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