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_fge] = ppir_op_ge,
 192    [nir_op_slt] = ppir_op_lt,
 193    [nir_op_flt] = ppir_op_lt,
 194    [nir_op_seq] = ppir_op_eq,
 195    [nir_op_feq] = ppir_op_eq,
 196    [nir_op_sne] = ppir_op_ne,
 197    [nir_op_fne] = ppir_op_ne,
 198    [nir_op_fcsel] = ppir_op_select,
 199    [nir_op_inot] = ppir_op_not,
 200    [nir_op_ftrunc] = ppir_op_trunc,
 201    [nir_op_fsat] = ppir_op_sat,
 202    [nir_op_fddx] = ppir_op_ddx,
 203    [nir_op_fddy] = ppir_op_ddy,
 204 };
 205
 206 static ppir_node *ppir_emit_alu(ppir_block *block, nir_instr *ni)
 207 {
 208    nir_alu_instr *instr = nir_instr_as_alu(ni);
 209    int op = nir_to_ppir_opcodes[instr->op];
 210
 211    if (op < 0) {
 212       ppir_error("unsupported nir_op: %s\n", nir_op_infos[instr->op].name);
 213       return NULL;
 214    }
 215
 216    ppir_alu_node *node = ppir_node_create_dest(block, op, &instr->dest.dest,
 217                                                instr->dest.write_mask);
 218    if (!node)
 219       return NULL;
 220
 221    ppir_dest *pd = &node->dest;
 222    nir_alu_dest *nd = &instr->dest;
 223    if (nd->saturate)
 224       pd->modifier = ppir_outmod_clamp_fraction;
 225
 226    unsigned src_mask;
 227    switch (op) {
 228    case ppir_op_sum3:
 229       src_mask = 0b0111;
 230       break;
 231    case ppir_op_sum4:
 232       src_mask = 0b1111;
 233       break;
 234    default:
 235       src_mask = pd->write_mask;
 236       break;
 237    }
 238
 239    unsigned num_child = nir_op_infos[instr->op].num_inputs;
 240    node->num_src = num_child;
 241
 242    for (int i = 0; i < num_child; i++) {
 243       nir_alu_src *ns = instr->src + i;
 244       ppir_src *ps = node->src + i;
 245       memcpy(ps->swizzle, ns->swizzle, sizeof(ps->swizzle));
 246       ppir_node_add_src(block->comp, &node->node, ps, &ns->src, src_mask);
 247
 248       ps->absolute = ns->abs;
 249       ps->negate = ns->negate;
 250    }
 251
 252    return &node->node;
 253 }
 254
 255 static ppir_block *ppir_block_create(ppir_compiler *comp);
 256
 257 static bool ppir_emit_discard_block(ppir_compiler *comp)
 258 {
 259    ppir_block *block = ppir_block_create(comp);
 260    ppir_discard_node *discard;
 261    if (!block)
 262       return false;
 263
 264    comp->discard_block = block;
 265    block->comp  = comp;
 266
 267    discard = ppir_node_create(block, ppir_op_discard, -1, 0);
 268    if (discard)
 269       list_addtail(&discard->node.list, &block->node_list);
 270    else
 271       return false;
 272
 273    return true;
 274 }
 275
 276 static ppir_node *ppir_emit_discard_if(ppir_block *block, nir_instr *ni)
 277 {
 278    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 279    ppir_node *node;
 280    ppir_compiler *comp = block->comp;
 281    ppir_branch_node *branch;
 282
 283    if (!comp->discard_block && !ppir_emit_discard_block(comp))
 284       return NULL;
 285
 286    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 287    if (!node)
 288       return NULL;
 289    branch = ppir_node_to_branch(node);
 290
 291    /* second src and condition will be updated during lowering */
 292    ppir_node_add_src(block->comp, node, &branch->src[0],
 293                      &instr->src[0], u_bit_consecutive(0, instr->num_components));
 294    branch->num_src = 1;
 295    branch->target = comp->discard_block;
 296
 297    return node;
 298 }
 299
 300 static ppir_node *ppir_emit_discard(ppir_block *block, nir_instr *ni)
 301 {
 302    ppir_node *node = ppir_node_create(block, ppir_op_discard, -1, 0);
 303
 304    return node;
 305 }
 306
 307 static ppir_node *ppir_emit_intrinsic(ppir_block *block, nir_instr *ni)
 308 {
 309    nir_intrinsic_instr *instr = nir_instr_as_intrinsic(ni);
 310    unsigned mask = 0;
 311    ppir_load_node *lnode;
 312    ppir_alu_node *alu_node;
 313
 314    switch (instr->intrinsic) {
 315    case nir_intrinsic_load_input:
 316       if (!instr->dest.is_ssa)
 317          mask = u_bit_consecutive(0, instr->num_components);
 318
 319       lnode = ppir_node_create_dest(block, ppir_op_load_varying, &instr->dest, mask);
 320       if (!lnode)
 321          return NULL;
 322
 323       lnode->num_components = instr->num_components;
 324       lnode->index = nir_intrinsic_base(instr) * 4 + nir_intrinsic_component(instr);
 325       return &lnode->node;
 326
 327    case nir_intrinsic_load_frag_coord:
 328    case nir_intrinsic_load_point_coord:
 329    case nir_intrinsic_load_front_face:
 330       if (!instr->dest.is_ssa)
 331          mask = u_bit_consecutive(0, instr->num_components);
 332
 333       ppir_op op;
 334       switch (instr->intrinsic) {
 335       case nir_intrinsic_load_frag_coord:
 336          op = ppir_op_load_fragcoord;
 337          break;
 338       case nir_intrinsic_load_point_coord:
 339          op = ppir_op_load_pointcoord;
 340          break;
 341       case nir_intrinsic_load_front_face:
 342          op = ppir_op_load_frontface;
 343          break;
 344       default:
 345          assert(0);
 346          break;
 347       }
 348
 349       lnode = ppir_node_create_dest(block, op, &instr->dest, mask);
 350       if (!lnode)
 351          return NULL;
 352
 353       lnode->num_components = instr->num_components;
 354       return &lnode->node;
 355
 356    case nir_intrinsic_load_uniform:
 357       if (!instr->dest.is_ssa)
 358          mask = u_bit_consecutive(0, instr->num_components);
 359
 360       lnode = ppir_node_create_dest(block, ppir_op_load_uniform, &instr->dest, mask);
 361       if (!lnode)
 362          return NULL;
 363
 364       lnode->num_components = instr->num_components;
 365       lnode->index = nir_intrinsic_base(instr);
 366       lnode->index += (uint32_t)nir_src_as_float(instr->src[0]);
 367
 368       return &lnode->node;
 369
 370    case nir_intrinsic_store_output: {
 371       alu_node = ppir_node_create_dest(block, ppir_op_store_color, NULL, 0);
 372       if (!alu_node)
 373          return NULL;
 374
 375       ppir_dest *dest = ppir_node_get_dest(&alu_node->node);
 376       dest->type = ppir_target_ssa;
 377       dest->ssa.num_components = instr->num_components;
 378       dest->ssa.live_in = INT_MAX;
 379       dest->ssa.live_out = 0;
 380       dest->ssa.index = 0;
 381       dest->write_mask = u_bit_consecutive(0, instr->num_components);
 382
 383       alu_node->num_src = 1;
 384
 385       for (int i = 0; i < instr->num_components; i++)
 386          alu_node->src[0].swizzle[i] = i;
 387
 388       ppir_node_add_src(block->comp, &alu_node->node, alu_node->src, instr->src,
 389                         u_bit_consecutive(0, instr->num_components));
 390
 391       return &alu_node->node;
 392    }
 393
 394    case nir_intrinsic_discard:
 395       return ppir_emit_discard(block, ni);
 396
 397    case nir_intrinsic_discard_if:
 398       return ppir_emit_discard_if(block, ni);
 399
 400    default:
 401       ppir_error("unsupported nir_intrinsic_instr %s\n",
 402                  nir_intrinsic_infos[instr->intrinsic].name);
 403       return NULL;
 404    }
 405 }
 406
 407 static ppir_node *ppir_emit_load_const(ppir_block *block, nir_instr *ni)
 408 {
 409    nir_load_const_instr *instr = nir_instr_as_load_const(ni);
 410    ppir_const_node *node = ppir_node_create_ssa(block, ppir_op_const, &instr->def);
 411    if (!node)
 412       return NULL;
 413
 414    assert(instr->def.bit_size == 32);
 415
 416    for (int i = 0; i < instr->def.num_components; i++)
 417       node->constant.value[i].i = instr->value[i].i32;
 418    node->constant.num = instr->def.num_components;
 419
 420    return &node->node;
 421 }
 422
 423 static ppir_node *ppir_emit_ssa_undef(ppir_block *block, nir_instr *ni)
 424 {
 425    ppir_error("nir_ssa_undef_instr not support\n");
 426    return NULL;
 427 }
 428
 429 static ppir_node *ppir_emit_tex(ppir_block *block, nir_instr *ni)
 430 {
 431    nir_tex_instr *instr = nir_instr_as_tex(ni);
 432    ppir_load_texture_node *node;
 433
 434    if (instr->op != nir_texop_tex) {
 435       ppir_error("unsupported texop %d\n", instr->op);
 436       return NULL;
 437    }
 438
 439    unsigned mask = 0;
 440    if (!instr->dest.is_ssa)
 441       mask = u_bit_consecutive(0, nir_tex_instr_dest_size(instr));
 442
 443    node = ppir_node_create_dest(block, ppir_op_load_texture, &instr->dest, mask);
 444    if (!node)
 445       return NULL;
 446
 447    node->sampler = instr->texture_index;
 448
 449    switch (instr->sampler_dim) {
 450    case GLSL_SAMPLER_DIM_2D:
 451    case GLSL_SAMPLER_DIM_RECT:
 452    case GLSL_SAMPLER_DIM_EXTERNAL:
 453       break;
 454    default:
 455       ppir_error("unsupported sampler dim: %d\n", instr->sampler_dim);
 456       return NULL;
 457    }
 458
 459    node->sampler_dim = instr->sampler_dim;
 460
 461    for (int i = 0; i < instr->coord_components; i++)
 462          node->src_coords.swizzle[i] = i;
 463
 464    for (int i = 0; i < instr->num_srcs; i++) {
 465       switch (instr->src[i].src_type) {
 466       case nir_tex_src_coord:
 467          ppir_node_add_src(block->comp, &node->node, &node->src_coords, &instr->src[i].src,
 468                            u_bit_consecutive(0, instr->coord_components));
 469          break;
 470       default:
 471          ppir_error("unsupported texture source type\n");
 472          assert(0);
 473          return NULL;
 474       }
 475    }
 476
 477    return &node->node;
 478 }
 479
 480 static ppir_block *ppir_get_block(ppir_compiler *comp, nir_block *nblock)
 481 {
 482    ppir_block *block = _mesa_hash_table_u64_search(comp->blocks, (uint64_t)nblock);
 483
 484    return block;
 485 }
 486
 487 static ppir_node *ppir_emit_jump(ppir_block *block, nir_instr *ni)
 488 {
 489    ppir_node *node;
 490    ppir_compiler *comp = block->comp;
 491    ppir_branch_node *branch;
 492    ppir_block *jump_block;
 493    nir_jump_instr *jump = nir_instr_as_jump(ni);
 494
 495    switch (jump->type) {
 496    case nir_jump_break: {
 497       assert(comp->current_block->successors[0]);
 498       assert(!comp->current_block->successors[1]);
 499       jump_block = comp->current_block->successors[0];
 500    }
 501    break;
 502    case nir_jump_continue:
 503       jump_block = comp->loop_cont_block;
 504    break;
 505    default:
 506       ppir_error("nir_jump_instr not support\n");
 507       return NULL;
 508    }
 509
 510    assert(jump_block != NULL);
 511
 512    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 513    if (!node)
 514       return NULL;
 515    branch = ppir_node_to_branch(node);
 516
 517    /* Unconditional */
 518    branch->num_src = 0;
 519    branch->target = jump_block;
 520
 521    return node;
 522 }
 523
 524 static ppir_node *(*ppir_emit_instr[nir_instr_type_phi])(ppir_block *, nir_instr *) = {
 525    [nir_instr_type_alu]        = ppir_emit_alu,
 526    [nir_instr_type_intrinsic]  = ppir_emit_intrinsic,
 527    [nir_instr_type_load_const] = ppir_emit_load_const,
 528    [nir_instr_type_ssa_undef]  = ppir_emit_ssa_undef,
 529    [nir_instr_type_tex]        = ppir_emit_tex,
 530    [nir_instr_type_jump]       = ppir_emit_jump,
 531 };
 532
 533 static ppir_block *ppir_block_create(ppir_compiler *comp)
 534 {
 535    ppir_block *block = rzalloc(comp, ppir_block);
 536    if (!block)
 537       return NULL;
 538
 539    list_inithead(&block->node_list);
 540    list_inithead(&block->instr_list);
 541
 542    block->comp = comp;
 543
 544    return block;
 545 }
 546
 547 static bool ppir_emit_block(ppir_compiler *comp, nir_block *nblock)
 548 {
 549    ppir_block *block = ppir_get_block(comp, nblock);
 550
 551    comp->current_block = block;
 552
 553    list_addtail(&block->list, &comp->block_list);
 554
 555    nir_foreach_instr(instr, nblock) {
 556       assert(instr->type < nir_instr_type_phi);
 557       ppir_node *node = ppir_emit_instr[instr->type](block, instr);
 558       if (!node)
 559          return false;
 560
 561       list_addtail(&node->list, &block->node_list);
 562    }
 563
 564    return true;
 565 }
 566
 567 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list);
 568
 569 static bool ppir_emit_if(ppir_compiler *comp, nir_if *if_stmt)
 570 {
 571    ppir_node *node;
 572    ppir_branch_node *else_branch, *after_branch;
 573    nir_block *nir_else_block = nir_if_first_else_block(if_stmt);
 574    bool empty_else_block =
 575       (nir_else_block == nir_if_last_else_block(if_stmt) &&
 576       exec_list_is_empty(&nir_else_block->instr_list));
 577    ppir_block *block = comp->current_block;
 578
 579    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 580    if (!node)
 581       return false;
 582    else_branch = ppir_node_to_branch(node);
 583    ppir_node_add_src(block->comp, node, &else_branch->src[0],
 584                      &if_stmt->condition, 1);
 585    else_branch->num_src = 1;
 586    /* Negate condition to minimize branching. We're generating following:
 587     * current_block: { ...; if (!statement) branch else_block; }
 588     * then_block: { ...; branch after_block; }
 589     * else_block: { ... }
 590     * after_block: { ... }
 591     *
 592     * or if else list is empty:
 593     * block: { if (!statement) branch else_block; }
 594     * then_block: { ... }
 595     * else_block: after_block: { ... }
 596     */
 597    else_branch->negate = true;
 598    list_addtail(&else_branch->node.list, &block->node_list);
 599
 600    ppir_emit_cf_list(comp, &if_stmt->then_list);
 601    if (empty_else_block) {
 602       nir_block *nblock = nir_if_last_else_block(if_stmt);
 603       assert(nblock->successors[0]);
 604       assert(!nblock->successors[1]);
 605       else_branch->target = ppir_get_block(comp, nblock->successors[0]);
 606       /* Add empty else block to the list */
 607       list_addtail(&block->successors[1]->list, &comp->block_list);
 608       return true;
 609    }
 610
 611    else_branch->target = ppir_get_block(comp, nir_if_first_else_block(if_stmt));
 612
 613    nir_block *last_then_block = nir_if_last_then_block(if_stmt);
 614    assert(last_then_block->successors[0]);
 615    assert(!last_then_block->successors[1]);
 616    block = ppir_get_block(comp, last_then_block);
 617    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 618    if (!node)
 619       return false;
 620    after_branch = ppir_node_to_branch(node);
 621    /* Unconditional */
 622    after_branch->num_src = 0;
 623    after_branch->target = ppir_get_block(comp, last_then_block->successors[0]);
 624    /* Target should be after_block, will fixup later */
 625    list_addtail(&after_branch->node.list, &block->node_list);
 626
 627    ppir_emit_cf_list(comp, &if_stmt->else_list);
 628
 629    return true;
 630 }
 631
 632 static bool ppir_emit_loop(ppir_compiler *comp, nir_loop *nloop)
 633 {
 634    ppir_block *save_loop_cont_block = comp->loop_cont_block;
 635    ppir_block *block;
 636    ppir_branch_node *loop_branch;
 637    nir_block *loop_last_block;
 638    ppir_node *node;
 639
 640    comp->loop_cont_block = ppir_get_block(comp, nir_loop_first_block(nloop));
 641
 642    ppir_emit_cf_list(comp, &nloop->body);
 643
 644    loop_last_block = nir_loop_last_block(nloop);
 645    block = ppir_get_block(comp, loop_last_block);
 646    node = ppir_node_create(block, ppir_op_branch, -1, 0);
 647    if (!node)
 648       return false;
 649    loop_branch = ppir_node_to_branch(node);
 650    /* Unconditional */
 651    loop_branch->num_src = 0;
 652    loop_branch->target = comp->loop_cont_block;
 653    list_addtail(&loop_branch->node.list, &block->node_list);
 654
 655    comp->loop_cont_block = save_loop_cont_block;
 656
 657    comp->num_loops++;
 658
 659    return true;
 660 }
 661
 662 static bool ppir_emit_function(ppir_compiler *comp, nir_function_impl *nfunc)
 663 {
 664    ppir_error("function nir_cf_node not support\n");
 665    return false;
 666 }
 667
 668 static bool ppir_emit_cf_list(ppir_compiler *comp, struct exec_list *list)
 669 {
 670    foreach_list_typed(nir_cf_node, node, node, list) {
 671       bool ret;
 672
 673       switch (node->type) {
 674       case nir_cf_node_block:
 675          ret = ppir_emit_block(comp, nir_cf_node_as_block(node));
 676          break;
 677       case nir_cf_node_if:
 678          ret = ppir_emit_if(comp, nir_cf_node_as_if(node));
 679          break;
 680       case nir_cf_node_loop:
 681          ret = ppir_emit_loop(comp, nir_cf_node_as_loop(node));
 682          break;
 683       case nir_cf_node_function:
 684          ret = ppir_emit_function(comp, nir_cf_node_as_function(node));
 685          break;
 686       default:
 687          ppir_error("unknown NIR node type %d\n", node->type);
 688          return false;
 689       }
 690
 691       if (!ret)
 692          return false;
 693    }
 694
 695    return true;
 696 }
 697
 698 static ppir_compiler *ppir_compiler_create(void *prog, unsigned num_reg, unsigned num_ssa)
 699 {
 700    ppir_compiler *comp = rzalloc_size(
 701       prog, sizeof(*comp) + ((num_reg << 2) + num_ssa) * sizeof(ppir_node *));
 702    if (!comp)
 703       return NULL;
 704
 705    list_inithead(&comp->block_list);
 706    list_inithead(&comp->reg_list);
 707    comp->blocks = _mesa_hash_table_u64_create(prog);
 708
 709    comp->var_nodes = (ppir_node **)(comp + 1);
 710    comp->reg_base = num_ssa;
 711    comp->prog = prog;
 712    return comp;
 713 }
 714
 715 static void ppir_add_ordering_deps(ppir_compiler *comp)
 716 {
 717    /* Some intrinsics do not have explicit dependencies and thus depend
 718     * on instructions order. Consider discard_if and store_ouput as
 719     * example. If we don't add fake dependency of discard_if to store_output
 720     * scheduler may put store_output first and since store_output terminates
 721     * shader on Utgard PP, rest of it will never be executed.
 722     * Add fake dependencies for discard/branch/store to preserve
 723     * instruction order.
 724     *
 725     * TODO: scheduler should schedule discard_if as early as possible otherwise
 726     * we may end up with suboptimal code for cases like this:
 727     *
 728     * s3 = s1 < s2
 729     * discard_if s3
 730     * s4 = s1 + s2
 731     * store s4
 732     *
 733     * In this case store depends on discard_if and s4, but since dependencies can
 734     * be scheduled in any order it can result in code like this:
 735     *
 736     * instr1: s3 = s1 < s3
 737     * instr2: s4 = s1 + s2
 738     * instr3: discard_if s3
 739     * instr4: store s4
 740     */
 741    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 742       ppir_node *prev_node = NULL;
 743       list_for_each_entry_rev(ppir_node, node, &block->node_list, list) {
 744          if (prev_node && ppir_node_is_root(node) && node->op != ppir_op_const) {
 745             ppir_node_add_dep(prev_node, node);
 746          }
 747          if (node->op == ppir_op_discard ||
 748              node->op == ppir_op_store_color ||
 749              node->op == ppir_op_store_temp ||
 750              node->op == ppir_op_branch) {
 751             prev_node = node;
 752          }
 753       }
 754    }
 755 }
 756
 757 static void ppir_print_shader_db(struct nir_shader *nir, ppir_compiler *comp,
 758                                  struct pipe_debug_callback *debug)
 759 {
 760    const struct shader_info *info = &nir->info;
 761    char *shaderdb;
 762    int ret = asprintf(&shaderdb,
 763                       "%s shader: %d inst, %d loops, %d:%d spills:fills\n",
 764                       gl_shader_stage_name(info->stage),
 765                       comp->cur_instr_index,
 766                       comp->num_loops,
 767                       comp->num_spills,
 768                       comp->num_fills);
 769    assert(ret >= 0);
 770
 771    if (lima_debug & LIMA_DEBUG_SHADERDB)
 772       fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
 773
 774    pipe_debug_message(debug, SHADER_INFO, "%s", shaderdb);
 775    free(shaderdb);
 776 }
 777
 778 static void ppir_add_write_after_read_deps(ppir_compiler *comp)
 779 {
 780    list_for_each_entry(ppir_block, block, &comp->block_list, list) {
 781       list_for_each_entry(ppir_reg, reg, &comp->reg_list, list) {
 782          ppir_node *write = NULL;
 783          list_for_each_entry_rev(ppir_node, node, &block->node_list, list) {
 784             for (int i = 0; i < ppir_node_get_src_num(node); i++) {
 785                ppir_src *src = ppir_node_get_src(node, i);
 786                if (src && src->type == ppir_target_register &&
 787                    src->reg == reg &&
 788                    write)
 789                   ppir_node_add_dep(write, node);
 790             }
 791             ppir_dest *dest = ppir_node_get_dest(node);
 792             if (dest && dest->type == ppir_target_register &&
 793                 dest->reg == reg)
 794                write = node;
 795          }
 796       }
 797    }
 798 }
 799
 800 bool ppir_compile_nir(struct lima_fs_shader_state *prog, struct nir_shader *nir,
 801                       struct ra_regs *ra,
 802                       struct pipe_debug_callback *debug)
 803 {
 804    nir_function_impl *func = nir_shader_get_entrypoint(nir);
 805    ppir_compiler *comp = ppir_compiler_create(prog, func->reg_alloc, func->ssa_alloc);
 806    if (!comp)
 807       return false;
 808
 809    comp->ra = ra;
 810
 811    /* 1st pass: create ppir blocks */
 812    nir_foreach_function(function, nir) {
 813       if (!function->impl)
 814          continue;
 815
 816       nir_foreach_block(nblock, function->impl) {
 817          ppir_block *block = ppir_block_create(comp);
 818          if (!block)
 819             return false;
 820          block->index = nblock->index;
 821          _mesa_hash_table_u64_insert(comp->blocks, (uint64_t)nblock, block);
 822       }
 823    }
 824
 825    /* 2nd pass: populate successors */
 826    nir_foreach_function(function, nir) {
 827       if (!function->impl)
 828          continue;
 829
 830       nir_foreach_block(nblock, function->impl) {
 831          ppir_block *block = ppir_get_block(comp, nblock);
 832          assert(block);
 833
 834          for (int i = 0; i < 2; i++) {
 835             if (nblock->successors[i])
 836                block->successors[i] = ppir_get_block(comp, nblock->successors[i]);
 837          }
 838       }
 839    }
 840
 841    /* Validate outputs, we support only gl_FragColor */
 842    nir_foreach_variable(var, &nir->outputs) {
 843       switch (var->data.location) {
 844       case FRAG_RESULT_COLOR:
 845       case FRAG_RESULT_DATA0:
 846          break;
 847       default:
 848          ppir_error("unsupported output type\n");
 849          goto err_out0;
 850          break;
 851       }
 852    }
 853
 854    foreach_list_typed(nir_register, reg, node, &func->registers) {
 855       ppir_reg *r = rzalloc(comp, ppir_reg);
 856       if (!r)
 857          return false;
 858
 859       r->index = reg->index;
 860       r->num_components = reg->num_components;
 861       r->live_in = INT_MAX;
 862       r->live_out = 0;
 863       r->is_head = false;
 864       list_addtail(&r->list, &comp->reg_list);
 865    }
 866
 867    if (!ppir_emit_cf_list(comp, &func->body))
 868       goto err_out0;
 869
 870    /* If we have discard block add it to the very end */
 871    if (comp->discard_block)
 872       list_addtail(&comp->discard_block->list, &comp->block_list);
 873
 874    ppir_node_print_prog(comp);
 875
 876    if (!ppir_lower_prog(comp))
 877       goto err_out0;
 878
 879    ppir_add_ordering_deps(comp);
 880    ppir_add_write_after_read_deps(comp);
 881
 882    ppir_node_print_prog(comp);
 883
 884    if (!ppir_node_to_instr(comp))
 885       goto err_out0;
 886
 887    if (!ppir_schedule_prog(comp))
 888       goto err_out0;
 889
 890    if (!ppir_regalloc_prog(comp))
 891       goto err_out0;
 892
 893    if (!ppir_codegen_prog(comp))
 894       goto err_out0;
 895
 896    ppir_print_shader_db(nir, comp, debug);
 897
 898    _mesa_hash_table_u64_destroy(comp->blocks, NULL);
 899    ralloc_free(comp);
 900    return true;
 901
 902 err_out0:
 903    _mesa_hash_table_u64_destroy(comp->blocks, NULL);
 904    ralloc_free(comp);
 905    return false;
 906 }
 907