src/gallium/drivers/radeonsi/si_shader_nir.c

   1 /*
   2  * Copyright 2017 Advanced Micro Devices, Inc.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 #include "ac_nir_to_llvm.h"
  26 #include "compiler/nir/nir.h"
  27 #include "compiler/nir/nir_builder.h"
  28 #include "compiler/nir/nir_deref.h"
  29 #include "compiler/nir_types.h"
  30 #include "si_pipe.h"
  31 #include "si_shader_internal.h"
  32 #include "tgsi/tgsi_from_mesa.h"
  33
  34 static const nir_deref_instr *tex_get_texture_deref(nir_tex_instr *instr)
  35 {
  36    for (unsigned i = 0; i < instr->num_srcs; i++) {
  37       switch (instr->src[i].src_type) {
  38       case nir_tex_src_texture_deref:
  39          return nir_src_as_deref(instr->src[i].src);
  40       default:
  41          break;
  42       }
  43    }
  44
  45    return NULL;
  46 }
  47
  48 static void scan_io_usage(struct si_shader_info *info, nir_intrinsic_instr *intr,
  49                           bool is_input)
  50 {
  51    unsigned interp = INTERP_MODE_FLAT; /* load_input uses flat shading */
  52
  53    if (intr->intrinsic == nir_intrinsic_load_interpolated_input) {
  54       nir_intrinsic_instr *baryc = nir_instr_as_intrinsic(intr->src[0].ssa->parent_instr);
  55
  56       if (baryc) {
  57          if (nir_intrinsic_infos[baryc->intrinsic].index_map[NIR_INTRINSIC_INTERP_MODE] > 0)
  58             interp = nir_intrinsic_interp_mode(baryc);
  59          else
  60             unreachable("unknown barycentric intrinsic");
  61       } else {
  62          unreachable("unknown barycentric expression");
  63       }
  64    }
  65
  66    unsigned mask, bit_size;
  67    bool dual_slot, is_output_load;
  68
  69    if (nir_intrinsic_infos[intr->intrinsic].index_map[NIR_INTRINSIC_WRMASK] > 0) {
  70       mask = nir_intrinsic_write_mask(intr); /* store */
  71       bit_size = nir_src_bit_size(intr->src[0]);
  72       dual_slot = bit_size == 64 && nir_src_num_components(intr->src[0]) >= 3;
  73       is_output_load = false;
  74    } else {
  75       mask = nir_ssa_def_components_read(&intr->dest.ssa); /* load */
  76       bit_size = intr->dest.ssa.bit_size;
  77       dual_slot = bit_size == 64 && intr->dest.ssa.num_components >= 3;
  78       is_output_load = !is_input;
  79    }
  80
  81    /* Convert the 64-bit component mask to a 32-bit component mask. */
  82    if (bit_size == 64) {
  83       unsigned new_mask = 0;
  84       for (unsigned i = 0; i < 4; i++) {
  85          if (mask & (1 << i))
  86             new_mask |= 0x3 << (2 * i);
  87       }
  88       mask = new_mask;
  89    }
  90
  91    /* Convert the 16-bit component mask to a 32-bit component mask. */
  92    if (bit_size == 16) {
  93       unsigned new_mask = 0;
  94       for (unsigned i = 0; i < 4; i++) {
  95          if (mask & (1 << i))
  96             new_mask |= 0x1 << (i / 2);
  97       }
  98       mask = new_mask;
  99    }
 100
 101    mask <<= nir_intrinsic_component(intr);
 102
 103    nir_src offset = *nir_get_io_offset_src(intr);
 104    bool indirect = !nir_src_is_const(offset);
 105    if (!indirect)
 106       assert(nir_src_as_uint(offset) == 0);
 107
 108    unsigned semantic = 0;
 109    /* VS doesn't have semantics. */
 110    if (info->stage != MESA_SHADER_VERTEX || !is_input)
 111       semantic = nir_intrinsic_io_semantics(intr).location;
 112
 113    if (info->stage == MESA_SHADER_FRAGMENT && !is_input) {
 114       /* Never use FRAG_RESULT_COLOR directly. */
 115       if (semantic == FRAG_RESULT_COLOR) {
 116          semantic = FRAG_RESULT_DATA0;
 117          info->color0_writes_all_cbufs = true;
 118       }
 119       semantic += nir_intrinsic_io_semantics(intr).dual_source_blend_index;
 120    }
 121
 122    unsigned driver_location = nir_intrinsic_base(intr);
 123    unsigned num_slots = indirect ? nir_intrinsic_io_semantics(intr).num_slots : (1 + dual_slot);
 124
 125    if (is_input) {
 126       assert(driver_location + num_slots <= ARRAY_SIZE(info->input_usage_mask));
 127
 128       for (unsigned i = 0; i < num_slots; i++) {
 129          unsigned loc = driver_location + i;
 130          unsigned slot_mask = (dual_slot && i % 2 ? mask >> 4 : mask) & 0xf;
 131
 132          info->input_semantic[loc] = semantic + i;
 133          info->input_interpolate[loc] = interp;
 134
 135          if (slot_mask) {
 136             info->input_usage_mask[loc] |= slot_mask;
 137             info->num_inputs = MAX2(info->num_inputs, loc + 1);
 138
 139             if (semantic == VARYING_SLOT_PRIMITIVE_ID)
 140                info->uses_primid = true;
 141          }
 142       }
 143    } else {
 144       /* Outputs. */
 145       assert(driver_location + num_slots <= ARRAY_SIZE(info->output_usagemask));
 146       assert(semantic + num_slots < ARRAY_SIZE(info->output_semantic_to_slot));
 147
 148       for (unsigned i = 0; i < num_slots; i++) {
 149          unsigned loc = driver_location + i;
 150          unsigned slot_mask = (dual_slot && i % 2 ? mask >> 4 : mask) & 0xf;
 151
 152          info->output_semantic[loc] = semantic + i;
 153          info->output_semantic_to_slot[semantic + i] = loc;
 154
 155          if (is_output_load) {
 156             /* Output loads have only a few things that we need to track. */
 157             info->output_readmask[loc] |= slot_mask;
 158
 159             if (info->stage == MESA_SHADER_FRAGMENT &&
 160                 nir_intrinsic_io_semantics(intr).fb_fetch_output)
 161                info->uses_fbfetch = true;
 162          } else if (slot_mask) {
 163             /* Output stores. */
 164             if (info->stage == MESA_SHADER_GEOMETRY) {
 165                unsigned gs_streams = (uint32_t)nir_intrinsic_io_semantics(intr).gs_streams <<
 166                                      (nir_intrinsic_component(intr) * 2);
 167                unsigned new_mask = slot_mask & ~info->output_usagemask[loc];
 168
 169                for (unsigned i = 0; i < 4; i++) {
 170                   unsigned stream = (gs_streams >> (i * 2)) & 0x3;
 171
 172                   if (new_mask & (1 << i)) {
 173                      info->output_streams[loc] |= stream << (i * 2);
 174                      info->num_stream_output_components[stream]++;
 175                   }
 176                }
 177             }
 178
 179             info->output_usagemask[loc] |= slot_mask;
 180             info->num_outputs = MAX2(info->num_outputs, loc + 1);
 181
 182             if (info->stage == MESA_SHADER_FRAGMENT) {
 183                switch (semantic) {
 184                case FRAG_RESULT_DEPTH:
 185                   info->writes_z = true;
 186                   break;
 187                case FRAG_RESULT_STENCIL:
 188                   info->writes_stencil = true;
 189                   break;
 190                case FRAG_RESULT_SAMPLE_MASK:
 191                   info->writes_samplemask = true;
 192                   break;
 193                default:
 194                   if (semantic >= FRAG_RESULT_DATA0 && semantic <= FRAG_RESULT_DATA7) {
 195                      unsigned index = semantic - FRAG_RESULT_DATA0;
 196                      info->colors_written |= 1 << (index + i);
 197                   }
 198                   break;
 199                }
 200             } else {
 201                switch (semantic) {
 202                case VARYING_SLOT_PRIMITIVE_ID:
 203                   info->writes_primid = true;
 204                   break;
 205                case VARYING_SLOT_VIEWPORT:
 206                   info->writes_viewport_index = true;
 207                   break;
 208                case VARYING_SLOT_LAYER:
 209                   info->writes_layer = true;
 210                   break;
 211                case VARYING_SLOT_PSIZ:
 212                   info->writes_psize = true;
 213                   break;
 214                case VARYING_SLOT_CLIP_VERTEX:
 215                   info->writes_clipvertex = true;
 216                   break;
 217                case VARYING_SLOT_EDGE:
 218                   info->writes_edgeflag = true;
 219                   break;
 220                case VARYING_SLOT_POS:
 221                   info->writes_position = true;
 222                   break;
 223                }
 224             }
 225          }
 226       }
 227    }
 228 }
 229
 230 static void scan_instruction(const struct nir_shader *nir, struct si_shader_info *info,
 231                              nir_instr *instr)
 232 {
 233    if (instr->type == nir_instr_type_alu) {
 234       nir_alu_instr *alu = nir_instr_as_alu(instr);
 235
 236       switch (alu->op) {
 237       case nir_op_fddx:
 238       case nir_op_fddy:
 239       case nir_op_fddx_fine:
 240       case nir_op_fddy_fine:
 241       case nir_op_fddx_coarse:
 242       case nir_op_fddy_coarse:
 243          info->uses_derivatives = true;
 244          break;
 245       default:
 246          break;
 247       }
 248    } else if (instr->type == nir_instr_type_tex) {
 249       nir_tex_instr *tex = nir_instr_as_tex(instr);
 250       const nir_deref_instr *deref = tex_get_texture_deref(tex);
 251       nir_variable *var = deref ? nir_deref_instr_get_variable(deref) : NULL;
 252
 253       if (var) {
 254          if (deref->mode != nir_var_uniform || var->data.bindless)
 255             info->uses_bindless_samplers = true;
 256       }
 257
 258       switch (tex->op) {
 259       case nir_texop_tex:
 260       case nir_texop_txb:
 261       case nir_texop_lod:
 262          info->uses_derivatives = true;
 263          break;
 264       default:
 265          break;
 266       }
 267    } else if (instr->type == nir_instr_type_intrinsic) {
 268       nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 269
 270       switch (intr->intrinsic) {
 271       case nir_intrinsic_load_front_face:
 272          info->uses_frontface = 1;
 273          break;
 274       case nir_intrinsic_load_instance_id:
 275          info->uses_instanceid = 1;
 276          break;
 277       case nir_intrinsic_load_invocation_id:
 278          info->uses_invocationid = true;
 279          break;
 280       case nir_intrinsic_load_num_work_groups:
 281          info->uses_grid_size = true;
 282          break;
 283       case nir_intrinsic_load_local_invocation_index:
 284       case nir_intrinsic_load_subgroup_id:
 285       case nir_intrinsic_load_num_subgroups:
 286          info->uses_subgroup_info = true;
 287          break;
 288       case nir_intrinsic_load_local_group_size:
 289          /* The block size is translated to IMM with a fixed block size. */
 290          if (info->base.cs.local_size[0] == 0)
 291             info->uses_block_size = true;
 292          break;
 293       case nir_intrinsic_load_local_invocation_id:
 294       case nir_intrinsic_load_work_group_id: {
 295          unsigned mask = nir_ssa_def_components_read(&intr->dest.ssa);
 296          while (mask) {
 297             unsigned i = u_bit_scan(&mask);
 298
 299             if (intr->intrinsic == nir_intrinsic_load_work_group_id)
 300                info->uses_block_id[i] = true;
 301             else
 302                info->uses_thread_id[i] = true;
 303          }
 304          break;
 305       }
 306       case nir_intrinsic_load_draw_id:
 307          info->uses_drawid = 1;
 308          break;
 309       case nir_intrinsic_load_primitive_id:
 310          info->uses_primid = 1;
 311          break;
 312       case nir_intrinsic_load_sample_mask_in:
 313          info->reads_samplemask = true;
 314          break;
 315       case nir_intrinsic_load_tess_level_inner:
 316       case nir_intrinsic_load_tess_level_outer:
 317          info->reads_tess_factors = true;
 318          break;
 319       case nir_intrinsic_bindless_image_load:
 320       case nir_intrinsic_bindless_image_size:
 321       case nir_intrinsic_bindless_image_samples:
 322          info->uses_bindless_images = true;
 323          break;
 324       case nir_intrinsic_bindless_image_store:
 325          info->uses_bindless_images = true;
 326          info->writes_memory = true;
 327          info->num_memory_instructions++; /* we only care about stores */
 328          break;
 329       case nir_intrinsic_image_deref_store:
 330          info->writes_memory = true;
 331          info->num_memory_instructions++; /* we only care about stores */
 332          break;
 333       case nir_intrinsic_bindless_image_atomic_add:
 334       case nir_intrinsic_bindless_image_atomic_imin:
 335       case nir_intrinsic_bindless_image_atomic_umin:
 336       case nir_intrinsic_bindless_image_atomic_imax:
 337       case nir_intrinsic_bindless_image_atomic_umax:
 338       case nir_intrinsic_bindless_image_atomic_and:
 339       case nir_intrinsic_bindless_image_atomic_or:
 340       case nir_intrinsic_bindless_image_atomic_xor:
 341       case nir_intrinsic_bindless_image_atomic_exchange:
 342       case nir_intrinsic_bindless_image_atomic_comp_swap:
 343          info->uses_bindless_images = true;
 344          info->writes_memory = true;
 345          info->num_memory_instructions++; /* we only care about stores */
 346          break;
 347       case nir_intrinsic_image_deref_atomic_add:
 348       case nir_intrinsic_image_deref_atomic_imin:
 349       case nir_intrinsic_image_deref_atomic_umin:
 350       case nir_intrinsic_image_deref_atomic_imax:
 351       case nir_intrinsic_image_deref_atomic_umax:
 352       case nir_intrinsic_image_deref_atomic_and:
 353       case nir_intrinsic_image_deref_atomic_or:
 354       case nir_intrinsic_image_deref_atomic_xor:
 355       case nir_intrinsic_image_deref_atomic_exchange:
 356       case nir_intrinsic_image_deref_atomic_comp_swap:
 357       case nir_intrinsic_image_deref_atomic_inc_wrap:
 358       case nir_intrinsic_image_deref_atomic_dec_wrap:
 359          info->writes_memory = true;
 360          info->num_memory_instructions++; /* we only care about stores */
 361          break;
 362       case nir_intrinsic_store_ssbo:
 363       case nir_intrinsic_ssbo_atomic_add:
 364       case nir_intrinsic_ssbo_atomic_imin:
 365       case nir_intrinsic_ssbo_atomic_umin:
 366       case nir_intrinsic_ssbo_atomic_imax:
 367       case nir_intrinsic_ssbo_atomic_umax:
 368       case nir_intrinsic_ssbo_atomic_and:
 369       case nir_intrinsic_ssbo_atomic_or:
 370       case nir_intrinsic_ssbo_atomic_xor:
 371       case nir_intrinsic_ssbo_atomic_exchange:
 372       case nir_intrinsic_ssbo_atomic_comp_swap:
 373          info->writes_memory = true;
 374          info->num_memory_instructions++; /* we only care about stores */
 375          break;
 376       case nir_intrinsic_load_color0:
 377       case nir_intrinsic_load_color1: {
 378          unsigned index = intr->intrinsic == nir_intrinsic_load_color1;
 379          uint8_t mask = nir_ssa_def_components_read(&intr->dest.ssa);
 380          info->colors_read |= mask << (index * 4);
 381          break;
 382       }
 383       case nir_intrinsic_load_barycentric_pixel:
 384       case nir_intrinsic_load_barycentric_centroid:
 385       case nir_intrinsic_load_barycentric_sample:
 386       case nir_intrinsic_load_barycentric_at_offset:   /* uses center */
 387       case nir_intrinsic_load_barycentric_at_sample: { /* uses center */
 388          unsigned mode = nir_intrinsic_interp_mode(intr);
 389
 390          if (mode == INTERP_MODE_FLAT)
 391             break;
 392
 393          if (mode == INTERP_MODE_NOPERSPECTIVE) {
 394             if (intr->intrinsic == nir_intrinsic_load_barycentric_sample)
 395                info->uses_linear_sample = true;
 396             else if (intr->intrinsic == nir_intrinsic_load_barycentric_centroid)
 397                info->uses_linear_centroid = true;
 398             else
 399                info->uses_linear_center = true;
 400
 401             if (intr->intrinsic == nir_intrinsic_load_barycentric_at_sample)
 402                info->uses_linear_opcode_interp_sample = true;
 403          } else {
 404             if (intr->intrinsic == nir_intrinsic_load_barycentric_sample)
 405                info->uses_persp_sample = true;
 406             else if (intr->intrinsic == nir_intrinsic_load_barycentric_centroid)
 407                info->uses_persp_centroid = true;
 408             else
 409                info->uses_persp_center = true;
 410
 411             if (intr->intrinsic == nir_intrinsic_load_barycentric_at_sample)
 412                info->uses_persp_opcode_interp_sample = true;
 413          }
 414          break;
 415       }
 416       case nir_intrinsic_load_input:
 417       case nir_intrinsic_load_per_vertex_input:
 418       case nir_intrinsic_load_input_vertex:
 419       case nir_intrinsic_load_interpolated_input:
 420          scan_io_usage(info, intr, true);
 421          break;
 422       case nir_intrinsic_load_output:
 423       case nir_intrinsic_load_per_vertex_output:
 424       case nir_intrinsic_store_output:
 425       case nir_intrinsic_store_per_vertex_output:
 426          scan_io_usage(info, intr, false);
 427          break;
 428       case nir_intrinsic_load_deref:
 429       case nir_intrinsic_store_deref:
 430       case nir_intrinsic_interp_deref_at_centroid:
 431       case nir_intrinsic_interp_deref_at_sample:
 432       case nir_intrinsic_interp_deref_at_offset:
 433          unreachable("these opcodes should have been lowered");
 434          break;
 435       default:
 436          break;
 437       }
 438    }
 439 }
 440
 441 void si_nir_scan_shader(const struct nir_shader *nir, struct si_shader_info *info)
 442 {
 443    nir_function *func;
 444
 445    info->base = nir->info;
 446    info->stage = nir->info.stage;
 447
 448    if (nir->info.stage == MESA_SHADER_TESS_EVAL) {
 449       if (info->base.tess.primitive_mode == GL_ISOLINES)
 450          info->base.tess.primitive_mode = GL_LINES;
 451    }
 452
 453    if (nir->info.stage == MESA_SHADER_FRAGMENT) {
 454       /* post_depth_coverage implies early_fragment_tests */
 455       info->base.fs.early_fragment_tests |= info->base.fs.post_depth_coverage;
 456
 457       info->color_interpolate[0] = nir->info.fs.color0_interp;
 458       info->color_interpolate[1] = nir->info.fs.color1_interp;
 459       for (unsigned i = 0; i < 2; i++) {
 460          if (info->color_interpolate[i] == INTERP_MODE_NONE)
 461             info->color_interpolate[i] = INTERP_MODE_COLOR;
 462       }
 463
 464       info->color_interpolate_loc[0] = nir->info.fs.color0_sample ? TGSI_INTERPOLATE_LOC_SAMPLE :
 465                                        nir->info.fs.color0_centroid ? TGSI_INTERPOLATE_LOC_CENTROID :
 466                                                                       TGSI_INTERPOLATE_LOC_CENTER;
 467       info->color_interpolate_loc[1] = nir->info.fs.color1_sample ? TGSI_INTERPOLATE_LOC_SAMPLE :
 468                                        nir->info.fs.color1_centroid ? TGSI_INTERPOLATE_LOC_CENTROID :
 469                                                                       TGSI_INTERPOLATE_LOC_CENTER;
 470    }
 471
 472    info->constbuf0_num_slots = nir->num_uniforms;
 473
 474    if (info->stage == MESA_SHADER_FRAGMENT)
 475       info->uses_kill = nir->info.fs.uses_discard;
 476
 477    if (nir->info.stage == MESA_SHADER_TESS_CTRL) {
 478       info->tessfactors_are_def_in_all_invocs = ac_are_tessfactors_def_in_all_invocs(nir);
 479    }
 480
 481    memset(info->output_semantic_to_slot, -1, sizeof(info->output_semantic_to_slot));
 482
 483    func = (struct nir_function *)exec_list_get_head_const(&nir->functions);
 484    nir_foreach_block (block, func->impl) {
 485       nir_foreach_instr (instr, block)
 486          scan_instruction(nir, info, instr);
 487    }
 488
 489    /* Add color inputs to the list of inputs. */
 490    if (nir->info.stage == MESA_SHADER_FRAGMENT) {
 491       for (unsigned i = 0; i < 2; i++) {
 492          if ((info->colors_read >> (i * 4)) & 0xf) {
 493             info->input_semantic[info->num_inputs] = VARYING_SLOT_COL0 + i;
 494             info->input_interpolate[info->num_inputs] = info->color_interpolate[i];
 495             info->input_usage_mask[info->num_inputs] = info->colors_read >> (i * 4);
 496             info->num_inputs++;
 497          }
 498       }
 499    }
 500
 501    /* Trim output read masks based on write masks. */
 502    for (unsigned i = 0; i < info->num_outputs; i++)
 503       info->output_readmask[i] &= info->output_usagemask[i];
 504 }
 505
 506 static void si_nir_opts(struct nir_shader *nir, bool first)
 507 {
 508    bool progress;
 509
 510    NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 511    NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL, NULL);
 512    NIR_PASS_V(nir, nir_lower_phis_to_scalar);
 513
 514    do {
 515       progress = false;
 516       bool lower_alu_to_scalar = false;
 517       bool lower_phis_to_scalar = false;
 518
 519       if (first) {
 520          bool opt_find_array_copies = false;
 521
 522          NIR_PASS(progress, nir, nir_split_array_vars, nir_var_function_temp);
 523          NIR_PASS(lower_alu_to_scalar, nir, nir_shrink_vec_array_vars, nir_var_function_temp);
 524          NIR_PASS(opt_find_array_copies, nir, nir_opt_find_array_copies);
 525          NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
 526
 527          /* Call nir_lower_var_copies() to remove any copies introduced
 528           * by nir_opt_find_array_copies().
 529           */
 530          if (opt_find_array_copies)
 531             NIR_PASS(progress, nir, nir_lower_var_copies);
 532          progress |= opt_find_array_copies;
 533       } else {
 534          NIR_PASS(progress, nir, nir_opt_copy_prop_vars);
 535       }
 536
 537       NIR_PASS(progress, nir, nir_opt_dead_write_vars);
 538
 539       NIR_PASS(lower_alu_to_scalar, nir, nir_opt_trivial_continues);
 540       /* (Constant) copy propagation is needed for txf with offsets. */
 541       NIR_PASS(progress, nir, nir_copy_prop);
 542       NIR_PASS(progress, nir, nir_opt_remove_phis);
 543       NIR_PASS(progress, nir, nir_opt_dce);
 544       NIR_PASS(lower_phis_to_scalar, nir, nir_opt_if, true);
 545       NIR_PASS(progress, nir, nir_opt_dead_cf);
 546
 547       if (lower_alu_to_scalar)
 548          NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL, NULL);
 549       if (lower_phis_to_scalar)
 550          NIR_PASS_V(nir, nir_lower_phis_to_scalar);
 551       progress |= lower_alu_to_scalar | lower_phis_to_scalar;
 552
 553       NIR_PASS(progress, nir, nir_opt_cse);
 554       NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true, true);
 555
 556       /* Needed for algebraic lowering */
 557       NIR_PASS(progress, nir, nir_opt_algebraic);
 558       NIR_PASS(progress, nir, nir_opt_constant_folding);
 559
 560       if (!nir->info.flrp_lowered) {
 561          unsigned lower_flrp = (nir->options->lower_flrp16 ? 16 : 0) |
 562                                (nir->options->lower_flrp32 ? 32 : 0) |
 563                                (nir->options->lower_flrp64 ? 64 : 0);
 564          assert(lower_flrp);
 565          bool lower_flrp_progress = false;
 566
 567          NIR_PASS(lower_flrp_progress, nir, nir_lower_flrp, lower_flrp, false /* always_precise */);
 568          if (lower_flrp_progress) {
 569             NIR_PASS(progress, nir, nir_opt_constant_folding);
 570             progress = true;
 571          }
 572
 573          /* Nothing should rematerialize any flrps, so we only
 574           * need to do this lowering once.
 575           */
 576          nir->info.flrp_lowered = true;
 577       }
 578
 579       NIR_PASS(progress, nir, nir_opt_undef);
 580       NIR_PASS(progress, nir, nir_opt_conditional_discard);
 581       if (nir->options->max_unroll_iterations) {
 582          NIR_PASS(progress, nir, nir_opt_loop_unroll, 0);
 583       }
 584    } while (progress);
 585 }
 586
 587 static int type_size_vec4(const struct glsl_type *type, bool bindless)
 588 {
 589    return glsl_count_attribute_slots(type, false);
 590 }
 591
 592 static void si_nir_lower_color(nir_shader *nir)
 593 {
 594    nir_function_impl *entrypoint = nir_shader_get_entrypoint(nir);
 595
 596    nir_builder b;
 597    nir_builder_init(&b, entrypoint);
 598
 599    nir_foreach_block (block, entrypoint) {
 600       nir_foreach_instr_safe (instr, block) {
 601          if (instr->type != nir_instr_type_intrinsic)
 602             continue;
 603
 604          nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 605
 606          if (intrin->intrinsic != nir_intrinsic_load_deref)
 607             continue;
 608
 609          nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
 610          if (deref->mode != nir_var_shader_in)
 611             continue;
 612
 613          b.cursor = nir_before_instr(instr);
 614          nir_variable *var = nir_deref_instr_get_variable(deref);
 615          nir_ssa_def *def;
 616
 617          if (var->data.location == VARYING_SLOT_COL0) {
 618             def = nir_load_color0(&b);
 619             nir->info.fs.color0_interp = var->data.interpolation;
 620             nir->info.fs.color0_sample = var->data.sample;
 621             nir->info.fs.color0_centroid = var->data.centroid;
 622          } else if (var->data.location == VARYING_SLOT_COL1) {
 623             def = nir_load_color1(&b);
 624             nir->info.fs.color1_interp = var->data.interpolation;
 625             nir->info.fs.color1_sample = var->data.sample;
 626             nir->info.fs.color1_centroid = var->data.centroid;
 627          } else {
 628             continue;
 629          }
 630
 631          nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(def));
 632          nir_instr_remove(instr);
 633       }
 634    }
 635 }
 636
 637 static void si_lower_io(struct nir_shader *nir)
 638 {
 639    /* HW supports indirect indexing for: | Enabled in driver
 640     * -------------------------------------------------------
 641     * VS inputs                          | No
 642     * TCS inputs                         | Yes
 643     * TES inputs                         | Yes
 644     * GS inputs                          | No
 645     * -------------------------------------------------------
 646     * VS outputs before TCS              | No
 647     * VS outputs before GS               | No
 648     * TCS outputs                        | Yes
 649     * TES outputs before GS              | No
 650     */
 651    bool has_indirect_inputs = nir->info.stage == MESA_SHADER_TESS_CTRL ||
 652                               nir->info.stage == MESA_SHADER_TESS_EVAL;
 653    bool has_indirect_outputs = nir->info.stage == MESA_SHADER_TESS_CTRL;
 654
 655    if (!has_indirect_inputs || !has_indirect_outputs) {
 656       NIR_PASS_V(nir, nir_lower_io_to_temporaries, nir_shader_get_entrypoint(nir),
 657                  !has_indirect_outputs, !has_indirect_inputs);
 658
 659       /* Since we're doing nir_lower_io_to_temporaries late, we need
 660        * to lower all the copy_deref's introduced by
 661        * lower_io_to_temporaries before calling nir_lower_io.
 662        */
 663       NIR_PASS_V(nir, nir_split_var_copies);
 664       NIR_PASS_V(nir, nir_lower_var_copies);
 665       NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 666    }
 667
 668    if (nir->info.stage == MESA_SHADER_FRAGMENT)
 669       si_nir_lower_color(nir);
 670
 671    NIR_PASS_V(nir, nir_lower_io, nir_var_shader_out | nir_var_shader_in,
 672               type_size_vec4, 0);
 673    nir->info.io_lowered = true;
 674
 675    /* This pass needs actual constants */
 676    NIR_PASS_V(nir, nir_opt_constant_folding);
 677    NIR_PASS_V(nir, nir_io_add_const_offset_to_base, nir_var_shader_in);
 678    NIR_PASS_V(nir, nir_io_add_const_offset_to_base, nir_var_shader_out);
 679
 680    /* Remove dead derefs, so that nir_validate doesn't fail. */
 681    NIR_PASS_V(nir, nir_opt_dce);
 682
 683    /* Remove input and output nir_variables, because we don't need them
 684     * anymore. Also remove uniforms, because those should have been lowered
 685     * to UBOs already.
 686     */
 687    unsigned modes = nir_var_shader_in | nir_var_shader_out | nir_var_uniform;
 688    nir_foreach_variable_with_modes_safe(var, nir, modes) {
 689       if (var->data.mode == nir_var_uniform &&
 690           (glsl_type_get_image_count(var->type) ||
 691            glsl_type_get_sampler_count(var->type)))
 692          continue;
 693
 694       exec_node_remove(&var->node);
 695    }
 696 }
 697
 698 /**
 699  * Perform "lowering" operations on the NIR that are run once when the shader
 700  * selector is created.
 701  */
 702 static void si_lower_nir(struct si_screen *sscreen, struct nir_shader *nir)
 703 {
 704    /* Perform lowerings (and optimizations) of code.
 705     *
 706     * Performance considerations aside, we must:
 707     * - lower certain ALU operations
 708     * - ensure constant offsets for texture instructions are folded
 709     *   and copy-propagated
 710     */
 711
 712    static const struct nir_lower_tex_options lower_tex_options = {
 713       .lower_txp = ~0u,
 714    };
 715    NIR_PASS_V(nir, nir_lower_tex, &lower_tex_options);
 716
 717    const nir_lower_subgroups_options subgroups_options = {
 718       .subgroup_size = 64,
 719       .ballot_bit_size = 64,
 720       .lower_to_scalar = true,
 721       .lower_subgroup_masks = true,
 722       .lower_vote_trivial = false,
 723       .lower_vote_eq_to_ballot = true,
 724    };
 725    NIR_PASS_V(nir, nir_lower_subgroups, &subgroups_options);
 726
 727    /* Lower load constants to scalar and then clean up the mess */
 728    NIR_PASS_V(nir, nir_lower_load_const_to_scalar);
 729    NIR_PASS_V(nir, nir_lower_var_copies);
 730    NIR_PASS_V(nir, nir_lower_pack);
 731    NIR_PASS_V(nir, nir_opt_access);
 732    si_nir_opts(nir, true);
 733
 734    /* Lower large variables that are always constant with load_constant
 735     * intrinsics, which get turned into PC-relative loads from a data
 736     * section next to the shader.
 737     *
 738     * st/mesa calls finalize_nir twice, but we can't call this pass twice.
 739     */
 740    bool changed = false;
 741    if (!nir->constant_data) {
 742       /* The pass crashes if there are dead temps of lowered IO interface types. */
 743       NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_function_temp, NULL);
 744       NIR_PASS(changed, nir, nir_opt_large_constants, glsl_get_natural_size_align_bytes, 16);
 745    }
 746
 747    changed |= ac_lower_indirect_derefs(nir, sscreen->info.chip_class);
 748    if (changed)
 749       si_nir_opts(nir, false);
 750
 751    NIR_PASS_V(nir, nir_lower_bool_to_int32);
 752    NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_function_temp, NULL);
 753
 754    if (sscreen->debug_flags & DBG(FS_CORRECT_DERIVS_AFTER_KILL))
 755       NIR_PASS_V(nir, nir_lower_discard_to_demote);
 756 }
 757
 758 void si_finalize_nir(struct pipe_screen *screen, void *nirptr, bool optimize)
 759 {
 760    struct si_screen *sscreen = (struct si_screen *)screen;
 761    struct nir_shader *nir = (struct nir_shader *)nirptr;
 762
 763    nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
 764    si_lower_io(nir);
 765    si_lower_nir(sscreen, nir);
 766 }