src/gallium/drivers/lima/ir/pp/nir.c

   1 /*
   2  * Copyright (c) 2017 Lima Project
   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, sub license,
   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
  12  * next paragraph) shall be included in all copies or substantial portions
  13  * of the 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 NON-INFRINGEMENT. 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
  21  * DEALINGS IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <string.h>
  26
  27 #include "util/hash_table.h"
  28 #include "util/ralloc.h"
  29 #include "util/bitscan.h"
  30 #include "compiler/nir/nir.h"
  31 #include "pipe/p_state.h"
  32
  33
  34 #include "ppir.h"
  35
  36 static void *ppir_node_create_ssa(ppir_block *block, ppir_op op, nir_ssa_def *ssa)
  37 {
  38    ppir_node *node = ppir_node_create(block, op, ssa->index, 0);
  39    if (!node)
  40       return NULL;
  41
  42    ppir_dest *dest = ppir_node_get_dest(node);
  43    dest->type = ppir_target_ssa;
  44    dest->ssa.num_components = ssa->num_components;
  45    dest->ssa.live_in = INT_MAX;
  46    dest->ssa.live_out = 0;
  47    dest->write_mask = u_bit_consecutive(0, ssa->num_components);
  48
  49    if (node->type == ppir_node_type_load ||
  50        node->type == ppir_node_type_store)
  51       dest->ssa.is_head = true;
  52
  53    return node;
  54 }
  55
  56 static void *ppir_node_create_reg(ppir_block *block, ppir_op op,
  57                                   nir_register *reg, unsigned mask)
  58 {
  59    ppir_node *node = ppir_node_create(block, op, reg->index, mask);
  60    if (!node)
  61       return NULL;
  62
  63    ppir_dest *dest = ppir_node_get_dest(node);
  64
  65    list_for_each_entry(ppir_reg, r, &block->comp->reg_list, list) {
  66       if (r->index == reg->index) {
  67          dest->reg = r;
  68          break;
  69       }
  70    }
  71
  72    dest->type = ppir_target_register;
  73    dest->write_mask = mask;
  74
  75    if (node->type == ppir_node_type_load ||
  76        node->type == ppir_node_type_store)
  77       dest->reg->is_head = true;
  78
  79    return node;
  80 }
  81
  82 static void *ppir_node_create_dest(ppir_block *block, ppir_op op,
  83                                    nir_dest *dest, unsigned mask)
  84 {
  85    unsigned index = -1;
  86
  87    if (dest) {
  88       if (dest->is_ssa)
  89          return ppir_node_create_ssa(block, op, &dest->ssa);
  90       else
  91          return ppir_node_create_reg(block, op, dest->reg.reg, mask);
  92    }
  93
  94    return ppir_node_create(block, op, index, 0);
  95 }
  96
  97 static void ppir_node_add_src(ppir_compiler *comp, ppir_node *node,
  98                               ppir_src *ps, nir_src *ns, unsigned mask)
  99 {
 100    ppir_node *child = NULL;
 101
 102    if (ns->is_ssa) {
 103       child = comp->var_nodes[ns->ssa->index];
 104       /* Clone consts for each successor */
 105       switch (child->op) {
 106       case ppir_op_const:
 107          child = ppir_node_clone(node->block, child);
 108          break;
 109       case ppir_op_load_varying: {
 110          bool is_load_coords = false;
 111          if (node->op == ppir_op_load_texture) {
 112             nir_tex_src *nts = (nir_tex_src *)ns;
 113             if (nts->src_type == nir_tex_src_coord)
 114                is_load_coords = true;
 115          }
 116
 117          if (!is_load_coords) {
 118             /* Clone varying loads for each block */
 119             if (child->block != node->block) {
 120                child = ppir_node_clone(node->block, child);
 121                comp->var_nodes[ns->ssa->index] = child;
 122             }
 123             break;
 124          }
 125          /* At least one successor is load_texture, promote it to load_coords
 126           * to ensure that is has exactly one successor */
 127          child->op = ppir_op_load_coords;
 128       }
 129          /* Fallthrough */
 130       case ppir_op_load_uniform:
 131       case ppir_op_load_coords:
 132       case ppir_op_load_coords_reg:
 133          /* Clone uniform and texture coord loads for each block.
 134           * Also ensure that each load has a single successor.
 135           * Let's do a fetch each time and hope for a cache hit instead
 136           * of increasing reg pressure.
 137           */
 138          if (child->block != node->block || !ppir_node_is_root(child)) {
 139             child = ppir_node_clone(node->block, child);
 140             comp->var_nodes[ns->ssa->index] = child;
 141          }
 142          break;
 143       default:
 144          break;
 145       }
 146
 147       if (child->op != ppir_op_undef)
 148          ppir_node_add_dep(node, child, ppir_dep_src);
 149    }
 150    else {
 151       nir_register *reg = ns->reg.reg;
 152       while (mask) {
 153          int swizzle = ps->swizzle[u_bit_scan(&mask)];
 154          child = comp->var_nodes[(reg->index << 2) + comp->reg_base + swizzle];
 155          /* Reg is read before it was written, create a dummy node for it */
 156          if (!child) {
 157             child = ppir_node_create_reg(node->block, ppir_op_undef, reg,
 158                u_bit_consecutive(0, 4));
 159             comp->var_nodes[(reg->index << 2) + comp->reg_base + swizzle] = child;
 160          }
 161          /* Don't add dummies or recursive deps for ops like r1 = r1 + ssa1 */
 162          if (child && node != child && child->op != ppir_op_undef)
 163             ppir_node_add_dep(node, child, ppir_dep_src);
 164       }
 165    }
 166
 167    ppir_node_target_assign(ps, child);
 168 }
 169
 170 static int nir_to_ppir_opcodes[nir_num_opcodes] = {
 171    /* not supported */
 172    [0 ... nir_last_opcode] = -1,
 173
 174    [nir_op_mov] = ppir_op_mov,
 175    [nir_op_fmul] = ppir_op_mul,
 176    [nir_op_fabs] = ppir_op_abs,
 177    [nir_op_fneg] = ppir_op_neg,
 178    [nir_op_fadd] = ppir_op_add,
 179    [nir_op_fsum3] = ppir_op_sum3,
 180    [nir_op_fsum4] = ppir_op_sum4,
 181    [nir_op_frsq] = ppir_op_rsqrt,
 182    [nir_op_flog2] = ppir_op_log2,
 183    [nir_op_fexp2] = ppir_op_exp2,
 184    [nir_op_fsqrt] = ppir_op_sqrt,
 185    [nir_op_fsin] = ppir_op_sin,
 186    [nir_op_fcos] = ppir_op_cos,
 187    [nir_op_fmax] = ppir_op_max,
 188    [nir_op_fmin] = ppir_op_min,
 189    [nir_op_frcp] = ppir_op_rcp,
 190    [nir_op_ffloor] = ppir_op_floor,
 191    [nir_op_fceil] = ppir_op_ceil,
 192    [nir_op_ffract] = ppir_op_fract,
 193    [nir_op_sge] = ppir_op_ge,
 194    [nir_op_slt] = ppir_op_lt,
 195    [nir_op_seq] = ppir_op_eq,
 196    [nir_op_sne] = ppir_op_ne,
 197    [nir_op_fcsel] = ppir_op_select,
 198    [nir_op_inot] = ppir_op_not,
 199    [nir_op_ftrunc] = ppir_op_trunc,
 200    [nir_op_fsat] = ppir_op_sat,
 201    [nir_op_fddx] = ppir_op_ddx,
 202    [nir_op_fddy] = ppir_op_ddy,
 203 };
 204
 205 static ppir_node *ppir_emit_alu(ppir_block *block, nir_instr *ni)
 206 {
 207    nir_alu_instr *instr = nir_instr_as_alu(ni);
 208    int op = nir_to_ppir_opcodes[instr->op];
 209
 210    if (op < 0) {
 211       ppir_error("unsupported nir_op: %s\n", nir_op_infos[instr->op].name);
 212       return NULL;
 213    }
 214
 215    ppir_alu_node *node = ppir_node_create_dest(block, op, &instr->dest.dest,
 216                                                instr->dest.write_mask);
 217    if (!node)
 218       return NULL;
 219
 220    ppir_dest *pd = &node->dest;
 221    nir_alu_dest *nd = &instr->dest;
 222    if (nd->saturate)
 223       pd->modifier = ppir_outmod_clamp_fraction;
 224
 225    unsigned src_mask;
 226    switch (op) {
 227    case ppir_op_sum3:
 228       src_mask = 0b0111;
 229       break;
 230    case ppir_op_sum4:
 231       src_mask = 0b1111;
 232       break;
 233    default:
 234       src_mask = pd->write_mask;
 235       break;
 236    }
 237
 238    unsigned num_child = nir_op_infos[instr->op].num_inputs;
 239    node->num_src = num_child;
 240
 241    for (int i = 0; i < num_child; i++) {
 242       nir_alu_src *ns = instr->src + i;
 243       ppir_src *ps = node->src + i;
 244       memcpy(ps->swizzle, ns->swizzle, sizeof(ps->swizzle));
 245       ppir_node_add_src(block->comp, &node->node, ps, &ns->src, src_mask);
 246
 247       ps->absolute = ns->abs;
 248       ps->negate = ns->negate;
 249    }
 250
 251    return &node->node;
 252 }
 253
 254 static ppir_block *ppir_block_create(ppir_compiler *comp);
 255
 256 static bool ppir_emit_discard_block(ppir_compiler *comp)
 257 {
 258    ppir_block *block = ppir_block_create(comp);
 259    ppir_discard_node *discard;
 260    if (!block)
 261       return false;
 262
 263    comp->discard_block = block;
 264    block->comp  = comp;
 265
 266    discard = ppir_node_create(block, ppir_op_discard, -1, 0);
 267    if (discard)
 268       list_addtail(&discard->node.list, &block->node_list);
 269    else
 270       return false;
 271
 272    return true;
 273 }
 274
 275 static ppir_node *ppir_emit_discard_if(ppir_block *block, nir_instr *ni)
 276 {
 277    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 278    ppir_node *node;
 279    ppir_compiler *comp = block->comp;
 280    ppir_branch_node *branch;
 281
 282    if (!comp->discard_block && !ppir_emit_discard_block(comp))
 283       return NULL;
 284
 285    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 286    if (!node)
 287       return NULL;
 288    branch = ppir_node_to_branch(node);
 289
 290    /* second src and condition will be updated during lowering */
 291    ppir_node_add_src(block->comp, node, &branch->src[0],
 292                      &instr->src[0], u_bit_consecutive(0, instr->num_components));
 293    branch->num_src = 1;
 294    branch->target = comp->discard_block;
 295
 296    return node;
 297 }
 298
 299 static ppir_node *ppir_emit_discard(ppir_block *block, nir_instr *ni)
 300 {
 301    ppir_node *node = ppir_node_create(block, ppir_op_discard, -1, 0);
 302
 303    return node;
 304 }
 305
 306 static ppir_node *ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
 307 {
 308    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 309    unsigned mask = 0;
 310    ppir_load_node *lnode;
 311    ppir_alu_node *alu_node;
 312
 313    switch (instr->intrinsic) {
 314    case nir_intrinsic_load_input:
 315       if (!instr->dest.is_ssa)
 316          mask = u_bit_consecutive(0, instr->num_components);
 317
 318       lnode = ppir_node_create_dest(block, ppir_op_load_varying, &instr->dest, mask);
 319       if (!lnode)
 320          return NULL;
 321
 322       lnode->num_components = instr->num_components;
 323       lnode->index = nir_intrinsic_base(instr) * 4 + nir_intrinsic_component(instr);
 324       if (nir_src_is_const(instr->src[0]))
 325          lnode->index += (uint32_t)(nir_src_as_float(instr->src[0]) * 4);
 326       else {
 327          lnode->num_src = 1;
 328          ppir_node_add_src(block->comp, &lnode->node, &lnode->src, instr->src, 1);
 329       }
 330       return &lnode->node;
 331
 332    case nir_intrinsic_load_frag_coord:
 333    case nir_intrinsic_load_point_coord:
 334    case nir_intrinsic_load_front_face:
 335       if (!instr->dest.is_ssa)
 336          mask = u_bit_consecutive(0, instr->num_components);
 337
 338       ppir_op op;
 339       switch (instr->intrinsic) {
 340       case nir_intrinsic_load_frag_coord:
 341          op = ppir_op_load_fragcoord;
 342          break;
 343       case nir_intrinsic_load_point_coord:
 344          op = ppir_op_load_pointcoord;
 345          break;
 346       case nir_intrinsic_load_front_face:
 347          op = ppir_op_load_frontface;
 348          break;
 349       default:
 350          assert(0);
 351          break;
 352       }
 353
 354       lnode = ppir_node_create_dest(block, op, &instr->dest, mask);
 355       if (!lnode)
 356          return NULL;
 357
 358       lnode->num_components = instr->num_components;
 359       return &lnode->node;
 360
 361    case nir_intrinsic_load_uniform:
 362       if (!instr->dest.is_ssa)
 363          mask = u_bit_consecutive(0, instr->num_components);
 364
 365       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 366       if (!lnode)
 367          return NULL;
 368
 369       lnode->num_components = instr->num_components;
 370       lnode->index = nir_intrinsic_base(instr);
 371       if (nir_src_is_const(instr->src[0]))
 372          lnode->index += (uint32_t)nir_src_as_float(instr->src[0]);
 373       else {
 374          lnode->num_src = 1;
 375          ppir_node_add_src(block->comp, &lnode->node, &lnode->src, instr->src, 1);
 376       }
 377
 378       return &lnode->node;
 379
 380    case nir_intrinsic_store_output: {
 381       alu_node = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 382       if (!alu_node)
 383          return NULL;
 384
 385       ppir_dest *dest = ppir_node_get_dest(&alu_node->node);
 386       dest->type = ppir_target_ssa;
 387       dest->ssa.num_components = instr->num_components;
 388       dest->ssa.live_in = INT_MAX;
 389       dest->ssa.live_out = 0;
 390       dest->ssa.index = 0;
 391       dest->write_mask = u_bit_consecutive(0, instr->num_components);
 392
 393       alu_node->num_src = 1;
 394
 395       for (int i = 0; i < instr->num_components; i++)
 396          alu_node->src[0].swizzle[i] = i;
 397
 398       ppir_node_add_src(block->comp, &alu_node->node, alu_node->src, instr->src,
 399                         u_bit_consecutive(0, instr->num_components));
 400
 401       return &alu_node->node;
 402    }
 403
 404    case nir_intrinsic_discard:
 405       return ppir_emit_discard(block, ni);
 406
 407    case nir_intrinsic_discard_if:
 408       return ppir_emit_discard_if(block, ni);
 409
 410    default:
 411       ppir_error("unsupported nir_intrinsic_instr %s\n",
 412                  nir_intrinsic_infos[instr->intrinsic].name);
 413       return NULL;
 414    }
 415 }
 416
 417 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 418 {
 419    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 420    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 421    if (!node)
 422       return NULL;
 423
 424    assert(instr->def.bit_size == 32);
 425
 426    for (int i = 0; i < instr->def.num_components; i++)
 427       node->constant.value[i].i = instr->value[i].i32;
 428    node->constant.num = instr->def.num_components;
 429
 430    return &node->node;
 431 }
 432
 433 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 434 {
 435    nir_ssa_undef_instr *undef = nir_instr_as_ssa_undef(ni);
 436    ppir_node *node = ppir_node_create_ssa(block, ppir_op_undef, &undef->def);
 437    if (!node)
 438       return NULL;
 439    ppir_alu_node *alu = ppir_node_to_alu(node);
 440
 441    ppir_dest *dest = &alu->dest;
 442    dest->ssa.undef = true;
 443    ppir_reg *ssa = &dest->ssa;
 444
 445    list_add(&ssa->list, &block->comp->reg_list);
 446
 447    return node;
 448 }
 449
 450 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 451 {
 452    nir_tex_instr *instr = nir_instr_as_tex(ni);
 453    ppir_load_texture_node *node;
 454
 455    switch (instr->op) {
 456    case nir_texop_tex:
 457    case nir_texop_txb:
 458    case nir_texop_txl:
 459       break;
 460    default:
 461       ppir_error("unsupported texop %d\n", instr->op);
 462       return NULL;
 463    }
 464
 465    unsigned mask = 0;
 466    if (!instr->dest.is_ssa)
 467       mask = u_bit_consecutive(0, nir_tex_instr_dest_size(instr));
 468
 469    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, mask);
 470    if (!node)
 471       return NULL;
 472
 473    node->sampler = instr->texture_index;
 474
 475    switch (instr->sampler_dim) {
 476    case GLSL_SAMPLER_DIM_2D:
 477    case GLSL_SAMPLER_DIM_CUBE:
 478    case GLSL_SAMPLER_DIM_RECT:
 479    case GLSL_SAMPLER_DIM_EXTERNAL:
 480       break;
 481    default:
 482       ppir_error("unsupported sampler dim: %d\n", instr->sampler_dim);
 483       return NULL;
 484    }
 485
 486    node->sampler_dim = instr->sampler_dim;
 487
 488    for (int i = 0; i < instr->coord_components; i++)
 489          node->src_coords.swizzle[i] = i;
 490
 491    for (int i = 0; i < instr->num_srcs; i++) {
 492       switch (instr->src[i].src_type) {
 493       case nir_tex_src_coord:
 494          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 495                            u_bit_consecutive(0, instr->coord_components));
 496          break;
 497       case nir_tex_src_bias:
 498       case nir_tex_src_lod:
 499          node->lod_bias_en = true;
 500          node->explicit_lod = (instr->src[i].src_type == nir_tex_src_lod);
 501          ppir_node_add_src(block->comp, &node->node, &node->lod_bias, &instr->src[i].src, 1);
 502          break;
 503       default:
 504          ppir_error("unsupported texture source type\n");
 505          assert(0);
 506          return NULL;
 507       }
 508    }
 509
 510    return &node->node;
 511 }
 512
 513 static ppir_block *ppir_get_block(ppir_compiler *comp, nir_block *nblock)
 514 {
 515    ppir_block *block = _mesa_hash_table_u64_search(comp->blocks, (uint64_t)nblock);
 516
 517    return block;
 518 }
 519
 520 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 521 {
 522    ppir_node *node;
 523    ppir_compiler *comp = block->comp;
 524    ppir_branch_node *branch;
 525    ppir_block *jump_block;
 526    nir_jump_instr *jump = nir_instr_as_jump(ni);
 527
 528    switch (jump->type) {
 529    case nir_jump_break: {
 530       assert(comp->current_block->successors[0]);
 531       assert(!comp->current_block->successors[1]);
 532       jump_block = comp->current_block->successors[0];
 533    }
 534    break;
 535    case nir_jump_continue:
 536       jump_block = comp->loop_cont_block;
 537    break;
 538    default:
 539       ppir_error("nir_jump_instr not support\n");
 540       return NULL;
 541    }
 542
 543    assert(jump_block != NULL);
 544
 545    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 546    if (!node)
 547       return NULL;
 548    branch = ppir_node_to_branch(node);
 549
 550    /* Unconditional */
 551    branch->num_src = 0;
 552    branch->target = jump_block;
 553
 554    return node;
 555 }
 556
 557 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 558    [nir_instr_type_alu]        = ppir_emit_alu,
 559    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 560    [nir_instr_type_load_const] = ppir_emit_load_const,
 561    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 562    [nir_instr_type_tex]        = ppir_emit_tex,
 563    [nir_instr_type_jump]       = ppir_emit_jump,
 564 };
 565
 566 static ppir_block *ppir_block_create(ppir_compiler *comp)
 567 {
 568    ppir_block *block = rzalloc(comp, ppir_block);
 569    if (!block)
 570       return NULL;
 571
 572    list_inithead(&block->node_list);
 573    list_inithead(&block->instr_list);
 574
 575    block->comp = comp;
 576
 577    return block;
 578 }
 579
 580 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 581 {
 582    ppir_block *block = ppir_get_block(comp, nblock);
 583
 584    comp->current_block = block;
 585
 586    list_addtail(&block->list, &comp->block_list);
 587
 588    nir_foreach_instr(instr, nblock) {
 589       assert(instr->type < nir_instr_type_phi);
 590       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 591       if (!node)
 592          return false;
 593
 594       list_addtail(&node->list, &block->node_list);
 595    }
 596
 597    return true;
 598 }
 599
 600 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list);
 601
 602 static bool ppir_emit_if(ppir_compiler *comp, nir_if *if_stmt)
 603 {
 604    ppir_node *node;
 605    ppir_branch_node *else_branch, *after_branch;
 606    nir_block *nir_else_block = nir_if_first_else_block(if_stmt);
 607    bool empty_else_block =
 608       (nir_else_block == nir_if_last_else_block(if_stmt) &&
 609       exec_list_is_empty(&nir_else_block->instr_list));
 610    ppir_block *block = comp->current_block;
 611
 612    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 613    if (!node)
 614       return false;
 615    else_branch = ppir_node_to_branch(node);
 616    ppir_node_add_src(block->comp, node, &else_branch->src[0],
 617                      &if_stmt->condition, 1);
 618    else_branch->num_src = 1;
 619    /* Negate condition to minimize branching. We're generating following:
 620     * current_block: { ...; if (!statement) branch else_block; }
 621     * then_block: { ...; branch after_block; }
 622     * else_block: { ... }
 623     * after_block: { ... }
 624     *
 625     * or if else list is empty:
 626     * block: { if (!statement) branch else_block; }
 627     * then_block: { ... }
 628     * else_block: after_block: { ... }
 629     */
 630    else_branch->negate = true;
 631    list_addtail(&else_branch->node.list, &block->node_list);
 632
 633    ppir_emit_cf_list(comp, &if_stmt->then_list);
 634    if (empty_else_block) {
 635       nir_block *nblock = nir_if_last_else_block(if_stmt);
 636       assert(nblock->successors[0]);
 637       assert(!nblock->successors[1]);
 638       else_branch->target = ppir_get_block(comp, nblock->successors[0]);
 639       /* Add empty else block to the list */
 640       list_addtail(&block->successors[1]->list, &comp->block_list);
 641       return true;
 642    }
 643
 644    else_branch->target = ppir_get_block(comp, nir_if_first_else_block(if_stmt));
 645
 646    nir_block *last_then_block = nir_if_last_then_block(if_stmt);
 647    assert(last_then_block->successors[0]);
 648    assert(!last_then_block->successors[1]);
 649    block = ppir_get_block(comp, last_then_block);
 650    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 651    if (!node)
 652       return false;
 653    after_branch = ppir_node_to_branch(node);
 654    /* Unconditional */
 655    after_branch->num_src = 0;
 656    after_branch->target = ppir_get_block(comp, last_then_block->successors[0]);
 657    /* Target should be after_block, will fixup later */
 658    list_addtail(&after_branch->node.list, &block->node_list);
 659
 660    ppir_emit_cf_list(comp, &if_stmt->else_list);
 661
 662    return true;
 663 }
 664
 665 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 666 {
 667    ppir_block *save_loop_cont_block = comp->loop_cont_block;
 668    ppir_block *block;
 669    ppir_branch_node *loop_branch;
 670    nir_block *loop_last_block;
 671    ppir_node *node;
 672
 673    comp->loop_cont_block = ppir_get_block(comp, nir_loop_first_block(nloop));
 674
 675    ppir_emit_cf_list(comp, &nloop->body);
 676
 677    loop_last_block = nir_loop_last_block(nloop);
 678    block = ppir_get_block(comp, loop_last_block);
 679    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 680    if (!node)
 681       return false;
 682    loop_branch = ppir_node_to_branch(node);
 683    /* Unconditional */
 684    loop_branch->num_src = 0;
 685    loop_branch->target = comp->loop_cont_block;
 686    list_addtail(&loop_branch->node.list, &block->node_list);
 687
 688    comp->loop_cont_block = save_loop_cont_block;
 689
 690    comp->num_loops++;
 691
 692    return true;
 693 }
 694
 695 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 696 {
 697    ppir_error("function nir_cf_node not support\n");
 698    return false;
 699 }
 700
 701 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 702 {
 703    foreach_list_typed(nir_cf_node, node, node, list) {
 704       bool ret;
 705
 706       switch (node->type) {
 707       case nir_cf_node_block:
 708          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 709          break;
 710       case nir_cf_node_if:
 711          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 712          break;
 713       case nir_cf_node_loop:
 714          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 715          break;
 716       case nir_cf_node_function:
 717          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 718          break;
 719       default:
 720          ppir_error("unknown NIR node type %d\n", node->type);
 721          return false;
 722       }
 723
 724       if (!ret)
 725          return false;
 726    }
 727
 728    return true;
 729 }
 730
 731 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 732 {
 733    ppir_compiler *comp = rzalloc_size(
 734       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 735    if (!comp)
 736       return NULL;
 737
 738    list_inithead(&comp->block_list);
 739    list_inithead(&comp->reg_list);
 740    comp->blocks = _mesa_hash_table_u64_create(prog);
 741
 742    comp->var_nodes = (ppir_node **)(comp + 1);
 743    comp->reg_base = num_ssa;
 744    comp->prog = prog;
 745    return comp;
 746 }
 747
 748 static void ppir_add_ordering_deps(ppir_compiler *comp)
 749 {
 750    /* Some intrinsics do not have explicit dependencies and thus depend
 751     * on instructions order. Consider discard_if and store_ouput as
 752     * example. If we don't add fake dependency of discard_if to store_output
 753     * scheduler may put store_output first and since store_output terminates
 754     * shader on Utgard PP, rest of it will never be executed.
 755     * Add fake dependencies for discard/branch/store to preserve
 756     * instruction order.
 757     *
 758     * TODO: scheduler should schedule discard_if as early as possible otherwise
 759     * we may end up with suboptimal code for cases like this:
 760     *
 761     * s3 = s1 < s2
 762     * discard_if s3
 763     * s4 = s1 + s2
 764     * store s4
 765     *
 766     * In this case store depends on discard_if and s4, but since dependencies can
 767     * be scheduled in any order it can result in code like this:
 768     *
 769     * instr1: s3 = s1 < s3
 770     * instr2: s4 = s1 + s2
 771     * instr3: discard_if s3
 772     * instr4: store s4
 773     */
 774    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 775       ppir_node *prev_node = NULL;
 776       list_for_each_entry_rev(ppir_node, node, &block->node_list, list) {
 777          if (prev_node && ppir_node_is_root(node) && node->op != ppir_op_const) {
 778             ppir_node_add_dep(prev_node, node, ppir_dep_sequence);
 779          }
 780          if (node->op == ppir_op_discard ||
 781              node->op == ppir_op_store_color ||
 782              node->op == ppir_op_store_temp ||
 783              node->op == ppir_op_branch) {
 784             prev_node = node;
 785          }
 786       }
 787    }
 788 }
 789
 790 static void ppir_print_shader_db(struct nir_shader *nir, ppir_compiler *comp,
 791                                  struct pipe_debug_callback *debug)
 792 {
 793    const struct shader_info *info = &nir->info;
 794    char *shaderdb;
 795    int ret = asprintf(&shaderdb,
 796                       "%s shader: %d inst, %d loops, %d:%d spills:fills\n",
 797                       gl_shader_stage_name(info->stage),
 798                       comp->cur_instr_index,
 799                       comp->num_loops,
 800                       comp->num_spills,
 801                       comp->num_fills);
 802    assert(ret >= 0);
 803
 804    if (lima_debug & LIMA_DEBUG_SHADERDB)
 805       fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
 806
 807    pipe_debug_message(debug, SHADER_INFO, "%s", shaderdb);
 808    free(shaderdb);
 809 }
 810
 811 static void ppir_add_write_after_read_deps(ppir_compiler *comp)
 812 {
 813    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 814       list_for_each_entry(ppir_reg, reg, &comp->reg_list, list) {
 815          ppir_node *write = NULL;
 816          list_for_each_entry_rev(ppir_node, node, &block->node_list, list) {
 817             for (int i = 0; i < ppir_node_get_src_num(node); i++) {
 818                ppir_src *src = ppir_node_get_src(node, i);
 819                if (src && src->type == ppir_target_register &&
 820                    src->reg == reg &&
 821                    write) {
 822                   ppir_debug("Adding dep %d for write %d\n", node->index, write->index);
 823                   ppir_node_add_dep(write, node, ppir_dep_write_after_read);
 824                }
 825             }
 826             ppir_dest *dest = ppir_node_get_dest(node);
 827             if (dest && dest->type == ppir_target_register &&
 828                 dest->reg == reg)
 829                write = node;
 830          }
 831       }
 832    }
 833 }
 834
 835 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 836                       struct ra_regs *ra,
 837                       struct pipe_debug_callback *debug)
 838 {
 839    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 840    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 841    if (!comp)
 842       return false;
 843
 844    comp->ra = ra;
 845
 846    /* 1st pass: create ppir blocks */
 847    nir_foreach_function(function, nir) {
 848       if (!function->impl)
 849          continue;
 850
 851       nir_foreach_block(nblock, function->impl) {
 852          ppir_block *block = ppir_block_create(comp);
 853          if (!block)
 854             return false;
 855          block->index = nblock->index;
 856          _mesa_hash_table_u64_insert(comp->blocks, (uint64_t)nblock, block);
 857       }
 858    }
 859
 860    /* 2nd pass: populate successors */
 861    nir_foreach_function(function, nir) {
 862       if (!function->impl)
 863          continue;
 864
 865       nir_foreach_block(nblock, function->impl) {
 866          ppir_block *block = ppir_get_block(comp, nblock);
 867          assert(block);
 868
 869          for (int i = 0; i < 2; i++) {
 870             if (nblock->successors[i])
 871                block->successors[i] = ppir_get_block(comp, nblock->successors[i]);
 872          }
 873       }
 874    }
 875
 876    /* Validate outputs, we support only gl_FragColor */
 877    nir_foreach_variable(var, &nir->outputs) {
 878       switch (var->data.location) {
 879       case FRAG_RESULT_COLOR:
 880       case FRAG_RESULT_DATA0:
 881          break;
 882       default:
 883          ppir_error("unsupported output type\n");
 884          goto err_out0;
 885          break;
 886       }
 887    }
 888
 889    foreach_list_typed(nir_register, reg, node, &func->registers) {
 890       ppir_reg *r = rzalloc(comp, ppir_reg);
 891       if (!r)
 892          return false;
 893
 894       r->index = reg->index;
 895       r->num_components = reg->num_components;
 896       r->live_in = INT_MAX;
 897       r->live_out = 0;
 898       r->is_head = false;
 899       list_addtail(&r->list, &comp->reg_list);
 900    }
 901
 902    if (!ppir_emit_cf_list(comp, &func->body))
 903       goto err_out0;
 904
 905    /* If we have discard block add it to the very end */
 906    if (comp->discard_block)
 907       list_addtail(&comp->discard_block->list, &comp->block_list);
 908
 909    ppir_node_print_prog(comp);
 910
 911    if (!ppir_lower_prog(comp))
 912       goto err_out0;
 913
 914    ppir_add_ordering_deps(comp);
 915    ppir_add_write_after_read_deps(comp);
 916
 917    ppir_node_print_prog(comp);
 918
 919    if (!ppir_node_to_instr(comp))
 920       goto err_out0;
 921
 922    if (!ppir_schedule_prog(comp))
 923       goto err_out0;
 924
 925    if (!ppir_regalloc_prog(comp))
 926       goto err_out0;
 927
 928    if (!ppir_codegen_prog(comp))
 929       goto err_out0;
 930
 931    ppir_print_shader_db(nir, comp, debug);
 932
 933    _mesa_hash_table_u64_destroy(comp->blocks, NULL);
 934    ralloc_free(comp);
 935    return true;
 936
 937 err_out0:
 938    _mesa_hash_table_u64_destroy(comp->blocks, NULL);
 939    ralloc_free(comp);
 940    return false;
 941 }
 942