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