src/mesa/state_tracker/st_program.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keithw@vmware.com>
  30   *   Brian Paul
  31   */
  32
  33
  34 #include "main/errors.h"
  35
  36 #include "main/hash.h"
  37 #include "main/mtypes.h"
  38 #include "program/prog_parameter.h"
  39 #include "program/prog_print.h"
  40 #include "program/prog_to_nir.h"
  41 #include "program/programopt.h"
  42
  43 #include "compiler/nir/nir.h"
  44 #include "compiler/nir/nir_serialize.h"
  45 #include "draw/draw_context.h"
  46
  47 #include "pipe/p_context.h"
  48 #include "pipe/p_defines.h"
  49 #include "pipe/p_shader_tokens.h"
  50 #include "draw/draw_context.h"
  51 #include "tgsi/tgsi_dump.h"
  52 #include "tgsi/tgsi_emulate.h"
  53 #include "tgsi/tgsi_parse.h"
  54 #include "tgsi/tgsi_ureg.h"
  55
  56 #include "util/u_memory.h"
  57
  58 #include "st_debug.h"
  59 #include "st_cb_bitmap.h"
  60 #include "st_cb_drawpixels.h"
  61 #include "st_context.h"
  62 #include "st_tgsi_lower_depth_clamp.h"
  63 #include "st_tgsi_lower_yuv.h"
  64 #include "st_program.h"
  65 #include "st_mesa_to_tgsi.h"
  66 #include "st_atifs_to_tgsi.h"
  67 #include "st_nir.h"
  68 #include "st_shader_cache.h"
  69 #include "st_util.h"
  70 #include "cso_cache/cso_context.h"
  71
  72
  73 static void
  74 destroy_program_variants(struct st_context *st, struct gl_program *target);
  75
  76 static void
  77 set_affected_state_flags(uint64_t *states,
  78                          struct gl_program *prog,
  79                          uint64_t new_constants,
  80                          uint64_t new_sampler_views,
  81                          uint64_t new_samplers,
  82                          uint64_t new_images,
  83                          uint64_t new_ubos,
  84                          uint64_t new_ssbos,
  85                          uint64_t new_atomics)
  86 {
  87    if (prog->Parameters->NumParameters)
  88       *states |= new_constants;
  89
  90    if (prog->info.num_textures)
  91       *states |= new_sampler_views | new_samplers;
  92
  93    if (prog->info.num_images)
  94       *states |= new_images;
  95
  96    if (prog->info.num_ubos)
  97       *states |= new_ubos;
  98
  99    if (prog->info.num_ssbos)
 100       *states |= new_ssbos;
 101
 102    if (prog->info.num_abos)
 103       *states |= new_atomics;
 104 }
 105
 106 /**
 107  * This determines which states will be updated when the shader is bound.
 108  */
 109 void
 110 st_set_prog_affected_state_flags(struct gl_program *prog)
 111 {
 112    uint64_t *states;
 113
 114    switch (prog->info.stage) {
 115    case MESA_SHADER_VERTEX:
 116       states = &((struct st_program*)prog)->affected_states;
 117
 118       *states = ST_NEW_VS_STATE |
 119                 ST_NEW_RASTERIZER |
 120                 ST_NEW_VERTEX_ARRAYS;
 121
 122       set_affected_state_flags(states, prog,
 123                                ST_NEW_VS_CONSTANTS,
 124                                ST_NEW_VS_SAMPLER_VIEWS,
 125                                ST_NEW_VS_SAMPLERS,
 126                                ST_NEW_VS_IMAGES,
 127                                ST_NEW_VS_UBOS,
 128                                ST_NEW_VS_SSBOS,
 129                                ST_NEW_VS_ATOMICS);
 130       break;
 131
 132    case MESA_SHADER_TESS_CTRL:
 133       states = &(st_program(prog))->affected_states;
 134
 135       *states = ST_NEW_TCS_STATE;
 136
 137       set_affected_state_flags(states, prog,
 138                                ST_NEW_TCS_CONSTANTS,
 139                                ST_NEW_TCS_SAMPLER_VIEWS,
 140                                ST_NEW_TCS_SAMPLERS,
 141                                ST_NEW_TCS_IMAGES,
 142                                ST_NEW_TCS_UBOS,
 143                                ST_NEW_TCS_SSBOS,
 144                                ST_NEW_TCS_ATOMICS);
 145       break;
 146
 147    case MESA_SHADER_TESS_EVAL:
 148       states = &(st_program(prog))->affected_states;
 149
 150       *states = ST_NEW_TES_STATE |
 151                 ST_NEW_RASTERIZER;
 152
 153       set_affected_state_flags(states, prog,
 154                                ST_NEW_TES_CONSTANTS,
 155                                ST_NEW_TES_SAMPLER_VIEWS,
 156                                ST_NEW_TES_SAMPLERS,
 157                                ST_NEW_TES_IMAGES,
 158                                ST_NEW_TES_UBOS,
 159                                ST_NEW_TES_SSBOS,
 160                                ST_NEW_TES_ATOMICS);
 161       break;
 162
 163    case MESA_SHADER_GEOMETRY:
 164       states = &(st_program(prog))->affected_states;
 165
 166       *states = ST_NEW_GS_STATE |
 167                 ST_NEW_RASTERIZER;
 168
 169       set_affected_state_flags(states, prog,
 170                                ST_NEW_GS_CONSTANTS,
 171                                ST_NEW_GS_SAMPLER_VIEWS,
 172                                ST_NEW_GS_SAMPLERS,
 173                                ST_NEW_GS_IMAGES,
 174                                ST_NEW_GS_UBOS,
 175                                ST_NEW_GS_SSBOS,
 176                                ST_NEW_GS_ATOMICS);
 177       break;
 178
 179    case MESA_SHADER_FRAGMENT:
 180       states = &((struct st_program*)prog)->affected_states;
 181
 182       /* gl_FragCoord and glDrawPixels always use constants. */
 183       *states = ST_NEW_FS_STATE |
 184                 ST_NEW_SAMPLE_SHADING |
 185                 ST_NEW_FS_CONSTANTS;
 186
 187       set_affected_state_flags(states, prog,
 188                                ST_NEW_FS_CONSTANTS,
 189                                ST_NEW_FS_SAMPLER_VIEWS,
 190                                ST_NEW_FS_SAMPLERS,
 191                                ST_NEW_FS_IMAGES,
 192                                ST_NEW_FS_UBOS,
 193                                ST_NEW_FS_SSBOS,
 194                                ST_NEW_FS_ATOMICS);
 195       break;
 196
 197    case MESA_SHADER_COMPUTE:
 198       states = &((struct st_program*)prog)->affected_states;
 199
 200       *states = ST_NEW_CS_STATE;
 201
 202       set_affected_state_flags(states, prog,
 203                                ST_NEW_CS_CONSTANTS,
 204                                ST_NEW_CS_SAMPLER_VIEWS,
 205                                ST_NEW_CS_SAMPLERS,
 206                                ST_NEW_CS_IMAGES,
 207                                ST_NEW_CS_UBOS,
 208                                ST_NEW_CS_SSBOS,
 209                                ST_NEW_CS_ATOMICS);
 210       break;
 211
 212    default:
 213       unreachable("unhandled shader stage");
 214    }
 215 }
 216
 217
 218 /**
 219  * Delete a shader variant.  Note the caller must unlink the variant from
 220  * the linked list.
 221  */
 222 static void
 223 delete_variant(struct st_context *st, struct st_variant *v, GLenum target)
 224 {
 225    if (v->driver_shader) {
 226       if (target == GL_VERTEX_PROGRAM_ARB &&
 227           ((struct st_common_variant*)v)->key.is_draw_shader) {
 228          /* Draw shader. */
 229          draw_delete_vertex_shader(st->draw, v->driver_shader);
 230       } else if (st->has_shareable_shaders || v->st == st) {
 231          /* The shader's context matches the calling context, or we
 232           * don't care.
 233           */
 234          switch (target) {
 235          case GL_VERTEX_PROGRAM_ARB:
 236             st->pipe->delete_vs_state(st->pipe, v->driver_shader);
 237             break;
 238          case GL_TESS_CONTROL_PROGRAM_NV:
 239             st->pipe->delete_tcs_state(st->pipe, v->driver_shader);
 240             break;
 241          case GL_TESS_EVALUATION_PROGRAM_NV:
 242             st->pipe->delete_tes_state(st->pipe, v->driver_shader);
 243             break;
 244          case GL_GEOMETRY_PROGRAM_NV:
 245             st->pipe->delete_gs_state(st->pipe, v->driver_shader);
 246             break;
 247          case GL_FRAGMENT_PROGRAM_ARB:
 248             st->pipe->delete_fs_state(st->pipe, v->driver_shader);
 249             break;
 250          case GL_COMPUTE_PROGRAM_NV:
 251             st->pipe->delete_compute_state(st->pipe, v->driver_shader);
 252             break;
 253          default:
 254             unreachable("bad shader type in delete_basic_variant");
 255          }
 256       } else {
 257          /* We can't delete a shader with a context different from the one
 258           * that created it.  Add it to the creating context's zombie list.
 259           */
 260          enum pipe_shader_type type =
 261             pipe_shader_type_from_mesa(_mesa_program_enum_to_shader_stage(target));
 262
 263          st_save_zombie_shader(v->st, type, v->driver_shader);
 264       }
 265    }
 266
 267    free(v);
 268 }
 269
 270 static void
 271 st_unbind_program(struct st_context *st, struct st_program *p)
 272 {
 273    /* Unbind the shader in cso_context and re-bind in st/mesa. */
 274    switch (p->Base.info.stage) {
 275    case MESA_SHADER_VERTEX:
 276       cso_set_vertex_shader_handle(st->cso_context, NULL);
 277       st->dirty |= ST_NEW_VS_STATE;
 278       break;
 279    case MESA_SHADER_TESS_CTRL:
 280       cso_set_tessctrl_shader_handle(st->cso_context, NULL);
 281       st->dirty |= ST_NEW_TCS_STATE;
 282       break;
 283    case MESA_SHADER_TESS_EVAL:
 284       cso_set_tesseval_shader_handle(st->cso_context, NULL);
 285       st->dirty |= ST_NEW_TES_STATE;
 286       break;
 287    case MESA_SHADER_GEOMETRY:
 288       cso_set_geometry_shader_handle(st->cso_context, NULL);
 289       st->dirty |= ST_NEW_GS_STATE;
 290       break;
 291    case MESA_SHADER_FRAGMENT:
 292       cso_set_fragment_shader_handle(st->cso_context, NULL);
 293       st->dirty |= ST_NEW_FS_STATE;
 294       break;
 295    case MESA_SHADER_COMPUTE:
 296       cso_set_compute_shader_handle(st->cso_context, NULL);
 297       st->dirty |= ST_NEW_CS_STATE;
 298       break;
 299    default:
 300       unreachable("invalid shader type");
 301    }
 302 }
 303
 304 /**
 305  * Free all basic program variants.
 306  */
 307 void
 308 st_release_variants(struct st_context *st, struct st_program *p)
 309 {
 310    struct st_variant *v;
 311
 312    /* If we are releasing shaders, re-bind them, because we don't
 313     * know which shaders are bound in the driver.
 314     */
 315    if (p->variants)
 316       st_unbind_program(st, p);
 317
 318    for (v = p->variants; v; ) {
 319       struct st_variant *next = v->next;
 320       delete_variant(st, v, p->Base.Target);
 321       v = next;
 322    }
 323
 324    p->variants = NULL;
 325
 326    if (p->state.tokens) {
 327       ureg_free_tokens(p->state.tokens);
 328       p->state.tokens = NULL;
 329    }
 330
 331    /* Note: Any setup of ->ir.nir that has had pipe->create_*_state called on
 332     * it has resulted in the driver taking ownership of the NIR.  Those
 333     * callers should be NULLing out the nir field in any pipe_shader_state
 334     * that might have this called in order to indicate that.
 335     *
 336     * GLSL IR and ARB programs will have set gl_program->nir to the same
 337     * shader as ir->ir.nir, so it will be freed by _mesa_delete_program().
 338     */
 339 }
 340
 341 /**
 342  * Free all basic program variants and unref program.
 343  */
 344 void
 345 st_release_program(struct st_context *st, struct st_program **p)
 346 {
 347    if (!*p)
 348       return;
 349
 350    destroy_program_variants(st, &((*p)->Base));
 351    st_reference_prog(st, p, NULL);
 352 }
 353
 354 void
 355 st_finalize_nir_before_variants(struct nir_shader *nir)
 356 {
 357    NIR_PASS_V(nir, nir_opt_access);
 358
 359    NIR_PASS_V(nir, nir_split_var_copies);
 360    NIR_PASS_V(nir, nir_lower_var_copies);
 361    if (nir->options->lower_all_io_to_temps ||
 362        nir->options->lower_all_io_to_elements ||
 363        nir->info.stage == MESA_SHADER_VERTEX ||
 364        nir->info.stage == MESA_SHADER_GEOMETRY) {
 365       NIR_PASS_V(nir, nir_lower_io_arrays_to_elements_no_indirects, false);
 366    } else if (nir->info.stage == MESA_SHADER_FRAGMENT) {
 367       NIR_PASS_V(nir, nir_lower_io_arrays_to_elements_no_indirects, true);
 368    }
 369
 370    st_nir_assign_vs_in_locations(nir);
 371 }
 372
 373 /**
 374  * Translate ARB (asm) program to NIR
 375  */
 376 static nir_shader *
 377 st_translate_prog_to_nir(struct st_context *st, struct gl_program *prog,
 378                          gl_shader_stage stage)
 379 {
 380    struct pipe_screen *screen = st->pipe->screen;
 381    const struct gl_shader_compiler_options *options =
 382       &st->ctx->Const.ShaderCompilerOptions[stage];
 383
 384    /* Translate to NIR */
 385    nir_shader *nir = prog_to_nir(prog, options->NirOptions);
 386    NIR_PASS_V(nir, nir_lower_regs_to_ssa); /* turn registers into SSA */
 387    nir_validate_shader(nir, "after st/ptn lower_regs_to_ssa");
 388
 389    NIR_PASS_V(nir, st_nir_lower_wpos_ytransform, prog, screen);
 390    NIR_PASS_V(nir, nir_lower_system_values);
 391
 392    /* Optimise NIR */
 393    NIR_PASS_V(nir, nir_opt_constant_folding);
 394    st_nir_opts(nir);
 395    st_finalize_nir_before_variants(nir);
 396
 397    if (st->allow_st_finalize_nir_twice)
 398       st_finalize_nir(st, prog, NULL, nir, true);
 399
 400    nir_validate_shader(nir, "after st/glsl finalize_nir");
 401
 402    return nir;
 403 }
 404
 405 void
 406 st_prepare_vertex_program(struct st_program *stp)
 407 {
 408    struct st_vertex_program *stvp = (struct st_vertex_program *)stp;
 409
 410    stvp->num_inputs = 0;
 411    memset(stvp->input_to_index, ~0, sizeof(stvp->input_to_index));
 412    memset(stvp->result_to_output, ~0, sizeof(stvp->result_to_output));
 413
 414    /* Determine number of inputs, the mappings between VERT_ATTRIB_x
 415     * and TGSI generic input indexes, plus input attrib semantic info.
 416     */
 417    for (unsigned attr = 0; attr < VERT_ATTRIB_MAX; attr++) {
 418       if ((stp->Base.info.inputs_read & BITFIELD64_BIT(attr)) != 0) {
 419          stvp->input_to_index[attr] = stvp->num_inputs;
 420          stvp->index_to_input[stvp->num_inputs] = attr;
 421          stvp->num_inputs++;
 422
 423          if ((stp->Base.DualSlotInputs & BITFIELD64_BIT(attr)) != 0) {
 424             /* add placeholder for second part of a double attribute */
 425             stvp->index_to_input[stvp->num_inputs] = ST_DOUBLE_ATTRIB_PLACEHOLDER;
 426             stvp->num_inputs++;
 427          }
 428       }
 429    }
 430    /* pre-setup potentially unused edgeflag input */
 431    stvp->input_to_index[VERT_ATTRIB_EDGEFLAG] = stvp->num_inputs;
 432    stvp->index_to_input[stvp->num_inputs] = VERT_ATTRIB_EDGEFLAG;
 433
 434    /* Compute mapping of vertex program outputs to slots. */
 435    unsigned num_outputs = 0;
 436    for (unsigned attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 437       if (stp->Base.info.outputs_written & BITFIELD64_BIT(attr))
 438          stvp->result_to_output[attr] = num_outputs++;
 439    }
 440    /* pre-setup potentially unused edgeflag output */
 441    stvp->result_to_output[VARYING_SLOT_EDGE] = num_outputs;
 442 }
 443
 444 void
 445 st_translate_stream_output_info(struct gl_program *prog)
 446 {
 447    struct gl_transform_feedback_info *info = prog->sh.LinkedTransformFeedback;
 448    if (!info)
 449       return;
 450
 451    /* Determine the (default) output register mapping for each output. */
 452    unsigned num_outputs = 0;
 453    ubyte output_mapping[VARYING_SLOT_TESS_MAX];
 454    memset(output_mapping, 0, sizeof(output_mapping));
 455
 456    for (unsigned attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 457       if (prog->info.outputs_written & BITFIELD64_BIT(attr))
 458          output_mapping[attr] = num_outputs++;
 459    }
 460
 461    /* Translate stream output info. */
 462    struct pipe_stream_output_info *so_info =
 463       &((struct st_program*)prog)->state.stream_output;
 464
 465    for (unsigned i = 0; i < info->NumOutputs; i++) {
 466       so_info->output[i].register_index =
 467          output_mapping[info->Outputs[i].OutputRegister];
 468       so_info->output[i].start_component = info->Outputs[i].ComponentOffset;
 469       so_info->output[i].num_components = info->Outputs[i].NumComponents;
 470       so_info->output[i].output_buffer = info->Outputs[i].OutputBuffer;
 471       so_info->output[i].dst_offset = info->Outputs[i].DstOffset;
 472       so_info->output[i].stream = info->Outputs[i].StreamId;
 473    }
 474
 475    for (unsigned i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
 476       so_info->stride[i] = info->Buffers[i].Stride;
 477    }
 478    so_info->num_outputs = info->NumOutputs;
 479 }
 480
 481 /**
 482  * Translate a vertex program.
 483  */
 484 bool
 485 st_translate_vertex_program(struct st_context *st,
 486                             struct st_program *stp)
 487 {
 488    struct ureg_program *ureg;
 489    enum pipe_error error;
 490    unsigned num_outputs = 0;
 491    unsigned attr;
 492    ubyte output_semantic_name[VARYING_SLOT_MAX] = {0};
 493    ubyte output_semantic_index[VARYING_SLOT_MAX] = {0};
 494
 495    if (stp->Base.arb.IsPositionInvariant)
 496       _mesa_insert_mvp_code(st->ctx, &stp->Base);
 497
 498    /* ARB_vp: */
 499    if (!stp->glsl_to_tgsi) {
 500       _mesa_remove_output_reads(&stp->Base, PROGRAM_OUTPUT);
 501
 502       /* This determines which states will be updated when the assembly
 503        * shader is bound.
 504        */
 505       stp->affected_states = ST_NEW_VS_STATE |
 506                               ST_NEW_RASTERIZER |
 507                               ST_NEW_VERTEX_ARRAYS;
 508
 509       if (stp->Base.Parameters->NumParameters)
 510          stp->affected_states |= ST_NEW_VS_CONSTANTS;
 511
 512       /* Translate to NIR if preferred. */
 513       if (st->pipe->screen->get_shader_param(st->pipe->screen,
 514                                              PIPE_SHADER_VERTEX,
 515                                              PIPE_SHADER_CAP_PREFERRED_IR)) {
 516          assert(!stp->glsl_to_tgsi);
 517
 518          if (stp->Base.nir)
 519             ralloc_free(stp->Base.nir);
 520
 521          if (stp->serialized_nir) {
 522             free(stp->serialized_nir);
 523             stp->serialized_nir = NULL;
 524          }
 525
 526          stp->state.type = PIPE_SHADER_IR_NIR;
 527          stp->Base.nir = st_translate_prog_to_nir(st, &stp->Base,
 528                                                   MESA_SHADER_VERTEX);
 529
 530          /* We must update stp->Base.info after translation and before
 531           * st_prepare_vertex_program is called, because inputs_read
 532           * may become outdated after NIR optimization passes.
 533           *
 534           * For ffvp/ARB_vp inputs_read is populated based
 535           * on declared attributes without taking their usage into
 536           * consideration. When creating shader variants we expect
 537           * that their inputs_read would match the base ones for
 538           * input mapping to work properly.
 539           */
 540          nir_shader_gather_info(stp->Base.nir,
 541                                 nir_shader_get_entrypoint(stp->Base.nir));
 542          st_nir_assign_vs_in_locations(stp->Base.nir);
 543          stp->Base.info = stp->Base.nir->info;
 544
 545          /* For st_draw_feedback, we need to generate TGSI too if draw doesn't
 546           * use LLVM.
 547           */
 548          if (draw_has_llvm()) {
 549             st_prepare_vertex_program(stp);
 550             return true;
 551          }
 552       }
 553    }
 554
 555    st_prepare_vertex_program(stp);
 556
 557    /* Get semantic names and indices. */
 558    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 559       if (stp->Base.info.outputs_written & BITFIELD64_BIT(attr)) {
 560          unsigned slot = num_outputs++;
 561          unsigned semantic_name, semantic_index;
 562          tgsi_get_gl_varying_semantic(attr, st->needs_texcoord_semantic,
 563                                       &semantic_name, &semantic_index);
 564          output_semantic_name[slot] = semantic_name;
 565          output_semantic_index[slot] = semantic_index;
 566       }
 567    }
 568    /* pre-setup potentially unused edgeflag output */
 569    output_semantic_name[num_outputs] = TGSI_SEMANTIC_EDGEFLAG;
 570    output_semantic_index[num_outputs] = 0;
 571
 572    ureg = ureg_create_with_screen(PIPE_SHADER_VERTEX, st->pipe->screen);
 573    if (ureg == NULL)
 574       return false;
 575
 576    if (stp->Base.info.clip_distance_array_size)
 577       ureg_property(ureg, TGSI_PROPERTY_NUM_CLIPDIST_ENABLED,
 578                     stp->Base.info.clip_distance_array_size);
 579    if (stp->Base.info.cull_distance_array_size)
 580       ureg_property(ureg, TGSI_PROPERTY_NUM_CULLDIST_ENABLED,
 581                     stp->Base.info.cull_distance_array_size);
 582
 583    if (ST_DEBUG & DEBUG_MESA) {
 584       _mesa_print_program(&stp->Base);
 585       _mesa_print_program_parameters(st->ctx, &stp->Base);
 586       debug_printf("\n");
 587    }
 588
 589    struct st_vertex_program *stvp = (struct st_vertex_program *)stp;
 590
 591    if (stp->glsl_to_tgsi) {
 592       error = st_translate_program(st->ctx,
 593                                    PIPE_SHADER_VERTEX,
 594                                    ureg,
 595                                    stp->glsl_to_tgsi,
 596                                    &stp->Base,
 597                                    /* inputs */
 598                                    stvp->num_inputs,
 599                                    stvp->input_to_index,
 600                                    NULL, /* inputSlotToAttr */
 601                                    NULL, /* input semantic name */
 602                                    NULL, /* input semantic index */
 603                                    NULL, /* interp mode */
 604                                    /* outputs */
 605                                    num_outputs,
 606                                    stvp->result_to_output,
 607                                    output_semantic_name,
 608                                    output_semantic_index);
 609
 610       st_translate_stream_output_info(&stp->Base);
 611
 612       free_glsl_to_tgsi_visitor(stp->glsl_to_tgsi);
 613    } else
 614       error = st_translate_mesa_program(st->ctx,
 615                                         PIPE_SHADER_VERTEX,
 616                                         ureg,
 617                                         &stp->Base,
 618                                         /* inputs */
 619                                         stvp->num_inputs,
 620                                         stvp->input_to_index,
 621                                         NULL, /* input semantic name */
 622                                         NULL, /* input semantic index */
 623                                         NULL,
 624                                         /* outputs */
 625                                         num_outputs,
 626                                         stvp->result_to_output,
 627                                         output_semantic_name,
 628                                         output_semantic_index);
 629
 630    if (error) {
 631       debug_printf("%s: failed to translate Mesa program:\n", __func__);
 632       _mesa_print_program(&stp->Base);
 633       debug_assert(0);
 634       return false;
 635    }
 636
 637    stp->state.tokens = ureg_get_tokens(ureg, NULL);
 638    ureg_destroy(ureg);
 639
 640    if (stp->glsl_to_tgsi) {
 641       stp->glsl_to_tgsi = NULL;
 642       st_store_ir_in_disk_cache(st, &stp->Base, false);
 643    }
 644
 645    return stp->state.tokens != NULL;
 646 }
 647
 648 static struct nir_shader *
 649 get_nir_shader(struct st_context *st, struct st_program *stp)
 650 {
 651    if (stp->Base.nir) {
 652       nir_shader *nir = stp->Base.nir;
 653
 654       /* The first shader variant takes ownership of NIR, so that there is
 655        * no cloning. Additional shader variants are always generated from
 656        * serialized NIR to save memory.
 657        */
 658       stp->Base.nir = NULL;
 659       assert(stp->serialized_nir && stp->serialized_nir_size);
 660       return nir;
 661    }
 662
 663    struct blob_reader blob_reader;
 664    const struct nir_shader_compiler_options *options =
 665       st->ctx->Const.ShaderCompilerOptions[stp->Base.info.stage].NirOptions;
 666
 667    blob_reader_init(&blob_reader, stp->serialized_nir, stp->serialized_nir_size);
 668    return nir_deserialize(NULL, options, &blob_reader);
 669 }
 670
 671 static const gl_state_index16 depth_range_state[STATE_LENGTH] =
 672    { STATE_DEPTH_RANGE };
 673
 674 static struct st_common_variant *
 675 st_create_vp_variant(struct st_context *st,
 676                      struct st_program *stvp,
 677                      const struct st_common_variant_key *key)
 678 {
 679    struct st_common_variant *vpv = CALLOC_STRUCT(st_common_variant);
 680    struct pipe_context *pipe = st->pipe;
 681    struct pipe_screen *screen = pipe->screen;
 682    struct pipe_shader_state state = {0};
 683
 684    static const gl_state_index16 point_size_state[STATE_LENGTH] =
 685       { STATE_INTERNAL, STATE_POINT_SIZE_CLAMPED, 0 };
 686    struct gl_program_parameter_list *params = stvp->Base.Parameters;
 687
 688    vpv->key = *key;
 689
 690    state.stream_output = stvp->state.stream_output;
 691
 692    if (stvp->state.type == PIPE_SHADER_IR_NIR &&
 693        (!key->is_draw_shader || draw_has_llvm())) {
 694       bool finalize = false;
 695
 696       state.type = PIPE_SHADER_IR_NIR;
 697       state.ir.nir = get_nir_shader(st, stvp);
 698       if (key->clamp_color) {
 699          NIR_PASS_V(state.ir.nir, nir_lower_clamp_color_outputs);
 700          finalize = true;
 701       }
 702       if (key->passthrough_edgeflags) {
 703          NIR_PASS_V(state.ir.nir, nir_lower_passthrough_edgeflags);
 704          finalize = true;
 705       }
 706
 707       if (key->lower_point_size) {
 708          _mesa_add_state_reference(params, point_size_state);
 709          NIR_PASS_V(state.ir.nir, nir_lower_point_size_mov,
 710                     point_size_state);
 711          finalize = true;
 712       }
 713
 714       if (key->lower_ucp) {
 715          bool can_compact = screen->get_param(screen,
 716                                               PIPE_CAP_NIR_COMPACT_ARRAYS);
 717
 718          bool use_eye = st->ctx->_Shader->CurrentProgram[MESA_SHADER_VERTEX] != NULL;
 719          gl_state_index16 clipplane_state[MAX_CLIP_PLANES][STATE_LENGTH];
 720          for (int i = 0; i < MAX_CLIP_PLANES; ++i) {
 721             if (use_eye) {
 722                clipplane_state[i][0] = STATE_CLIPPLANE;
 723                clipplane_state[i][1] = i;
 724             } else {
 725                clipplane_state[i][0] = STATE_INTERNAL;
 726                clipplane_state[i][1] = STATE_CLIP_INTERNAL;
 727                clipplane_state[i][2] = i;
 728             }
 729             _mesa_add_state_reference(params, clipplane_state[i]);
 730          }
 731
 732          NIR_PASS_V(state.ir.nir, nir_lower_clip_vs, key->lower_ucp,
 733                     true, can_compact, clipplane_state);
 734          NIR_PASS_V(state.ir.nir, nir_lower_io_to_temporaries,
 735                     nir_shader_get_entrypoint(state.ir.nir), true, false);
 736          NIR_PASS_V(state.ir.nir, nir_lower_global_vars_to_local);
 737          finalize = true;
 738       }
 739
 740       if (finalize || !st->allow_st_finalize_nir_twice) {
 741          st_finalize_nir(st, &stvp->Base, stvp->shader_program, state.ir.nir,
 742                          true);
 743
 744          /* Some of the lowering above may have introduced new varyings */
 745          nir_shader_gather_info(state.ir.nir,
 746                                 nir_shader_get_entrypoint(state.ir.nir));
 747       }
 748
 749       if (ST_DEBUG & DEBUG_PRINT_IR)
 750          nir_print_shader(state.ir.nir, stderr);
 751
 752       if (key->is_draw_shader)
 753          vpv->base.driver_shader = draw_create_vertex_shader(st->draw, &state);
 754       else
 755          vpv->base.driver_shader = pipe->create_vs_state(pipe, &state);
 756
 757       return vpv;
 758    }
 759
 760    state.type = PIPE_SHADER_IR_TGSI;
 761    state.tokens = tgsi_dup_tokens(stvp->state.tokens);
 762
 763    /* Emulate features. */
 764    if (key->clamp_color || key->passthrough_edgeflags) {
 765       const struct tgsi_token *tokens;
 766       unsigned flags =
 767          (key->clamp_color ? TGSI_EMU_CLAMP_COLOR_OUTPUTS : 0) |
 768          (key->passthrough_edgeflags ? TGSI_EMU_PASSTHROUGH_EDGEFLAG : 0);
 769
 770       tokens = tgsi_emulate(state.tokens, flags);
 771
 772       if (tokens) {
 773          tgsi_free_tokens(state.tokens);
 774          state.tokens = tokens;
 775       } else {
 776          fprintf(stderr, "mesa: cannot emulate deprecated features\n");
 777       }
 778    }
 779
 780    if (key->lower_depth_clamp) {
 781       unsigned depth_range_const =
 782             _mesa_add_state_reference(params, depth_range_state);
 783
 784       const struct tgsi_token *tokens;
 785       tokens = st_tgsi_lower_depth_clamp(state.tokens, depth_range_const,
 786                                          key->clip_negative_one_to_one);
 787       if (tokens != state.tokens)
 788          tgsi_free_tokens(state.tokens);
 789       state.tokens = tokens;
 790    }
 791
 792    if (ST_DEBUG & DEBUG_PRINT_IR)
 793       tgsi_dump(state.tokens, 0);
 794
 795    if (key->is_draw_shader)
 796       vpv->base.driver_shader = draw_create_vertex_shader(st->draw, &state);
 797    else
 798       vpv->base.driver_shader = pipe->create_vs_state(pipe, &state);
 799
 800    if (state.tokens) {
 801       tgsi_free_tokens(state.tokens);
 802    }
 803
 804    return vpv;
 805 }
 806
 807
 808 /**
 809  * Find/create a vertex program variant.
 810  */
 811 struct st_common_variant *
 812 st_get_vp_variant(struct st_context *st,
 813                   struct st_program *stp,
 814                   const struct st_common_variant_key *key)
 815 {
 816    struct st_vertex_program *stvp = (struct st_vertex_program *)stp;
 817    struct st_common_variant *vpv;
 818
 819    /* Search for existing variant */
 820    for (vpv = st_common_variant(stp->variants); vpv;
 821         vpv = st_common_variant(vpv->base.next)) {
 822       if (memcmp(&vpv->key, key, sizeof(*key)) == 0) {
 823          break;
 824       }
 825    }
 826
 827    if (!vpv) {
 828       /* create now */
 829       vpv = st_create_vp_variant(st, stp, key);
 830       if (vpv) {
 831          vpv->base.st = key->st;
 832
 833          unsigned num_inputs = stvp->num_inputs + key->passthrough_edgeflags;
 834          for (unsigned index = 0; index < num_inputs; ++index) {
 835             unsigned attr = stvp->index_to_input[index];
 836             if (attr == ST_DOUBLE_ATTRIB_PLACEHOLDER)
 837                continue;
 838             vpv->vert_attrib_mask |= 1u << attr;
 839          }
 840
 841          /* insert into list */
 842          vpv->base.next = stp->variants;
 843          stp->variants = &vpv->base;
 844       }
 845    }
 846
 847    return vpv;
 848 }
 849
 850
 851 /**
 852  * Translate a Mesa fragment shader into a TGSI shader.
 853  */
 854 bool
 855 st_translate_fragment_program(struct st_context *st,
 856                               struct st_program *stfp)
 857 {
 858    /* Non-GLSL programs: */
 859    if (!stfp->glsl_to_tgsi) {
 860       _mesa_remove_output_reads(&stfp->Base, PROGRAM_OUTPUT);
 861       if (st->ctx->Const.GLSLFragCoordIsSysVal)
 862          _mesa_program_fragment_position_to_sysval(&stfp->Base);
 863
 864       /* This determines which states will be updated when the assembly
 865        * shader is bound.
 866        *
 867        * fragment.position and glDrawPixels always use constants.
 868        */
 869       stfp->affected_states = ST_NEW_FS_STATE |
 870                               ST_NEW_SAMPLE_SHADING |
 871                               ST_NEW_FS_CONSTANTS;
 872
 873       if (stfp->ati_fs) {
 874          /* Just set them for ATI_fs unconditionally. */
 875          stfp->affected_states |= ST_NEW_FS_SAMPLER_VIEWS |
 876                                   ST_NEW_FS_SAMPLERS;
 877       } else {
 878          /* ARB_fp */
 879          if (stfp->Base.SamplersUsed)
 880             stfp->affected_states |= ST_NEW_FS_SAMPLER_VIEWS |
 881                                      ST_NEW_FS_SAMPLERS;
 882       }
 883
 884       /* Translate to NIR. */
 885       if (!stfp->ati_fs &&
 886           st->pipe->screen->get_shader_param(st->pipe->screen,
 887                                              PIPE_SHADER_FRAGMENT,
 888                                              PIPE_SHADER_CAP_PREFERRED_IR)) {
 889          nir_shader *nir =
 890             st_translate_prog_to_nir(st, &stfp->Base, MESA_SHADER_FRAGMENT);
 891
 892          if (stfp->Base.nir)
 893             ralloc_free(stfp->Base.nir);
 894          if (stfp->serialized_nir) {
 895             free(stfp->serialized_nir);
 896             stfp->serialized_nir = NULL;
 897          }
 898          stfp->state.type = PIPE_SHADER_IR_NIR;
 899          stfp->Base.nir = nir;
 900          return true;
 901       }
 902    }
 903
 904    ubyte outputMapping[2 * FRAG_RESULT_MAX];
 905    ubyte inputMapping[VARYING_SLOT_MAX];
 906    ubyte inputSlotToAttr[VARYING_SLOT_MAX];
 907    ubyte interpMode[PIPE_MAX_SHADER_INPUTS];  /* XXX size? */
 908    GLuint attr;
 909    GLbitfield64 inputsRead;
 910    struct ureg_program *ureg;
 911
 912    GLboolean write_all = GL_FALSE;
 913
 914    ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS];
 915    ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
 916    uint fs_num_inputs = 0;
 917
 918    ubyte fs_output_semantic_name[PIPE_MAX_SHADER_OUTPUTS];
 919    ubyte fs_output_semantic_index[PIPE_MAX_SHADER_OUTPUTS];
 920    uint fs_num_outputs = 0;
 921
 922    memset(inputSlotToAttr, ~0, sizeof(inputSlotToAttr));
 923
 924    /*
 925     * Convert Mesa program inputs to TGSI input register semantics.
 926     */
 927    inputsRead = stfp->Base.info.inputs_read;
 928    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 929       if ((inputsRead & BITFIELD64_BIT(attr)) != 0) {
 930          const GLuint slot = fs_num_inputs++;
 931
 932          inputMapping[attr] = slot;
 933          inputSlotToAttr[slot] = attr;
 934
 935          switch (attr) {
 936          case VARYING_SLOT_POS:
 937             input_semantic_name[slot] = TGSI_SEMANTIC_POSITION;
 938             input_semantic_index[slot] = 0;
 939             interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
 940             break;
 941          case VARYING_SLOT_COL0:
 942             input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 943             input_semantic_index[slot] = 0;
 944             interpMode[slot] = stfp->glsl_to_tgsi ?
 945                TGSI_INTERPOLATE_COUNT : TGSI_INTERPOLATE_COLOR;
 946             break;
 947          case VARYING_SLOT_COL1:
 948             input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 949             input_semantic_index[slot] = 1;
 950             interpMode[slot] = stfp->glsl_to_tgsi ?
 951                TGSI_INTERPOLATE_COUNT : TGSI_INTERPOLATE_COLOR;
 952             break;
 953          case VARYING_SLOT_FOGC:
 954             input_semantic_name[slot] = TGSI_SEMANTIC_FOG;
 955             input_semantic_index[slot] = 0;
 956             interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
 957             break;
 958          case VARYING_SLOT_FACE:
 959             input_semantic_name[slot] = TGSI_SEMANTIC_FACE;
 960             input_semantic_index[slot] = 0;
 961             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 962             break;
 963          case VARYING_SLOT_PRIMITIVE_ID:
 964             input_semantic_name[slot] = TGSI_SEMANTIC_PRIMID;
 965             input_semantic_index[slot] = 0;
 966             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 967             break;
 968          case VARYING_SLOT_LAYER:
 969             input_semantic_name[slot] = TGSI_SEMANTIC_LAYER;
 970             input_semantic_index[slot] = 0;
 971             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 972             break;
 973          case VARYING_SLOT_VIEWPORT:
 974             input_semantic_name[slot] = TGSI_SEMANTIC_VIEWPORT_INDEX;
 975             input_semantic_index[slot] = 0;
 976             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 977             break;
 978          case VARYING_SLOT_CLIP_DIST0:
 979             input_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 980             input_semantic_index[slot] = 0;
 981             interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
 982             break;
 983          case VARYING_SLOT_CLIP_DIST1:
 984             input_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 985             input_semantic_index[slot] = 1;
 986             interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
 987             break;
 988          case VARYING_SLOT_CULL_DIST0:
 989          case VARYING_SLOT_CULL_DIST1:
 990             /* these should have been lowered by GLSL */
 991             assert(0);
 992             break;
 993             /* In most cases, there is nothing special about these
 994              * inputs, so adopt a convention to use the generic
 995              * semantic name and the mesa VARYING_SLOT_ number as the
 996              * index.
 997              *
 998              * All that is required is that the vertex shader labels
 999              * its own outputs similarly, and that the vertex shader
1000              * generates at least every output required by the
1001              * fragment shader plus fixed-function hardware (such as
1002              * BFC).
1003              *
1004              * However, some drivers may need us to identify the PNTC and TEXi
1005              * varyings if, for example, their capability to replace them with
1006              * sprite coordinates is limited.
1007              */
1008          case VARYING_SLOT_PNTC:
1009             if (st->needs_texcoord_semantic) {
1010                input_semantic_name[slot] = TGSI_SEMANTIC_PCOORD;
1011                input_semantic_index[slot] = 0;
1012                interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
1013                break;
1014             }
1015             /* fall through */
1016          case VARYING_SLOT_TEX0:
1017          case VARYING_SLOT_TEX1:
1018          case VARYING_SLOT_TEX2:
1019          case VARYING_SLOT_TEX3:
1020          case VARYING_SLOT_TEX4:
1021          case VARYING_SLOT_TEX5:
1022          case VARYING_SLOT_TEX6:
1023          case VARYING_SLOT_TEX7:
1024             if (st->needs_texcoord_semantic) {
1025                input_semantic_name[slot] = TGSI_SEMANTIC_TEXCOORD;
1026                input_semantic_index[slot] = attr - VARYING_SLOT_TEX0;
1027                interpMode[slot] = stfp->glsl_to_tgsi ?
1028                   TGSI_INTERPOLATE_COUNT : TGSI_INTERPOLATE_PERSPECTIVE;
1029                break;
1030             }
1031             /* fall through */
1032          case VARYING_SLOT_VAR0:
1033          default:
1034             /* Semantic indices should be zero-based because drivers may choose
1035              * to assign a fixed slot determined by that index.
1036              * This is useful because ARB_separate_shader_objects uses location
1037              * qualifiers for linkage, and if the semantic index corresponds to
1038              * these locations, linkage passes in the driver become unecessary.
1039              *
1040              * If needs_texcoord_semantic is true, no semantic indices will be
1041              * consumed for the TEXi varyings, and we can base the locations of
1042              * the user varyings on VAR0.  Otherwise, we use TEX0 as base index.
1043              */
1044             assert(attr >= VARYING_SLOT_VAR0 || attr == VARYING_SLOT_PNTC ||
1045                    (attr >= VARYING_SLOT_TEX0 && attr <= VARYING_SLOT_TEX7));
1046             input_semantic_name[slot] = TGSI_SEMANTIC_GENERIC;
1047             input_semantic_index[slot] = st_get_generic_varying_index(st, attr);
1048             if (attr == VARYING_SLOT_PNTC)
1049                interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
1050             else {
1051                interpMode[slot] = stfp->glsl_to_tgsi ?
1052                   TGSI_INTERPOLATE_COUNT : TGSI_INTERPOLATE_PERSPECTIVE;
1053             }
1054             break;
1055          }
1056       }
1057       else {
1058          inputMapping[attr] = -1;
1059       }
1060    }
1061
1062    /*
1063     * Semantics and mapping for outputs
1064     */
1065    GLbitfield64 outputsWritten = stfp->Base.info.outputs_written;
1066
1067    /* if z is written, emit that first */
1068    if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
1069       fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_POSITION;
1070       fs_output_semantic_index[fs_num_outputs] = 0;
1071       outputMapping[FRAG_RESULT_DEPTH] = fs_num_outputs;
1072       fs_num_outputs++;
1073       outputsWritten &= ~(1 << FRAG_RESULT_DEPTH);
1074    }
1075
1076    if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_STENCIL)) {
1077       fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_STENCIL;
1078       fs_output_semantic_index[fs_num_outputs] = 0;
1079       outputMapping[FRAG_RESULT_STENCIL] = fs_num_outputs;
1080       fs_num_outputs++;
1081       outputsWritten &= ~(1 << FRAG_RESULT_STENCIL);
1082    }
1083
1084    if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK)) {
1085       fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_SAMPLEMASK;
1086       fs_output_semantic_index[fs_num_outputs] = 0;
1087       outputMapping[FRAG_RESULT_SAMPLE_MASK] = fs_num_outputs;
1088       fs_num_outputs++;
1089       outputsWritten &= ~(1 << FRAG_RESULT_SAMPLE_MASK);
1090    }
1091
1092    /* handle remaining outputs (color) */
1093    for (attr = 0; attr < ARRAY_SIZE(outputMapping); attr++) {
1094       const GLbitfield64 written = attr < FRAG_RESULT_MAX ? outputsWritten :
1095          stfp->Base.SecondaryOutputsWritten;
1096       const unsigned loc = attr % FRAG_RESULT_MAX;
1097
1098       if (written & BITFIELD64_BIT(loc)) {
1099          switch (loc) {
1100          case FRAG_RESULT_DEPTH:
1101          case FRAG_RESULT_STENCIL:
1102          case FRAG_RESULT_SAMPLE_MASK:
1103             /* handled above */
1104             assert(0);
1105             break;
1106          case FRAG_RESULT_COLOR:
1107             write_all = GL_TRUE; /* fallthrough */
1108          default: {
1109             int index;
1110             assert(loc == FRAG_RESULT_COLOR ||
1111                    (FRAG_RESULT_DATA0 <= loc && loc < FRAG_RESULT_MAX));
1112
1113             index = (loc == FRAG_RESULT_COLOR) ? 0 : (loc - FRAG_RESULT_DATA0);
1114
1115             if (attr >= FRAG_RESULT_MAX) {
1116                /* Secondary color for dual source blending. */
1117                assert(index == 0);
1118                index++;
1119             }
1120
1121             fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_COLOR;
1122             fs_output_semantic_index[fs_num_outputs] = index;
1123             outputMapping[attr] = fs_num_outputs;
1124             break;
1125          }
1126          }
1127
1128          fs_num_outputs++;
1129       }
1130    }
1131
1132    ureg = ureg_create_with_screen(PIPE_SHADER_FRAGMENT, st->pipe->screen);
1133    if (ureg == NULL)
1134       return false;
1135
1136    if (ST_DEBUG & DEBUG_MESA) {
1137       _mesa_print_program(&stfp->Base);
1138       _mesa_print_program_parameters(st->ctx, &stfp->Base);
1139       debug_printf("\n");
1140    }
1141    if (write_all == GL_TRUE)
1142       ureg_property(ureg, TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS, 1);
1143
1144    if (stfp->Base.info.fs.depth_layout != FRAG_DEPTH_LAYOUT_NONE) {
1145       switch (stfp->Base.info.fs.depth_layout) {
1146       case FRAG_DEPTH_LAYOUT_ANY:
1147          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
1148                        TGSI_FS_DEPTH_LAYOUT_ANY);
1149          break;
1150       case FRAG_DEPTH_LAYOUT_GREATER:
1151          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
1152                        TGSI_FS_DEPTH_LAYOUT_GREATER);
1153          break;
1154       case FRAG_DEPTH_LAYOUT_LESS:
1155          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
1156                        TGSI_FS_DEPTH_LAYOUT_LESS);
1157          break;
1158       case FRAG_DEPTH_LAYOUT_UNCHANGED:
1159          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
1160                        TGSI_FS_DEPTH_LAYOUT_UNCHANGED);
1161          break;
1162       default:
1163          assert(0);
1164       }
1165    }
1166
1167    if (stfp->glsl_to_tgsi) {
1168       st_translate_program(st->ctx,
1169                            PIPE_SHADER_FRAGMENT,
1170                            ureg,
1171                            stfp->glsl_to_tgsi,
1172                            &stfp->Base,
1173                            /* inputs */
1174                            fs_num_inputs,
1175                            inputMapping,
1176                            inputSlotToAttr,
1177                            input_semantic_name,
1178                            input_semantic_index,
1179                            interpMode,
1180                            /* outputs */
1181                            fs_num_outputs,
1182                            outputMapping,
1183                            fs_output_semantic_name,
1184                            fs_output_semantic_index);
1185
1186       free_glsl_to_tgsi_visitor(stfp->glsl_to_tgsi);
1187    } else if (stfp->ati_fs)
1188       st_translate_atifs_program(ureg,
1189                                  stfp->ati_fs,
1190                                  &stfp->Base,
1191                                  /* inputs */
1192                                  fs_num_inputs,
1193                                  inputMapping,
1194                                  input_semantic_name,
1195                                  input_semantic_index,
1196                                  interpMode,
1197                                  /* outputs */
1198                                  fs_num_outputs,
1199                                  outputMapping,
1200                                  fs_output_semantic_name,
1201                                  fs_output_semantic_index);
1202    else
1203       st_translate_mesa_program(st->ctx,
1204                                 PIPE_SHADER_FRAGMENT,
1205                                 ureg,
1206                                 &stfp->Base,
1207                                 /* inputs */
1208                                 fs_num_inputs,
1209                                 inputMapping,
1210                                 input_semantic_name,
1211                                 input_semantic_index,
1212                                 interpMode,
1213                                 /* outputs */
1214                                 fs_num_outputs,
1215                                 outputMapping,
1216                                 fs_output_semantic_name,
1217                                 fs_output_semantic_index);
1218
1219    stfp->state.tokens = ureg_get_tokens(ureg, NULL);
1220    ureg_destroy(ureg);
1221
1222    if (stfp->glsl_to_tgsi) {
1223       stfp->glsl_to_tgsi = NULL;
1224       st_store_ir_in_disk_cache(st, &stfp->Base, false);
1225    }
1226
1227    return stfp->state.tokens != NULL;
1228 }
1229
1230 static struct st_fp_variant *
1231 st_create_fp_variant(struct st_context *st,
1232                      struct st_program *stfp,
1233                      const struct st_fp_variant_key *key)
1234 {
1235    struct pipe_context *pipe = st->pipe;
1236    struct st_fp_variant *variant = CALLOC_STRUCT(st_fp_variant);
1237    struct pipe_shader_state state = {0};
1238    struct gl_program_parameter_list *params = stfp->Base.Parameters;
1239    static const gl_state_index16 texcoord_state[STATE_LENGTH] =
1240       { STATE_INTERNAL, STATE_CURRENT_ATTRIB, VERT_ATTRIB_TEX0 };
1241    static const gl_state_index16 scale_state[STATE_LENGTH] =
1242       { STATE_INTERNAL, STATE_PT_SCALE };
1243    static const gl_state_index16 bias_state[STATE_LENGTH] =
1244       { STATE_INTERNAL, STATE_PT_BIAS };
1245    static const gl_state_index16 alpha_ref_state[STATE_LENGTH] =
1246       { STATE_INTERNAL, STATE_ALPHA_REF };
1247
1248    if (!variant)
1249       return NULL;
1250
1251    if (stfp->state.type == PIPE_SHADER_IR_NIR) {
1252       bool finalize = false;
1253
1254       state.type = PIPE_SHADER_IR_NIR;
1255       state.ir.nir = get_nir_shader(st, stfp);
1256
1257       if (key->clamp_color) {
1258          NIR_PASS_V(state.ir.nir, nir_lower_clamp_color_outputs);
1259          finalize = true;
1260       }
1261
1262       if (key->lower_flatshade) {
1263          NIR_PASS_V(state.ir.nir, nir_lower_flatshade);
1264          finalize = true;
1265       }
1266
1267       if (key->lower_alpha_func != COMPARE_FUNC_NEVER) {
1268          _mesa_add_state_reference(params, alpha_ref_state);
1269          NIR_PASS_V(state.ir.nir, nir_lower_alpha_test, key->lower_alpha_func,
1270                     false, alpha_ref_state);
1271          finalize = true;
1272       }
1273
1274       if (key->lower_two_sided_color) {
1275          NIR_PASS_V(state.ir.nir, nir_lower_two_sided_color);
1276          finalize = true;
1277       }
1278
1279       if (key->persample_shading) {
1280           nir_shader *shader = state.ir.nir;
1281           nir_foreach_variable(var, &shader->inputs)
1282              var->data.sample = true;
1283           finalize = true;
1284       }
1285
1286       assert(!(key->bitmap && key->drawpixels));
1287
1288       /* glBitmap */
1289       if (key->bitmap) {
1290          nir_lower_bitmap_options options = {0};
1291
1292          variant->bitmap_sampler = ffs(~stfp->Base.SamplersUsed) - 1;
1293          options.sampler = variant->bitmap_sampler;
1294          options.swizzle_xxxx = st->bitmap.tex_format == PIPE_FORMAT_R8_UNORM;
1295
1296          NIR_PASS_V(state.ir.nir, nir_lower_bitmap, &options);
1297          finalize = true;
1298       }
1299
1300       /* glDrawPixels (color only) */
1301       if (key->drawpixels) {
1302          nir_lower_drawpixels_options options = {{0}};
1303          unsigned samplers_used = stfp->Base.SamplersUsed;
1304
1305          /* Find the first unused slot. */
1306          variant->drawpix_sampler = ffs(~samplers_used) - 1;
1307          options.drawpix_sampler = variant->drawpix_sampler;
1308          samplers_used |= (1 << variant->drawpix_sampler);
1309
1310          options.pixel_maps = key->pixelMaps;
1311          if (key->pixelMaps) {
1312             variant->pixelmap_sampler = ffs(~samplers_used) - 1;
1313             options.pixelmap_sampler = variant->pixelmap_sampler;
1314          }
1315
1316          options.scale_and_bias = key->scaleAndBias;
1317          if (key->scaleAndBias) {
1318             _mesa_add_state_reference(params, scale_state);
1319             memcpy(options.scale_state_tokens, scale_state,
1320                    sizeof(options.scale_state_tokens));
1321             _mesa_add_state_reference(params, bias_state);
1322             memcpy(options.bias_state_tokens, bias_state,
1323                    sizeof(options.bias_state_tokens));
1324          }
1325
1326          _mesa_add_state_reference(params, texcoord_state);
1327          memcpy(options.texcoord_state_tokens, texcoord_state,
1328                 sizeof(options.texcoord_state_tokens));
1329
1330          NIR_PASS_V(state.ir.nir, nir_lower_drawpixels, &options);
1331          finalize = true;
1332       }
1333
1334       if (unlikely(key->external.lower_nv12 || key->external.lower_iyuv ||
1335                    key->external.lower_xy_uxvx || key->external.lower_yx_xuxv ||
1336                    key->external.lower_ayuv || key->external.lower_xyuv)) {
1337
1338          st_nir_lower_samplers(pipe->screen, state.ir.nir,
1339                                stfp->shader_program, &stfp->Base);
1340
1341          nir_lower_tex_options options = {0};
1342          options.lower_y_uv_external = key->external.lower_nv12;
1343          options.lower_y_u_v_external = key->external.lower_iyuv;
1344          options.lower_xy_uxvx_external = key->external.lower_xy_uxvx;
1345          options.lower_yx_xuxv_external = key->external.lower_yx_xuxv;
1346          options.lower_ayuv_external = key->external.lower_ayuv;
1347          options.lower_xyuv_external = key->external.lower_xyuv;
1348          NIR_PASS_V(state.ir.nir, nir_lower_tex, &options);
1349          finalize = true;
1350       }
1351
1352       if (finalize || !st->allow_st_finalize_nir_twice) {
1353          st_finalize_nir(st, &stfp->Base, stfp->shader_program, state.ir.nir,
1354                          false);
1355       }
1356
1357       /* This pass needs to happen *after* nir_lower_sampler */
1358       if (unlikely(key->external.lower_nv12 || key->external.lower_iyuv ||
1359                    key->external.lower_xy_uxvx || key->external.lower_yx_xuxv ||
1360                    key->external.lower_ayuv || key->external.lower_xyuv)) {
1361          NIR_PASS_V(state.ir.nir, st_nir_lower_tex_src_plane,
1362                     ~stfp->Base.SamplersUsed,
1363                     key->external.lower_nv12 || key->external.lower_xy_uxvx ||
1364                        key->external.lower_yx_xuxv,
1365                     key->external.lower_iyuv);
1366          finalize = true;
1367       }
1368
1369       if (finalize || !st->allow_st_finalize_nir_twice) {
1370          /* Some of the lowering above may have introduced new varyings */
1371          nir_shader_gather_info(state.ir.nir,
1372                                 nir_shader_get_entrypoint(state.ir.nir));
1373
1374          struct pipe_screen *screen = pipe->screen;
1375          if (screen->finalize_nir)
1376             screen->finalize_nir(screen, state.ir.nir, false);
1377       }
1378
1379       if (ST_DEBUG & DEBUG_PRINT_IR)
1380          nir_print_shader(state.ir.nir, stderr);
1381
1382       variant->base.driver_shader = pipe->create_fs_state(pipe, &state);
1383       variant->key = *key;
1384
1385       return variant;
1386    }
1387
1388    state.tokens = stfp->state.tokens;
1389
1390    assert(!(key->bitmap && key->drawpixels));
1391
1392    /* Fix texture targets and add fog for ATI_fs */
1393    if (stfp->ati_fs) {
1394       const struct tgsi_token *tokens = st_fixup_atifs(state.tokens, key);
1395
1396       if (tokens)
1397          state.tokens = tokens;
1398       else
1399          fprintf(stderr, "mesa: cannot post-process ATI_fs\n");
1400    }
1401
1402    /* Emulate features. */
1403    if (key->clamp_color || key->persample_shading) {
1404       const struct tgsi_token *tokens;
1405       unsigned flags =
1406          (key->clamp_color ? TGSI_EMU_CLAMP_COLOR_OUTPUTS : 0) |
1407          (key->persample_shading ? TGSI_EMU_FORCE_PERSAMPLE_INTERP : 0);
1408
1409       tokens = tgsi_emulate(state.tokens, flags);
1410
1411       if (tokens) {
1412          if (state.tokens != stfp->state.tokens)
1413             tgsi_free_tokens(state.tokens);
1414          state.tokens = tokens;
1415       } else
1416          fprintf(stderr, "mesa: cannot emulate deprecated features\n");
1417    }
1418
1419    /* glBitmap */
1420    if (key->bitmap) {
1421       const struct tgsi_token *tokens;
1422
1423       variant->bitmap_sampler = ffs(~stfp->Base.SamplersUsed) - 1;
1424
1425       tokens = st_get_bitmap_shader(state.tokens,
1426                                     st->internal_target,
1427                                     variant->bitmap_sampler,
1428                                     st->needs_texcoord_semantic,
1429                                     st->bitmap.tex_format ==
1430                                     PIPE_FORMAT_R8_UNORM);
1431
1432       if (tokens) {
1433          if (state.tokens != stfp->state.tokens)
1434             tgsi_free_tokens(state.tokens);
1435          state.tokens = tokens;
1436       } else
1437          fprintf(stderr, "mesa: cannot create a shader for glBitmap\n");
1438    }
1439
1440    /* glDrawPixels (color only) */
1441    if (key->drawpixels) {
1442       const struct tgsi_token *tokens;
1443       unsigned scale_const = 0, bias_const = 0, texcoord_const = 0;
1444
1445       /* Find the first unused slot. */
1446       variant->drawpix_sampler = ffs(~stfp->Base.SamplersUsed) - 1;
1447
1448       if (key->pixelMaps) {
1449          unsigned samplers_used = stfp->Base.SamplersUsed |
1450                                   (1 << variant->drawpix_sampler);
1451
1452          variant->pixelmap_sampler = ffs(~samplers_used) - 1;
1453       }
1454
1455       if (key->scaleAndBias) {
1456          scale_const = _mesa_add_state_reference(params, scale_state);
1457          bias_const = _mesa_add_state_reference(params, bias_state);
1458       }
1459
1460       texcoord_const = _mesa_add_state_reference(params, texcoord_state);
1461
1462       tokens = st_get_drawpix_shader(state.tokens,
1463                                      st->needs_texcoord_semantic,
1464                                      key->scaleAndBias, scale_const,
1465                                      bias_const, key->pixelMaps,
1466                                      variant->drawpix_sampler,
1467                                      variant->pixelmap_sampler,
1468                                      texcoord_const, st->internal_target);
1469
1470       if (tokens) {
1471          if (state.tokens != stfp->state.tokens)
1472             tgsi_free_tokens(state.tokens);
1473          state.tokens = tokens;
1474       } else
1475          fprintf(stderr, "mesa: cannot create a shader for glDrawPixels\n");
1476    }
1477
1478    if (unlikely(key->external.lower_nv12 || key->external.lower_iyuv ||
1479                 key->external.lower_xy_uxvx || key->external.lower_yx_xuxv)) {
1480       const struct tgsi_token *tokens;
1481
1482       /* samplers inserted would conflict, but this should be unpossible: */
1483       assert(!(key->bitmap || key->drawpixels));
1484
1485       tokens = st_tgsi_lower_yuv(state.tokens,
1486                                  ~stfp->Base.SamplersUsed,
1487                                  key->external.lower_nv12 ||
1488                                     key->external.lower_xy_uxvx ||
1489                                     key->external.lower_yx_xuxv,
1490                                  key->external.lower_iyuv);
1491       if (tokens) {
1492          if (state.tokens != stfp->state.tokens)
1493             tgsi_free_tokens(state.tokens);
1494          state.tokens = tokens;
1495       } else {
1496          fprintf(stderr, "mesa: cannot create a shader for samplerExternalOES\n");
1497       }
1498    }
1499
1500    if (key->lower_depth_clamp) {
1501       unsigned depth_range_const = _mesa_add_state_reference(params, depth_range_state);
1502
1503       const struct tgsi_token *tokens;
1504       tokens = st_tgsi_lower_depth_clamp_fs(state.tokens, depth_range_const);
1505       if (state.tokens != stfp->state.tokens)
1506          tgsi_free_tokens(state.tokens);
1507       state.tokens = tokens;
1508    }
1509
1510    if (ST_DEBUG & DEBUG_PRINT_IR)
1511       tgsi_dump(state.tokens, 0);
1512
1513    /* fill in variant */
1514    variant->base.driver_shader = pipe->create_fs_state(pipe, &state);
1515    variant->key = *key;
1516
1517    if (state.tokens != stfp->state.tokens)
1518       tgsi_free_tokens(state.tokens);
1519    return variant;
1520 }
1521
1522 /**
1523  * Translate fragment program if needed.
1524  */
1525 struct st_fp_variant *
1526 st_get_fp_variant(struct st_context *st,
1527                   struct st_program *stfp,
1528                   const struct st_fp_variant_key *key)
1529 {
1530    struct st_fp_variant *fpv;
1531
1532    /* Search for existing variant */
1533    for (fpv = st_fp_variant(stfp->variants); fpv;
1534         fpv = st_fp_variant(fpv->base.next)) {
1535       if (memcmp(&fpv->key, key, sizeof(*key)) == 0) {
1536          break;
1537       }
1538    }
1539
1540    if (!fpv) {
1541       /* create new */
1542       fpv = st_create_fp_variant(st, stfp, key);
1543       if (fpv) {
1544          fpv->base.st = key->st;
1545
1546          if (key->bitmap || key->drawpixels) {
1547             /* Regular variants should always come before the
1548              * bitmap & drawpixels variants, (unless there
1549              * are no regular variants) so that
1550              * st_update_fp can take a fast path when
1551              * shader_has_one_variant is set.
1552              */
1553             if (!stfp->variants) {
1554                stfp->variants = &fpv->base;
1555             } else {
1556                /* insert into list after the first one */
1557                fpv->base.next = stfp->variants->next;
1558                stfp->variants->next = &fpv->base;
1559             }
1560          } else {
1561             /* insert into list */
1562             fpv->base.next = stfp->variants;
1563             stfp->variants = &fpv->base;
1564          }
1565       }
1566    }
1567
1568    return fpv;
1569 }
1570
1571 /**
1572  * Translate a program. This is common code for geometry and tessellation
1573  * shaders.
1574  */
1575 bool
1576 st_translate_common_program(struct st_context *st,
1577                             struct st_program *stp)
1578 {
1579    struct gl_program *prog = &stp->Base;
1580    enum pipe_shader_type stage =
1581       pipe_shader_type_from_mesa(stp->Base.info.stage);
1582    struct ureg_program *ureg = ureg_create_with_screen(stage, st->pipe->screen);
1583
1584    if (ureg == NULL)
1585       return false;
1586
1587    switch (stage) {
1588    case PIPE_SHADER_TESS_CTRL:
1589       ureg_property(ureg, TGSI_PROPERTY_TCS_VERTICES_OUT,
1590                     stp->Base.info.tess.tcs_vertices_out);
1591       break;
1592
1593    case PIPE_SHADER_TESS_EVAL:
1594       if (stp->Base.info.tess.primitive_mode == GL_ISOLINES)
1595          ureg_property(ureg, TGSI_PROPERTY_TES_PRIM_MODE, GL_LINES);
1596       else
1597          ureg_property(ureg, TGSI_PROPERTY_TES_PRIM_MODE,
1598                        stp->Base.info.tess.primitive_mode);
1599
1600       STATIC_ASSERT((TESS_SPACING_EQUAL + 1) % 3 == PIPE_TESS_SPACING_EQUAL);
1601       STATIC_ASSERT((TESS_SPACING_FRACTIONAL_ODD + 1) % 3 ==
1602                     PIPE_TESS_SPACING_FRACTIONAL_ODD);
1603       STATIC_ASSERT((TESS_SPACING_FRACTIONAL_EVEN + 1) % 3 ==
1604                     PIPE_TESS_SPACING_FRACTIONAL_EVEN);
1605
1606       ureg_property(ureg, TGSI_PROPERTY_TES_SPACING,
1607                     (stp->Base.info.tess.spacing + 1) % 3);
1608
1609       ureg_property(ureg, TGSI_PROPERTY_TES_VERTEX_ORDER_CW,
1610                     !stp->Base.info.tess.ccw);
1611       ureg_property(ureg, TGSI_PROPERTY_TES_POINT_MODE,
1612                     stp->Base.info.tess.point_mode);
1613       break;
1614
1615    case PIPE_SHADER_GEOMETRY:
1616       ureg_property(ureg, TGSI_PROPERTY_GS_INPUT_PRIM,
1617                     stp->Base.info.gs.input_primitive);
1618       ureg_property(ureg, TGSI_PROPERTY_GS_OUTPUT_PRIM,
1619                     stp->Base.info.gs.output_primitive);
1620       ureg_property(ureg, TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES,
1621                     stp->Base.info.gs.vertices_out);
1622       ureg_property(ureg, TGSI_PROPERTY_GS_INVOCATIONS,
1623                     stp->Base.info.gs.invocations);
1624       break;
1625
1626    default:
1627       break;
1628    }
1629
1630    ubyte inputSlotToAttr[VARYING_SLOT_TESS_MAX];
1631    ubyte inputMapping[VARYING_SLOT_TESS_MAX];
1632    ubyte outputMapping[VARYING_SLOT_TESS_MAX];
1633    GLuint attr;
1634
1635    ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS];
1636    ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
1637    uint num_inputs = 0;
1638
1639    ubyte output_semantic_name[PIPE_MAX_SHADER_OUTPUTS];
1640    ubyte output_semantic_index[PIPE_MAX_SHADER_OUTPUTS];
1641    uint num_outputs = 0;
1642
1643    GLint i;
1644
1645    memset(inputSlotToAttr, 0, sizeof(inputSlotToAttr));
1646    memset(inputMapping, 0, sizeof(inputMapping));
1647    memset(outputMapping, 0, sizeof(outputMapping));
1648    memset(&stp->state, 0, sizeof(stp->state));
1649
1650    if (prog->info.clip_distance_array_size)
1651       ureg_property(ureg, TGSI_PROPERTY_NUM_CLIPDIST_ENABLED,
1652                     prog->info.clip_distance_array_size);
1653    if (prog->info.cull_distance_array_size)
1654       ureg_property(ureg, TGSI_PROPERTY_NUM_CULLDIST_ENABLED,
1655                     prog->info.cull_distance_array_size);
1656
1657    /*
1658     * Convert Mesa program inputs to TGSI input register semantics.
1659     */
1660    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
1661       if ((prog->info.inputs_read & BITFIELD64_BIT(attr)) == 0)
1662          continue;
1663
1664       unsigned slot = num_inputs++;
1665
1666       inputMapping[attr] = slot;
1667       inputSlotToAttr[slot] = attr;
1668
1669       unsigned semantic_name, semantic_index;
1670       tgsi_get_gl_varying_semantic(attr, st->needs_texcoord_semantic,
1671                                    &semantic_name, &semantic_index);
1672       input_semantic_name[slot] = semantic_name;
1673       input_semantic_index[slot] = semantic_index;
1674    }
1675
1676    /* Also add patch inputs. */
1677    for (attr = 0; attr < 32; attr++) {
1678       if (prog->info.patch_inputs_read & (1u << attr)) {
1679          GLuint slot = num_inputs++;
1680          GLuint patch_attr = VARYING_SLOT_PATCH0 + attr;
1681
1682          inputMapping[patch_attr] = slot;
1683          inputSlotToAttr[slot] = patch_attr;
1684          input_semantic_name[slot] = TGSI_SEMANTIC_PATCH;
1685          input_semantic_index[slot] = attr;
1686       }
1687    }
1688
1689    /* initialize output semantics to defaults */
1690    for (i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
1691       output_semantic_name[i] = TGSI_SEMANTIC_GENERIC;
1692       output_semantic_index[i] = 0;
1693    }
1694
1695    /*
1696     * Determine number of outputs, the (default) output register
1697     * mapping and the semantic information for each output.
1698     */
1699    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
1700       if (prog->info.outputs_written & BITFIELD64_BIT(attr)) {
1701          GLuint slot = num_outputs++;
1702
1703          outputMapping[attr] = slot;
1704
1705          unsigned semantic_name, semantic_index;
1706          tgsi_get_gl_varying_semantic(attr, st->needs_texcoord_semantic,
1707                                       &semantic_name, &semantic_index);
1708          output_semantic_name[slot] = semantic_name;
1709          output_semantic_index[slot] = semantic_index;
1710       }
1711    }
1712
1713    /* Also add patch outputs. */
1714    for (attr = 0; attr < 32; attr++) {
1715       if (prog->info.patch_outputs_written & (1u << attr)) {
1716          GLuint slot = num_outputs++;
1717          GLuint patch_attr = VARYING_SLOT_PATCH0 + attr;
1718
1719          outputMapping[patch_attr] = slot;
1720          output_semantic_name[slot] = TGSI_SEMANTIC_PATCH;
1721          output_semantic_index[slot] = attr;
1722       }
1723    }
1724
1725    st_translate_program(st->ctx,
1726                         stage,
1727                         ureg,
1728                         stp->glsl_to_tgsi,
1729                         prog,
1730                         /* inputs */
1731                         num_inputs,
1732                         inputMapping,
1733                         inputSlotToAttr,
1734                         input_semantic_name,
1735                         input_semantic_index,
1736                         NULL,
1737                         /* outputs */
1738                         num_outputs,
1739                         outputMapping,
1740                         output_semantic_name,
1741                         output_semantic_index);
1742
1743    stp->state.tokens = ureg_get_tokens(ureg, NULL);
1744
1745    ureg_destroy(ureg);
1746
1747    st_translate_stream_output_info(prog);
1748
1749    st_store_ir_in_disk_cache(st, prog, false);
1750
1751    if (ST_DEBUG & DEBUG_PRINT_IR && ST_DEBUG & DEBUG_MESA)
1752       _mesa_print_program(prog);
1753
1754    free_glsl_to_tgsi_visitor(stp->glsl_to_tgsi);
1755    stp->glsl_to_tgsi = NULL;
1756    return true;
1757 }
1758
1759
1760 /**
1761  * Get/create a basic program variant.
1762  */
1763 struct st_variant *
1764 st_get_common_variant(struct st_context *st,
1765                       struct st_program *prog,
1766                       const struct st_common_variant_key *key)
1767 {
1768    struct pipe_context *pipe = st->pipe;
1769    struct st_variant *v;
1770    struct pipe_shader_state state = {0};
1771
1772    /* Search for existing variant */
1773    for (v = prog->variants; v; v = v->next) {
1774       if (memcmp(&st_common_variant(v)->key, key, sizeof(*key)) == 0)
1775          break;
1776    }
1777
1778    if (!v) {
1779       /* create new */
1780       v = (struct st_variant*)CALLOC_STRUCT(st_common_variant);
1781       if (v) {
1782          if (prog->state.type == PIPE_SHADER_IR_NIR) {
1783             bool finalize = false;
1784
1785             state.type = PIPE_SHADER_IR_NIR;
1786             state.ir.nir = get_nir_shader(st, prog);
1787
1788             if (key->clamp_color) {
1789                NIR_PASS_V(state.ir.nir, nir_lower_clamp_color_outputs);
1790                finalize = true;
1791             }
1792
1793             state.stream_output = prog->state.stream_output;
1794
1795             if (finalize || !st->allow_st_finalize_nir_twice) {
1796                st_finalize_nir(st, &prog->Base, prog->shader_program,
1797                                state.ir.nir, true);
1798             }
1799
1800             if (ST_DEBUG & DEBUG_PRINT_IR)
1801                nir_print_shader(state.ir.nir, stderr);
1802          } else {
1803             if (key->lower_depth_clamp) {
1804                struct gl_program_parameter_list *params = prog->Base.Parameters;
1805
1806                unsigned depth_range_const =
1807                      _mesa_add_state_reference(params, depth_range_state);
1808
1809                const struct tgsi_token *tokens;
1810                tokens =
1811                      st_tgsi_lower_depth_clamp(prog->state.tokens,
1812                                                depth_range_const,
1813                                                key->clip_negative_one_to_one);
1814
1815                if (tokens != prog->state.tokens)
1816                   tgsi_free_tokens(prog->state.tokens);
1817
1818                prog->state.tokens = tokens;
1819             }
1820             state = prog->state;
1821
1822             if (ST_DEBUG & DEBUG_PRINT_IR)
1823                tgsi_dump(state.tokens, 0);
1824          }
1825          /* fill in new variant */
1826          switch (prog->Base.info.stage) {
1827          case MESA_SHADER_TESS_CTRL:
1828             v->driver_shader = pipe->create_tcs_state(pipe, &state);
1829             break;
1830          case MESA_SHADER_TESS_EVAL:
1831             v->driver_shader = pipe->create_tes_state(pipe, &state);
1832             break;
1833          case MESA_SHADER_GEOMETRY:
1834             v->driver_shader = pipe->create_gs_state(pipe, &state);
1835             break;
1836          case MESA_SHADER_COMPUTE: {
1837             struct pipe_compute_state cs = {0};
1838             cs.ir_type = state.type;
1839             cs.req_local_mem = prog->Base.info.cs.shared_size;
1840
1841             if (state.type == PIPE_SHADER_IR_NIR)
1842                cs.prog = state.ir.nir;
1843             else
1844                cs.prog = state.tokens;
1845
1846             v->driver_shader = pipe->create_compute_state(pipe, &cs);
1847             break;
1848          }
1849          default:
1850             assert(!"unhandled shader type");
1851             free(v);
1852             return NULL;
1853          }
1854
1855          st_common_variant(v)->key = *key;
1856          v->st = key->st;
1857
1858          /* insert into list */
1859          v->next = prog->variants;
1860          prog->variants = v;
1861       }
1862    }
1863
1864    return v;
1865 }
1866
1867
1868 /**
1869  * Vert/Geom/Frag programs have per-context variants.  Free all the
1870  * variants attached to the given program which match the given context.
1871  */
1872 static void
1873 destroy_program_variants(struct st_context *st, struct gl_program *target)
1874 {
1875    if (!target || target == &_mesa_DummyProgram)
1876       return;
1877
1878    struct st_program *p = st_program(target);
1879    struct st_variant *v, **prevPtr = &p->variants;
1880    bool unbound = false;
1881
1882    for (v = p->variants; v; ) {
1883       struct st_variant *next = v->next;
1884       if (v->st == st) {
1885          if (!unbound) {
1886             st_unbind_program(st, p);
1887             unbound = true;
1888          }
1889
1890          /* unlink from list */
1891          *prevPtr = next;
1892          /* destroy this variant */
1893          delete_variant(st, v, target->Target);
1894       }
1895       else {
1896          prevPtr = &v->next;
1897       }
1898       v = next;
1899    }
1900 }
1901
1902
1903 /**
1904  * Callback for _mesa_HashWalk.  Free all the shader's program variants
1905  * which match the given context.
1906  */
1907 static void
1908 destroy_shader_program_variants_cb(GLuint key, void *data, void *userData)
1909 {
1910    struct st_context *st = (struct st_context *) userData;
1911    struct gl_shader *shader = (struct gl_shader *) data;
1912
1913    switch (shader->Type) {
1914    case GL_SHADER_PROGRAM_MESA:
1915       {
1916          struct gl_shader_program *shProg = (struct gl_shader_program *) data;
1917          GLuint i;
1918
1919          for (i = 0; i < ARRAY_SIZE(shProg->_LinkedShaders); i++) {
1920             if (shProg->_LinkedShaders[i])
1921                destroy_program_variants(st, shProg->_LinkedShaders[i]->Program);
1922          }
1923       }
1924       break;
1925    case GL_VERTEX_SHADER:
1926    case GL_FRAGMENT_SHADER:
1927    case GL_GEOMETRY_SHADER:
1928    case GL_TESS_CONTROL_SHADER:
1929    case GL_TESS_EVALUATION_SHADER:
1930    case GL_COMPUTE_SHADER:
1931       break;
1932    default:
1933       assert(0);
1934    }
1935 }
1936
1937
1938 /**
1939  * Callback for _mesa_HashWalk.  Free all the program variants which match
1940  * the given context.
1941  */
1942 static void
1943 destroy_program_variants_cb(GLuint key, void *data, void *userData)
1944 {
1945    struct st_context *st = (struct st_context *) userData;
1946    struct gl_program *program = (struct gl_program *) data;
1947    destroy_program_variants(st, program);
1948 }
1949
1950
1951 /**
1952  * Walk over all shaders and programs to delete any variants which
1953  * belong to the given context.
1954  * This is called during context tear-down.
1955  */
1956 void
1957 st_destroy_program_variants(struct st_context *st)
1958 {
1959    /* If shaders can be shared with other contexts, the last context will
1960     * call DeleteProgram on all shaders, releasing everything.
1961     */
1962    if (st->has_shareable_shaders)
1963       return;
1964
1965    /* ARB vert/frag program */
1966    _mesa_HashWalk(st->ctx->Shared->Programs,
1967                   destroy_program_variants_cb, st);
1968
1969    /* GLSL vert/frag/geom shaders */
1970    _mesa_HashWalk(st->ctx->Shared->ShaderObjects,
1971                   destroy_shader_program_variants_cb, st);
1972 }
1973
1974
1975 /**
1976  * Compile one shader variant.
1977  */
1978 static void
1979 st_precompile_shader_variant(struct st_context *st,
1980                              struct gl_program *prog)
1981 {
1982    switch (prog->Target) {
1983    case GL_VERTEX_PROGRAM_ARB: {
1984       struct st_program *p = (struct st_program *)prog;
1985       struct st_common_variant_key key;
1986
1987       memset(&key, 0, sizeof(key));
1988
1989       key.st = st->has_shareable_shaders ? NULL : st;
1990       st_get_vp_variant(st, p, &key);
1991       break;
1992    }
1993
1994    case GL_FRAGMENT_PROGRAM_ARB: {
1995       struct st_program *p = (struct st_program *)prog;
1996       struct st_fp_variant_key key;
1997
1998       memset(&key, 0, sizeof(key));
1999
2000       key.st = st->has_shareable_shaders ? NULL : st;
2001       st_get_fp_variant(st, p, &key);
2002       break;
2003    }
2004
2005    case GL_TESS_CONTROL_PROGRAM_NV:
2006    case GL_TESS_EVALUATION_PROGRAM_NV:
2007    case GL_GEOMETRY_PROGRAM_NV:
2008    case GL_COMPUTE_PROGRAM_NV: {
2009       struct st_program *p = st_program(prog);
2010       struct st_common_variant_key key;
2011
2012       memset(&key, 0, sizeof(key));
2013
2014       key.st = st->has_shareable_shaders ? NULL : st;
2015       st_get_common_variant(st, p, &key);
2016       break;
2017    }
2018
2019    default:
2020       assert(0);
2021    }
2022 }
2023
2024 void
2025 st_serialize_nir(struct st_program *stp)
2026 {
2027    if (!stp->serialized_nir) {
2028       struct blob blob;
2029       size_t size;
2030
2031       blob_init(&blob);
2032       nir_serialize(&blob, stp->Base.nir, false);
2033       blob_finish_get_buffer(&blob, &stp->serialized_nir, &size);
2034       stp->serialized_nir_size = size;
2035    }
2036 }
2037
2038 void
2039 st_finalize_program(struct st_context *st, struct gl_program *prog)
2040 {
2041    if (st->current_program[prog->info.stage] == prog) {
2042       if (prog->info.stage == MESA_SHADER_VERTEX)
2043          st->dirty |= ST_NEW_VERTEX_PROGRAM(st, (struct st_program *)prog);
2044       else
2045          st->dirty |= ((struct st_program *)prog)->affected_states;
2046    }
2047
2048    if (prog->nir) {
2049       nir_sweep(prog->nir);
2050
2051       /* This is only needed for ARB_vp/fp programs and when the disk cache
2052        * is disabled. If the disk cache is enabled, GLSL programs are
2053        * serialized in write_nir_to_cache.
2054        */
2055       st_serialize_nir(st_program(prog));
2056    }
2057
2058    /* Create Gallium shaders now instead of on demand. */
2059    if (ST_DEBUG & DEBUG_PRECOMPILE ||
2060        st->shader_has_one_variant[prog->info.stage])
2061       st_precompile_shader_variant(st, prog);
2062 }