_instr_type_parallel_copy:
1091       if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr),
1092                                    cb, state))
1093          return false;
1094       break;
1095    case nir_instr_type_jump:
1096    case nir_instr_type_ssa_undef:
1097       return true;
1098
1099    default:
1100       unreachable("Invalid instruction type");
1101       break;
1102    }
1103
1104    visit_dest_indirect_state dest_state;
1105    dest_state.state = state;
1106    dest_state.cb = cb;
1107    return nir_foreach_dest(instr, visit_dest_indirect, &dest_state);
1108 }
1109
1110 nir_const_value *
1111 nir_src_as_const_value(nir_src src)
1112 {
1113    if (!src.is_ssa)
1114       return NULL;
1115
1116    if (src.ssa->parent_instr->type != nir_instr_type_load_const)
1117       return NULL;
1118
1119    nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
1120
1121    return &load->value;
1122 }
1123
1124 bool
1125 nir_srcs_equal(nir_src src1, nir_src src2)
1126 {
1127    if (src1.is_ssa) {
1128       if (src2.is_ssa) {
1129          return src1.ssa == src2.ssa;
1130       } else {
1131          return false;
1132       }
1133    } else {
1134       if (src2.is_ssa) {
1135          return false;
1136       } else {
1137          if ((src1.reg.indirect == NULL) != (src2.reg.indirect == NULL))
1138             return false;
1139
1140          if (src1.reg.indirect) {
1141             if (!nir_srcs_equal(*src1.reg.indirect, *src2.reg.indirect))
1142                return false;
1143          }
1144
1145          return src1.reg.reg == src2.reg.reg &&
1146                 src1.reg.base_offset == src2.reg.base_offset;
1147       }
1148    }
1149 }
1150
1151 static bool
1152 src_is_valid(const nir_src *src)
1153 {
1154    return src->is_ssa ? (src->ssa != NULL) : (src->reg.reg != NULL);
1155 }
1156
1157 static void
1158 src_remove_all_uses(nir_src *src)
1159 {
1160    for (; src; src = src->is_ssa ? NULL : src->reg.indirect) {
1161       if (!src_is_valid(src))
1162          continue;
1163
1164       list_del(&src->use_link);
1165    }
1166 }
1167
1168 static void
1169 src_add_all_uses(nir_src *src, nir_instr *parent_instr, nir_if *parent_if)
1170 {
1171    for (; src; src = src->is_ssa ? NULL : src->reg.indirect) {
1172       if (!src_is_valid(src))
1173          continue;
1174
1175       if (parent_instr) {
1176          src->parent_instr = parent_instr;
1177          if (src->is_ssa)
1178             list_addtail(&src->use_link, &src->ssa->uses);
1179          else
1180             list_addtail(&src->use_link, &src->reg.reg->uses);
1181       } else {
1182          assert(parent_if);
1183          src->parent_if = parent_if;
1184          if (src->is_ssa)
1185             list_addtail(&src->use_link, &src->ssa->if_uses);
1186          else
1187             list_addtail(&src->use_link, &src->reg.reg->if_uses);
1188       }
1189    }
1190 }
1191
1192 void
1193 nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src)
1194 {
1195    assert(!src_is_valid(src) || src->parent_instr == instr);
1196
1197    src_remove_all_uses(src);
1198    *src = new_src;
1199    src_add_all_uses(src, instr, NULL);
1200 }
1201
1202 void
1203 nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src)
1204 {
1205    assert(!src_is_valid(dest) || dest->parent_instr == dest_instr);
1206
1207    src_remove_all_uses(dest);
1208    src_remove_all_uses(src);
1209    *dest = *src;
1210    *src = NIR_SRC_INIT;
1211    src_add_all_uses(dest, dest_instr, NULL);
1212 }
1213
1214 void
1215 nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src)
1216 {
1217    nir_src *src = &if_stmt->condition;
1218    assert(!src_is_valid(src) || src->parent_if == if_stmt);
1219
1220    src_remove_all_uses(src);
1221    *src = new_src;
1222    src_add_all_uses(src, NULL, if_stmt);
1223 }
1224
1225 void
1226 nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def,
1227                  unsigned num_components, const char *name)
1228 {
1229    def->name = name;
1230    def->parent_instr = instr;
1231    list_inithead(&def->uses);
1232    list_inithead(&def->if_uses);
1233    def->num_components = num_components;
1234
1235    if (instr->block) {
1236       nir_function_impl *impl =
1237          nir_cf_node_get_function(&instr->block->cf_node);
1238
1239       def->index = impl->ssa_alloc++;
1240    } else {
1241       def->index = UINT_MAX;
1242    }
1243 }
1244
1245 void
1246 nir_ssa_dest_init(nir_instr *instr, nir_dest *dest,
1247                  unsigned num_components, const char *name)
1248 {
1249    dest->is_ssa = true;
1250    nir_ssa_def_init(instr, &dest->ssa, num_components, name);
1251 }
1252
1253 void
1254 nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src, void *mem_ctx)
1255 {
1256    assert(!new_src.is_ssa || def != new_src.ssa);
1257
1258    nir_foreach_use_safe(def, use_src) {
1259       nir_instr *src_parent_instr = use_src->parent_instr;
1260       list_del(&use_src->use_link);
1261       nir_src_copy(use_src, &new_src, mem_ctx);
1262       src_add_all_uses(use_src, src_parent_instr, NULL);
1263    }
1264
1265    nir_foreach_if_use_safe(def, use_src) {
1266       nir_if *src_parent_if = use_src->parent_if;
1267       list_del(&use_src->use_link);
1268       nir_src_copy(use_src, &new_src, mem_ctx);
1269       src_add_all_uses(use_src, NULL, src_parent_if);
1270    }
1271 }
1272
1273
1274 static bool foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb,
1275                             bool reverse, void *state);
1276
1277 static inline bool
1278 foreach_if(nir_if *if_stmt, nir_foreach_block_cb cb, bool reverse, void *state)
1279 {
1280    if (reverse) {
1281       foreach_list_typed_safe_reverse(nir_cf_node, node, node,
1282                                       &if_stmt->else_list) {
1283          if (!foreach_cf_node(node, cb, reverse, state))
1284             return false;
1285       }
1286
1287       foreach_list_typed_safe_reverse(nir_cf_node, node, node,
1288                                       &if_stmt->then_list) {
1289          if (!foreach_cf_node(node, cb, reverse, state))
1290             return false;
1291       }
1292    } else {
1293       foreach_list_typed_safe(nir_cf_node, node, node, &if_stmt->then_list) {
1294          if (!foreach_cf_node(node, cb, reverse, state))
1295             return false;
1296       }
1297
1298       foreach_list_typed_safe(nir_cf_node, node, node, &if_stmt->else_list) {
1299          if (!foreach_cf_node(node, cb, reverse, state))
1300             return false;
1301       }
1302    }
1303
1304    return true;
1305 }
1306
1307 static inline bool
1308 foreach_loop(nir_loop *loop, nir_foreach_block_cb cb, bool reverse, void *state)
1309 {
1310    if (reverse) {
1311       foreach_list_typed_safe_reverse(nir_cf_node, node, node, &loop->body) {
1312          if (!foreach_cf_node(node, cb, reverse, state))
1313             return false;
1314       }
1315    } else {
1316       foreach_list_typed_safe(nir_cf_node, node, node, &loop->body) {
1317          if (!foreach_cf_node(node, cb, reverse, state))
1318             return false;
1319       }
1320    }
1321
1322    return true;
1323 }
1324
1325 static bool
1326 foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb,
1327                 bool reverse, void *state)
1328 {
1329    switch (node->type) {
1330    case nir_cf_node_block:
1331       return cb(nir_cf_node_as_block(node), state);
1332    case nir_cf_node_if:
1333       return foreach_if(nir_cf_node_as_if(node), cb, reverse, state);
1334    case nir_cf_node_loop:
1335       return foreach_loop(nir_cf_node_as_loop(node), cb, reverse, state);
1336       break;
1337
1338    default:
1339       unreachable("Invalid CFG node type");
1340       break;
1341    }
1342
1343    return false;
1344 }
1345
1346 bool
1347 nir_foreach_block(nir_function_impl *impl, nir_foreach_block_cb cb, void *state)
1348 {
1349    foreach_list_typed_safe(nir_cf_node, node, node, &impl->body) {
1350       if (!foreach_cf_node(node, cb, false, state))
1351          return false;
1352    }
1353
1354    return cb(impl->end_block, state);
1355 }
1356
1357 bool
1358 nir_foreach_block_reverse(nir_function_impl *impl, nir_foreach_block_cb cb,
1359                           void *state)
1360 {
1361    if (!cb(impl->end_block, state))
1362       return false;
1363
1364    foreach_list_typed_safe_reverse(nir_cf_node, node, node, &impl->body) {
1365       if (!foreach_cf_node(node, cb, true, state))
1366          return false;
1367    }
1368
1369    return true;
1370 }
1371
1372 nir_if *
1373 nir_block_get_following_if(nir_block *block)
1374 {
1375    if (exec_node_is_tail_sentinel(&block->cf_node.node))
1376       return NULL;
1377
1378    if (nir_cf_node_is_last(&block->cf_node))
1379       return NULL;
1380
1381    nir_cf_node *next_node = nir_cf_node_next(&block->cf_node);
1382
1383    if (next_node->type != nir_cf_node_if)
1384       return NULL;
1385
1386    return nir_cf_node_as_if(next_node);
1387 }
1388
1389 static bool
1390 index_block(nir_block *block, void *state)
1391 {
1392    unsigned *index = state;
1393    block->index = (*index)++;
1394    return true;
1395 }
1396
1397 void
1398 nir_index_blocks(nir_function_impl *impl)
1399 {
1400    unsigned index = 0;
1401
1402    if (impl->valid_metadata & nir_metadata_block_index)
1403       return;
1404
1405    nir_foreach_block(impl, index_block, &index);
1406
1407    impl->num_blocks = index;
1408 }
1409
1410 static bool
1411 index_ssa_def_cb(nir_ssa_def *def, void *state)
1412 {
1413    unsigned *index = (unsigned *) state;
1414    def->index = (*index)++;
1415
1416    return true;
1417 }
1418
1419 static bool
1420 index_ssa_block(nir_block *block, void *state)
1421 {
1422    nir_foreach_instr(block, instr)
1423       nir_foreach_ssa_def(instr, index_ssa_def_cb, state);
1424
1425    return true;
1426 }
1427
1428 void
1429 nir_index_ssa_defs(nir_function_impl *impl)
1430 {
1431    unsigned index = 0;
1432    nir_foreach_block(impl, index_ssa_block, &index);
1433    impl->ssa_alloc = index;
1434 }