src/mesa/drivers/dri/i965/brw_nir.c

   1 /*
   2  * Copyright © 2014 Intel Corporation
   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, sublicense,
   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 next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * 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 NONINFRINGEMENT.  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 DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "brw_nir.h"
  25 #include "brw_shader.h"
  26 #include "compiler/glsl_types.h"
  27 #include "compiler/nir/nir_builder.h"
  28
  29 static bool
  30 is_input(nir_intrinsic_instr *intrin)
  31 {
  32    return intrin->intrinsic == nir_intrinsic_load_input ||
  33           intrin->intrinsic == nir_intrinsic_load_per_vertex_input ||
  34           intrin->intrinsic == nir_intrinsic_load_interpolated_input;
  35 }
  36
  37 static bool
  38 is_output(nir_intrinsic_instr *intrin)
  39 {
  40    return intrin->intrinsic == nir_intrinsic_load_output ||
  41           intrin->intrinsic == nir_intrinsic_load_per_vertex_output ||
  42           intrin->intrinsic == nir_intrinsic_store_output ||
  43           intrin->intrinsic == nir_intrinsic_store_per_vertex_output;
  44 }
  45
  46 /**
  47  * In many cases, we just add the base and offset together, so there's no
  48  * reason to keep them separate.  Sometimes, combining them is essential:
  49  * if a shader only accesses part of a compound variable (such as a matrix
  50  * or array), the variable's base may not actually exist in the VUE map.
  51  *
  52  * This pass adds constant offsets to instr->const_index[0], and resets
  53  * the offset source to 0.  Non-constant offsets remain unchanged - since
  54  * we don't know what part of a compound variable is accessed, we allocate
  55  * storage for the entire thing.
  56  */
  57
  58 static bool
  59 add_const_offset_to_base_block(nir_block *block, nir_builder *b,
  60                                nir_variable_mode mode)
  61 {
  62    nir_foreach_instr_safe(instr, block) {
  63       if (instr->type != nir_instr_type_intrinsic)
  64          continue;
  65
  66       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
  67
  68       if ((mode == nir_var_shader_in && is_input(intrin)) ||
  69           (mode == nir_var_shader_out && is_output(intrin))) {
  70          nir_src *offset = nir_get_io_offset_src(intrin);
  71          nir_const_value *const_offset = nir_src_as_const_value(*offset);
  72
  73          if (const_offset) {
  74             intrin->const_index[0] += const_offset->u32[0];
  75             b->cursor = nir_before_instr(&intrin->instr);
  76             nir_instr_rewrite_src(&intrin->instr, offset,
  77                                   nir_src_for_ssa(nir_imm_int(b, 0)));
  78          }
  79       }
  80    }
  81    return true;
  82 }
  83
  84 static void
  85 add_const_offset_to_base(nir_shader *nir, nir_variable_mode mode)
  86 {
  87    nir_foreach_function(f, nir) {
  88       if (f->impl) {
  89          nir_builder b;
  90          nir_builder_init(&b, f->impl);
  91          nir_foreach_block(block, f->impl) {
  92             add_const_offset_to_base_block(block, &b, mode);
  93          }
  94       }
  95    }
  96 }
  97
  98 static bool
  99 remap_vs_attrs(nir_block *block, shader_info *nir_info)
 100 {
 101    nir_foreach_instr(instr, block) {
 102       if (instr->type != nir_instr_type_intrinsic)
 103          continue;
 104
 105       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 106
 107       if (intrin->intrinsic == nir_intrinsic_load_input) {
 108          /* Attributes come in a contiguous block, ordered by their
 109           * gl_vert_attrib value.  That means we can compute the slot
 110           * number for an attribute by masking out the enabled attributes
 111           * before it and counting the bits.
 112           */
 113          int attr = intrin->const_index[0];
 114          int slot = _mesa_bitcount_64(nir_info->inputs_read &
 115                                       BITFIELD64_MASK(attr));
 116          intrin->const_index[0] = 4 * slot;
 117       }
 118    }
 119    return true;
 120 }
 121
 122 static bool
 123 remap_inputs_with_vue_map(nir_block *block, const struct brw_vue_map *vue_map)
 124 {
 125    nir_foreach_instr(instr, block) {
 126       if (instr->type != nir_instr_type_intrinsic)
 127          continue;
 128
 129       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 130
 131       if (intrin->intrinsic == nir_intrinsic_load_input ||
 132           intrin->intrinsic == nir_intrinsic_load_per_vertex_input) {
 133          int vue_slot = vue_map->varying_to_slot[intrin->const_index[0]];
 134          assert(vue_slot != -1);
 135          intrin->const_index[0] = vue_slot;
 136       }
 137    }
 138    return true;
 139 }
 140
 141 static bool
 142 remap_tess_levels(nir_builder *b, nir_intrinsic_instr *intr,
 143                   GLenum primitive_mode)
 144 {
 145    const int location = nir_intrinsic_base(intr);
 146    const unsigned component = nir_intrinsic_component(intr);
 147    bool out_of_bounds;
 148
 149    if (location == VARYING_SLOT_TESS_LEVEL_INNER) {
 150       switch (primitive_mode) {
 151       case GL_QUADS:
 152          /* gl_TessLevelInner[0..1] lives at DWords 3-2 (reversed). */
 153          nir_intrinsic_set_base(intr, 0);
 154          nir_intrinsic_set_component(intr, 3 - component);
 155          out_of_bounds = false;
 156          break;
 157       case GL_TRIANGLES:
 158          /* gl_TessLevelInner[0] lives at DWord 4. */
 159          nir_intrinsic_set_base(intr, 1);
 160          out_of_bounds = component > 0;
 161          break;
 162       case GL_ISOLINES:
 163          out_of_bounds = true;
 164          break;
 165       default:
 166          unreachable("Bogus tessellation domain");
 167       }
 168    } else if (location == VARYING_SLOT_TESS_LEVEL_OUTER) {
 169       if (primitive_mode == GL_ISOLINES) {
 170          /* gl_TessLevelOuter[0..1] lives at DWords 6-7 (in order). */
 171          nir_intrinsic_set_base(intr, 1);
 172          nir_intrinsic_set_component(intr, 2 + nir_intrinsic_component(intr));
 173          out_of_bounds = component > 1;
 174       } else {
 175          /* Triangles use DWords 7-5 (reversed); Quads use 7-4 (reversed) */
 176          nir_intrinsic_set_base(intr, 1);
 177          nir_intrinsic_set_component(intr, 3 - nir_intrinsic_component(intr));
 178          out_of_bounds = component == 3 && primitive_mode == GL_TRIANGLES;
 179       }
 180    } else {
 181       return false;
 182    }
 183
 184    if (out_of_bounds) {
 185       if (nir_intrinsic_infos[intr->intrinsic].has_dest) {
 186          b->cursor = nir_before_instr(&intr->instr);
 187          nir_ssa_def *undef = nir_ssa_undef(b, 1, 32);
 188          nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(undef));
 189       }
 190       nir_instr_remove(&intr->instr);
 191    }
 192
 193    return true;
 194 }
 195
 196 static bool
 197 remap_patch_urb_offsets(nir_block *block, nir_builder *b,
 198                         const struct brw_vue_map *vue_map,
 199                         GLenum tes_primitive_mode)
 200 {
 201    const bool is_passthrough_tcs = b->shader->info->name &&
 202       strcmp(b->shader->info->name, "passthrough") == 0;
 203
 204    nir_foreach_instr_safe(instr, block) {
 205       if (instr->type != nir_instr_type_intrinsic)
 206          continue;
 207
 208       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 209
 210       gl_shader_stage stage = b->shader->stage;
 211
 212       if ((stage == MESA_SHADER_TESS_CTRL && is_output(intrin)) ||
 213           (stage == MESA_SHADER_TESS_EVAL && is_input(intrin))) {
 214
 215          if (!is_passthrough_tcs &&
 216              remap_tess_levels(b, intrin, tes_primitive_mode))
 217             continue;
 218
 219          int vue_slot = vue_map->varying_to_slot[intrin->const_index[0]];
 220          assert(vue_slot != -1);
 221          intrin->const_index[0] = vue_slot;
 222
 223          nir_src *vertex = nir_get_io_vertex_index_src(intrin);
 224          if (vertex) {
 225             nir_const_value *const_vertex = nir_src_as_const_value(*vertex);
 226             if (const_vertex) {
 227                intrin->const_index[0] += const_vertex->u32[0] *
 228                                          vue_map->num_per_vertex_slots;
 229             } else {
 230                b->cursor = nir_before_instr(&intrin->instr);
 231
 232                /* Multiply by the number of per-vertex slots. */
 233                nir_ssa_def *vertex_offset =
 234                   nir_imul(b,
 235                            nir_ssa_for_src(b, *vertex, 1),
 236                            nir_imm_int(b,
 237                                        vue_map->num_per_vertex_slots));
 238
 239                /* Add it to the existing offset */
 240                nir_src *offset = nir_get_io_offset_src(intrin);
 241                nir_ssa_def *total_offset =
 242                   nir_iadd(b, vertex_offset,
 243                            nir_ssa_for_src(b, *offset, 1));
 244
 245                nir_instr_rewrite_src(&intrin->instr, offset,
 246                                      nir_src_for_ssa(total_offset));
 247             }
 248          }
 249       }
 250    }
 251    return true;
 252 }
 253
 254 void
 255 brw_nir_lower_vs_inputs(nir_shader *nir,
 256                         bool is_scalar,
 257                         bool use_legacy_snorm_formula,
 258                         const uint8_t *vs_attrib_wa_flags)
 259 {
 260    /* Start with the location of the variable's base. */
 261    foreach_list_typed(nir_variable, var, node, &nir->inputs) {
 262       var->data.driver_location = var->data.location;
 263    }
 264
 265    /* Now use nir_lower_io to walk dereference chains.  Attribute arrays are
 266     * loaded as one vec4 or dvec4 per element (or matrix column), depending on
 267     * whether it is a double-precision type or not.
 268     */
 269    nir_lower_io(nir, nir_var_shader_in, type_size_vec4, 0);
 270
 271    /* This pass needs actual constants */
 272    nir_opt_constant_folding(nir);
 273
 274    add_const_offset_to_base(nir, nir_var_shader_in);
 275
 276    brw_nir_apply_attribute_workarounds(nir, use_legacy_snorm_formula,
 277                                        vs_attrib_wa_flags);
 278
 279    if (is_scalar) {
 280       /* Finally, translate VERT_ATTRIB_* values into the actual registers. */
 281
 282       nir_foreach_function(function, nir) {
 283          if (function->impl) {
 284             nir_foreach_block(block, function->impl) {
 285                remap_vs_attrs(block, nir->info);
 286             }
 287          }
 288       }
 289    }
 290 }
 291
 292 void
 293 brw_nir_lower_vue_inputs(nir_shader *nir, bool is_scalar,
 294                          const struct brw_vue_map *vue_map)
 295 {
 296    foreach_list_typed(nir_variable, var, node, &nir->inputs) {
 297       var->data.driver_location = var->data.location;
 298    }
 299
 300    /* Inputs are stored in vec4 slots, so use type_size_vec4(). */
 301    nir_lower_io(nir, nir_var_shader_in, type_size_vec4, 0);
 302
 303    if (is_scalar || nir->stage != MESA_SHADER_GEOMETRY) {
 304       /* This pass needs actual constants */
 305       nir_opt_constant_folding(nir);
 306
 307       add_const_offset_to_base(nir, nir_var_shader_in);
 308
 309       nir_foreach_function(function, nir) {
 310          if (function->impl) {
 311             nir_foreach_block(block, function->impl) {
 312                remap_inputs_with_vue_map(block, vue_map);
 313             }
 314          }
 315       }
 316    }
 317 }
 318
 319 void
 320 brw_nir_lower_tes_inputs(nir_shader *nir, const struct brw_vue_map *vue_map)
 321 {
 322    foreach_list_typed(nir_variable, var, node, &nir->inputs) {
 323       var->data.driver_location = var->data.location;
 324    }
 325
 326    nir_lower_io(nir, nir_var_shader_in, type_size_vec4, 0);
 327
 328    /* This pass needs actual constants */
 329    nir_opt_constant_folding(nir);
 330
 331    add_const_offset_to_base(nir, nir_var_shader_in);
 332
 333    nir_foreach_function(function, nir) {
 334       if (function->impl) {
 335          nir_builder b;
 336          nir_builder_init(&b, function->impl);
 337          nir_foreach_block(block, function->impl) {
 338             remap_patch_urb_offsets(block, &b, vue_map,
 339                                     nir->info->tes.primitive_mode);
 340          }
 341       }
 342    }
 343 }
 344
 345 void
 346 brw_nir_lower_fs_inputs(nir_shader *nir, struct brw_vue_map *vue_map,
 347                         struct gl_program *prog,
 348                         const struct gen_device_info *devinfo,
 349                         const struct brw_wm_prog_key *key)
 350 {
 351    foreach_list_typed(nir_variable, var, node, &nir->inputs) {
 352       var->data.driver_location = var->data.location;
 353
 354       /* Apply default interpolation mode.
 355        *
 356        * Everything defaults to smooth except for the legacy GL color
 357        * built-in variables, which might be flat depending on API state.
 358        */
 359       if (var->data.interpolation == INTERP_MODE_NONE) {
 360          const bool flat = key->flat_shade &&
 361             (var->data.location == VARYING_SLOT_COL0 ||
 362              var->data.location == VARYING_SLOT_COL1);
 363
 364          var->data.interpolation = flat ? INTERP_MODE_FLAT
 365                                         : INTERP_MODE_SMOOTH;
 366       }
 367
 368       /* On Ironlake and below, there is only one interpolation mode.
 369        * Centroid interpolation doesn't mean anything on this hardware --
 370        * there is no multisampling.
 371        */
 372       if (devinfo->gen < 6) {
 373          var->data.centroid = false;
 374          var->data.sample = false;
 375       }
 376    }
 377
 378    if (devinfo->gen < 6) {
 379       assert(prog); /* prog will be NULL when called from Vulkan */
 380       brw_setup_vue_interpolation(vue_map, nir, prog, devinfo);
 381    }
 382
 383    nir_lower_io_options lower_io_options = 0;
 384    if (key->persample_interp)
 385       lower_io_options |= nir_lower_io_force_sample_interpolation;
 386
 387    nir_lower_io(nir, nir_var_shader_in, type_size_vec4, lower_io_options);
 388
 389    /* This pass needs actual constants */
 390    nir_opt_constant_folding(nir);
 391
 392    add_const_offset_to_base(nir, nir_var_shader_in);
 393 }
 394
 395 void
 396 brw_nir_lower_vue_outputs(nir_shader *nir,
 397                           bool is_scalar)
 398 {
 399    nir_foreach_variable(var, &nir->outputs) {
 400       var->data.driver_location = var->data.location;
 401    }
 402
 403    nir_lower_io(nir, nir_var_shader_out, type_size_vec4, 0);
 404 }
 405
 406 void
 407 brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map,
 408                           GLenum tes_primitive_mode)
 409 {
 410    nir_foreach_variable(var, &nir->outputs) {
 411       var->data.driver_location = var->data.location;
 412    }
 413
 414    nir_lower_io(nir, nir_var_shader_out, type_size_vec4, 0);
 415
 416    /* This pass needs actual constants */
 417    nir_opt_constant_folding(nir);
 418
 419    add_const_offset_to_base(nir, nir_var_shader_out);
 420
 421    nir_foreach_function(function, nir) {
 422       if (function->impl) {
 423          nir_builder b;
 424          nir_builder_init(&b, function->impl);
 425          nir_foreach_block(block, function->impl) {
 426             remap_patch_urb_offsets(block, &b, vue_map, tes_primitive_mode);
 427          }
 428       }
 429    }
 430 }
 431
 432 void
 433 brw_nir_lower_fs_outputs(nir_shader *nir)
 434 {
 435    nir_foreach_variable(var, &nir->outputs) {
 436       var->data.driver_location =
 437          SET_FIELD(var->data.index, BRW_NIR_FRAG_OUTPUT_INDEX) |
 438          SET_FIELD(var->data.location, BRW_NIR_FRAG_OUTPUT_LOCATION);
 439    }
 440
 441    nir_lower_io(nir, nir_var_shader_out, type_size_dvec4, 0);
 442 }
 443
 444 void
 445 brw_nir_lower_cs_shared(nir_shader *nir)
 446 {
 447    nir_assign_var_locations(&nir->shared, &nir->num_shared,
 448                             type_size_scalar_bytes);
 449    nir_lower_io(nir, nir_var_shared, type_size_scalar_bytes, 0);
 450 }
 451
 452 #define OPT(pass, ...) ({                                  \
 453    bool this_progress = false;                             \
 454    NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
 455    if (this_progress)                                      \
 456       progress = true;                                     \
 457    this_progress;                                          \
 458 })
 459
 460 #define OPT_V(pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
 461
 462 static nir_shader *
 463 nir_optimize(nir_shader *nir, const struct brw_compiler *compiler,
 464              bool is_scalar)
 465 {
 466    nir_variable_mode indirect_mask = 0;
 467    if (compiler->glsl_compiler_options[nir->stage].EmitNoIndirectInput)
 468       indirect_mask |= nir_var_shader_in;
 469    if (compiler->glsl_compiler_options[nir->stage].EmitNoIndirectOutput)
 470       indirect_mask |= nir_var_shader_out;
 471    if (compiler->glsl_compiler_options[nir->stage].EmitNoIndirectTemp)
 472       indirect_mask |= nir_var_local;
 473
 474    bool progress;
 475    do {
 476       progress = false;
 477       OPT_V(nir_lower_vars_to_ssa);
 478       OPT(nir_opt_copy_prop_vars);
 479
 480       if (is_scalar) {
 481          OPT(nir_lower_alu_to_scalar);
 482       }
 483
 484       OPT(nir_copy_prop);
 485
 486       if (is_scalar) {
 487          OPT(nir_lower_phis_to_scalar);
 488       }
 489
 490       OPT(nir_copy_prop);
 491       OPT(nir_opt_dce);
 492       OPT(nir_opt_cse);
 493       OPT(nir_opt_peephole_select, 0);
 494       OPT(nir_opt_algebraic);
 495       OPT(nir_opt_constant_folding);
 496       OPT(nir_opt_dead_cf);
 497       if (OPT(nir_opt_trivial_continues)) {
 498          /* If nir_opt_trivial_continues makes progress, then we need to clean
 499           * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
 500           * to make progress.
 501           */
 502          OPT(nir_copy_prop);
 503          OPT(nir_opt_dce);
 504       }
 505       OPT(nir_opt_if);
 506       if (nir->options->max_unroll_iterations != 0) {
 507          OPT(nir_opt_loop_unroll, indirect_mask);
 508       }
 509       OPT(nir_opt_remove_phis);
 510       OPT(nir_opt_undef);
 511       OPT_V(nir_lower_doubles, nir_lower_drcp |
 512                                nir_lower_dsqrt |
 513                                nir_lower_drsq |
 514                                nir_lower_dtrunc |
 515                                nir_lower_dfloor |
 516                                nir_lower_dceil |
 517                                nir_lower_dfract |
 518                                nir_lower_dround_even |
 519                                nir_lower_dmod);
 520       OPT_V(nir_lower_double_pack);
 521    } while (progress);
 522
 523    return nir;
 524 }
 525
 526 /* Does some simple lowering and runs the standard suite of optimizations
 527  *
 528  * This is intended to be called more-or-less directly after you get the
 529  * shader out of GLSL or some other source.  While it is geared towards i965,
 530  * it is not at all generator-specific except for the is_scalar flag.  Even
 531  * there, it is safe to call with is_scalar = false for a shader that is
 532  * intended for the FS backend as long as nir_optimize is called again with
 533  * is_scalar = true to scalarize everything prior to code gen.
 534  */
 535 nir_shader *
 536 brw_preprocess_nir(const struct brw_compiler *compiler, nir_shader *nir)
 537 {
 538    const struct gen_device_info *devinfo = compiler->devinfo;
 539    bool progress; /* Written by OPT and OPT_V */
 540    (void)progress;
 541
 542    const bool is_scalar = compiler->scalar_stage[nir->stage];
 543
 544    if (nir->stage == MESA_SHADER_GEOMETRY)
 545       OPT(nir_lower_gs_intrinsics);
 546
 547    /* See also brw_nir_trig_workarounds.py */
 548    if (compiler->precise_trig &&
 549        !(devinfo->gen >= 10 || devinfo->is_kabylake))
 550       OPT(brw_nir_apply_trig_workarounds);
 551
 552    static const nir_lower_tex_options tex_options = {
 553       .lower_txp = ~0,
 554       .lower_txf_offset = true,
 555       .lower_rect_offset = true,
 556       .lower_txd_cube_map = true,
 557    };
 558
 559    OPT(nir_lower_tex, &tex_options);
 560    OPT(nir_normalize_cubemap_coords);
 561
 562    OPT(nir_lower_global_vars_to_local);
 563
 564    OPT(nir_split_var_copies);
 565
 566    nir = nir_optimize(nir, compiler, is_scalar);
 567
 568    if (is_scalar) {
 569       OPT_V(nir_lower_load_const_to_scalar);
 570    }
 571
 572    /* Lower a bunch of stuff */
 573    OPT_V(nir_lower_var_copies);
 574
 575    OPT_V(nir_lower_clip_cull_distance_arrays);
 576
 577    nir_variable_mode indirect_mask = 0;
 578    if (compiler->glsl_compiler_options[nir->stage].EmitNoIndirectInput)
 579       indirect_mask |= nir_var_shader_in;
 580    if (compiler->glsl_compiler_options[nir->stage].EmitNoIndirectOutput)
 581       indirect_mask |= nir_var_shader_out;
 582    if (compiler->glsl_compiler_options[nir->stage].EmitNoIndirectTemp)
 583       indirect_mask |= nir_var_local;
 584
 585    nir_lower_indirect_derefs(nir, indirect_mask);
 586
 587    /* Get rid of split copies */
 588    nir = nir_optimize(nir, compiler, is_scalar);
 589
 590    OPT(nir_remove_dead_variables, nir_var_local);
 591
 592    return nir;
 593 }
 594
 595 /* Prepare the given shader for codegen
 596  *
 597  * This function is intended to be called right before going into the actual
 598  * backend and is highly backend-specific.  Also, once this function has been
 599  * called on a shader, it will no longer be in SSA form so most optimizations
 600  * will not work.
 601  */
 602 nir_shader *
 603 brw_postprocess_nir(nir_shader *nir, const struct brw_compiler *compiler,
 604                     bool is_scalar)
 605 {
 606    const struct gen_device_info *devinfo = compiler->devinfo;
 607    bool debug_enabled =
 608       (INTEL_DEBUG & intel_debug_flag_for_shader_stage(nir->stage));
 609
 610    bool progress; /* Written by OPT and OPT_V */
 611    (void)progress;
 612
 613    nir = nir_optimize(nir, compiler, is_scalar);
 614
 615    if (devinfo->gen >= 6) {
 616       /* Try and fuse multiply-adds */
 617       OPT(brw_nir_opt_peephole_ffma);
 618    }
 619
 620    OPT(nir_opt_algebraic_late);
 621
 622    OPT(nir_lower_locals_to_regs);
 623
 624    OPT_V(nir_lower_to_source_mods);
 625    OPT(nir_copy_prop);
 626    OPT(nir_opt_dce);
 627
 628    if (unlikely(debug_enabled)) {
 629       /* Re-index SSA defs so we print more sensible numbers. */
 630       nir_foreach_function(function, nir) {
 631          if (function->impl)
 632             nir_index_ssa_defs(function->impl);
 633       }
 634
 635       fprintf(stderr, "NIR (SSA form) for %s shader:\n",
 636               _mesa_shader_stage_to_string(nir->stage));
 637       nir_print_shader(nir, stderr);
 638    }
 639
 640    OPT_V(nir_convert_from_ssa, true);
 641
 642    if (!is_scalar) {
 643       OPT_V(nir_move_vec_src_uses_to_dest);
 644       OPT(nir_lower_vec_to_movs);
 645    }
 646
 647    /* This is the last pass we run before we start emitting stuff.  It
 648     * determines when we need to insert boolean resolves on Gen <= 5.  We
 649     * run it last because it stashes data in instr->pass_flags and we don't
 650     * want that to be squashed by other NIR passes.
 651     */
 652    if (devinfo->gen <= 5)
 653       brw_nir_analyze_boolean_resolves(nir);
 654
 655    nir_sweep(nir);
 656
 657    if (unlikely(debug_enabled)) {
 658       fprintf(stderr, "NIR (final form) for %s shader:\n",
 659               _mesa_shader_stage_to_string(nir->stage));
 660       nir_print_shader(nir, stderr);
 661    }
 662
 663    return nir;
 664 }
 665
 666 nir_shader *
 667 brw_nir_apply_sampler_key(nir_shader *nir,
 668                           const struct brw_compiler *compiler,
 669                           const struct brw_sampler_prog_key_data *key_tex,
 670                           bool is_scalar)
 671 {
 672    const struct gen_device_info *devinfo = compiler->devinfo;
 673    nir_lower_tex_options tex_options = { 0 };
 674
 675    /* Iron Lake and prior require lowering of all rectangle textures */
 676    if (devinfo->gen < 6)
 677       tex_options.lower_rect = true;
 678
 679    /* Prior to Broadwell, our hardware can't actually do GL_CLAMP */
 680    if (devinfo->gen < 8) {
 681       tex_options.saturate_s = key_tex->gl_clamp_mask[0];
 682       tex_options.saturate_t = key_tex->gl_clamp_mask[1];
 683       tex_options.saturate_r = key_tex->gl_clamp_mask[2];
 684    }
 685
 686    /* Prior to Haswell, we have to fake texture swizzle */
 687    for (unsigned s = 0; s < MAX_SAMPLERS; s++) {
 688       if (key_tex->swizzles[s] == SWIZZLE_NOOP)
 689          continue;
 690
 691       tex_options.swizzle_result |= (1 << s);
 692       for (unsigned c = 0; c < 4; c++)
 693          tex_options.swizzles[s][c] = GET_SWZ(key_tex->swizzles[s], c);
 694    }
 695
 696    /* Prior to Haswell, we have to lower gradients on shadow samplers */
 697    tex_options.lower_txd_shadow = devinfo->gen < 8 && !devinfo->is_haswell;
 698
 699    tex_options.lower_y_uv_external = key_tex->y_uv_image_mask;
 700    tex_options.lower_y_u_v_external = key_tex->y_u_v_image_mask;
 701    tex_options.lower_yx_xuxv_external = key_tex->yx_xuxv_image_mask;
 702
 703    if (nir_lower_tex(nir, &tex_options)) {
 704       nir_validate_shader(nir);
 705       nir = nir_optimize(nir, compiler, is_scalar);
 706    }
 707
 708    return nir;
 709 }
 710
 711 enum brw_reg_type
 712 brw_type_for_nir_type(nir_alu_type type)
 713 {
 714    switch (type) {
 715    case nir_type_uint:
 716    case nir_type_uint32:
 717       return BRW_REGISTER_TYPE_UD;
 718    case nir_type_bool:
 719    case nir_type_int:
 720    case nir_type_bool32:
 721    case nir_type_int32:
 722       return BRW_REGISTER_TYPE_D;
 723    case nir_type_float:
 724    case nir_type_float32:
 725       return BRW_REGISTER_TYPE_F;
 726    case nir_type_float64:
 727       return BRW_REGISTER_TYPE_DF;
 728    case nir_type_int64:
 729    case nir_type_uint64:
 730       /* TODO we should only see these in moves, so for now it's ok, but when
 731        * we add actual 64-bit integer support we should fix this.
 732        */
 733       return BRW_REGISTER_TYPE_DF;
 734    default:
 735       unreachable("unknown type");
 736    }
 737
 738    return BRW_REGISTER_TYPE_F;
 739 }
 740
 741 /* Returns the glsl_base_type corresponding to a nir_alu_type.
 742  * This is used by both brw_vec4_nir and brw_fs_nir.
 743  */
 744 enum glsl_base_type
 745 brw_glsl_base_type_for_nir_type(nir_alu_type type)
 746 {
 747    switch (type) {
 748    case nir_type_float:
 749    case nir_type_float32:
 750       return GLSL_TYPE_FLOAT;
 751
 752    case nir_type_float64:
 753       return GLSL_TYPE_DOUBLE;
 754
 755    case nir_type_int:
 756    case nir_type_int32:
 757       return GLSL_TYPE_INT;
 758
 759    case nir_type_uint:
 760    case nir_type_uint32:
 761       return GLSL_TYPE_UINT;
 762
 763    default:
 764       unreachable("bad type");
 765    }
 766 }