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/imports.h"
  35 #include "main/hash.h"
  36 #include "main/mtypes.h"
  37 #include "program/prog_parameter.h"
  38 #include "program/prog_print.h"
  39 #include "program/programopt.h"
  40
  41 #include "pipe/p_context.h"
  42 #include "pipe/p_defines.h"
  43 #include "pipe/p_shader_tokens.h"
  44 #include "draw/draw_context.h"
  45 #include "tgsi/tgsi_dump.h"
  46 #include "tgsi/tgsi_ureg.h"
  47
  48 #include "st_debug.h"
  49 #include "st_cb_bitmap.h"
  50 #include "st_cb_drawpixels.h"
  51 #include "st_context.h"
  52 #include "st_program.h"
  53 #include "st_mesa_to_tgsi.h"
  54 #include "cso_cache/cso_context.h"
  55
  56
  57
  58 /**
  59  * Delete a vertex program variant.  Note the caller must unlink
  60  * the variant from the linked list.
  61  */
  62 static void
  63 delete_vp_variant(struct st_context *st, struct st_vp_variant *vpv)
  64 {
  65    if (vpv->driver_shader)
  66       cso_delete_vertex_shader(st->cso_context, vpv->driver_shader);
  67
  68    if (vpv->draw_shader)
  69       draw_delete_vertex_shader( st->draw, vpv->draw_shader );
  70
  71    if (vpv->tgsi.tokens)
  72       ureg_free_tokens(vpv->tgsi.tokens);
  73
  74    free( vpv );
  75 }
  76
  77
  78
  79 /**
  80  * Clean out any old compilations:
  81  */
  82 void
  83 st_release_vp_variants( struct st_context *st,
  84                         struct st_vertex_program *stvp )
  85 {
  86    struct st_vp_variant *vpv;
  87
  88    for (vpv = stvp->variants; vpv; ) {
  89       struct st_vp_variant *next = vpv->next;
  90       delete_vp_variant(st, vpv);
  91       vpv = next;
  92    }
  93
  94    stvp->variants = NULL;
  95 }
  96
  97
  98
  99 /**
 100  * Delete a fragment program variant.  Note the caller must unlink
 101  * the variant from the linked list.
 102  */
 103 static void
 104 delete_fp_variant(struct st_context *st, struct st_fp_variant *fpv)
 105 {
 106    if (fpv->driver_shader)
 107       cso_delete_fragment_shader(st->cso_context, fpv->driver_shader);
 108    if (fpv->parameters)
 109       _mesa_free_parameter_list(fpv->parameters);
 110    if (fpv->tgsi.tokens)
 111       ureg_free_tokens(fpv->tgsi.tokens);
 112    free(fpv);
 113 }
 114
 115
 116 /**
 117  * Free all variants of a fragment program.
 118  */
 119 void
 120 st_release_fp_variants(struct st_context *st, struct st_fragment_program *stfp)
 121 {
 122    struct st_fp_variant *fpv;
 123
 124    for (fpv = stfp->variants; fpv; ) {
 125       struct st_fp_variant *next = fpv->next;
 126       delete_fp_variant(st, fpv);
 127       fpv = next;
 128    }
 129
 130    stfp->variants = NULL;
 131 }
 132
 133
 134 /**
 135  * Delete a geometry program variant.  Note the caller must unlink
 136  * the variant from the linked list.
 137  */
 138 static void
 139 delete_gp_variant(struct st_context *st, struct st_gp_variant *gpv)
 140 {
 141    if (gpv->driver_shader)
 142       cso_delete_geometry_shader(st->cso_context, gpv->driver_shader);
 143
 144    free(gpv);
 145 }
 146
 147
 148 /**
 149  * Free all variants of a geometry program.
 150  */
 151 void
 152 st_release_gp_variants(struct st_context *st, struct st_geometry_program *stgp)
 153 {
 154    struct st_gp_variant *gpv;
 155
 156    for (gpv = stgp->variants; gpv; ) {
 157       struct st_gp_variant *next = gpv->next;
 158       delete_gp_variant(st, gpv);
 159       gpv = next;
 160    }
 161
 162    stgp->variants = NULL;
 163 }
 164
 165
 166
 167
 168 /**
 169  * Translate a Mesa vertex shader into a TGSI shader.
 170  * \param outputMapping  to map vertex program output registers (VARYING_SLOT_x)
 171  *       to TGSI output slots
 172  * \param tokensOut  destination for TGSI tokens
 173  * \return  pointer to cached pipe_shader object.
 174  */
 175 void
 176 st_prepare_vertex_program(struct gl_context *ctx,
 177                             struct st_vertex_program *stvp)
 178 {
 179    struct st_context *st = st_context(ctx);
 180    GLuint attr;
 181
 182    stvp->num_inputs = 0;
 183    stvp->num_outputs = 0;
 184
 185    if (stvp->Base.IsPositionInvariant)
 186       _mesa_insert_mvp_code(ctx, &stvp->Base);
 187
 188    if (!stvp->glsl_to_tgsi)
 189       assert(stvp->Base.Base.NumInstructions > 1);
 190
 191    /*
 192     * Determine number of inputs, the mappings between VERT_ATTRIB_x
 193     * and TGSI generic input indexes, plus input attrib semantic info.
 194     */
 195    for (attr = 0; attr < VERT_ATTRIB_MAX; attr++) {
 196       if ((stvp->Base.Base.InputsRead & BITFIELD64_BIT(attr)) != 0) {
 197          stvp->input_to_index[attr] = stvp->num_inputs;
 198          stvp->index_to_input[stvp->num_inputs] = attr;
 199          stvp->num_inputs++;
 200       }
 201    }
 202    /* bit of a hack, presetup potentially unused edgeflag input */
 203    stvp->input_to_index[VERT_ATTRIB_EDGEFLAG] = stvp->num_inputs;
 204    stvp->index_to_input[stvp->num_inputs] = VERT_ATTRIB_EDGEFLAG;
 205
 206    /* Compute mapping of vertex program outputs to slots.
 207     */
 208    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 209       if ((stvp->Base.Base.OutputsWritten & BITFIELD64_BIT(attr)) == 0) {
 210          stvp->result_to_output[attr] = ~0;
 211       }
 212       else {
 213          unsigned slot = stvp->num_outputs++;
 214
 215          stvp->result_to_output[attr] = slot;
 216
 217          switch (attr) {
 218          case VARYING_SLOT_POS:
 219             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_POSITION;
 220             stvp->output_semantic_index[slot] = 0;
 221             break;
 222          case VARYING_SLOT_COL0:
 223             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 224             stvp->output_semantic_index[slot] = 0;
 225             break;
 226          case VARYING_SLOT_COL1:
 227             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 228             stvp->output_semantic_index[slot] = 1;
 229             break;
 230          case VARYING_SLOT_BFC0:
 231             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_BCOLOR;
 232             stvp->output_semantic_index[slot] = 0;
 233             break;
 234          case VARYING_SLOT_BFC1:
 235             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_BCOLOR;
 236             stvp->output_semantic_index[slot] = 1;
 237             break;
 238          case VARYING_SLOT_FOGC:
 239             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_FOG;
 240             stvp->output_semantic_index[slot] = 0;
 241             break;
 242          case VARYING_SLOT_PSIZ:
 243             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_PSIZE;
 244             stvp->output_semantic_index[slot] = 0;
 245             break;
 246          case VARYING_SLOT_CLIP_DIST0:
 247             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 248             stvp->output_semantic_index[slot] = 0;
 249             break;
 250          case VARYING_SLOT_CLIP_DIST1:
 251             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 252             stvp->output_semantic_index[slot] = 1;
 253             break;
 254          case VARYING_SLOT_EDGE:
 255             assert(0);
 256             break;
 257          case VARYING_SLOT_CLIP_VERTEX:
 258             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_CLIPVERTEX;
 259             stvp->output_semantic_index[slot] = 0;
 260             break;
 261          case VARYING_SLOT_LAYER:
 262             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_LAYER;
 263             stvp->output_semantic_index[slot] = 0;
 264             break;
 265          case VARYING_SLOT_VIEWPORT:
 266             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_VIEWPORT_INDEX;
 267             stvp->output_semantic_index[slot] = 0;
 268             break;
 269
 270          case VARYING_SLOT_TEX0:
 271          case VARYING_SLOT_TEX1:
 272          case VARYING_SLOT_TEX2:
 273          case VARYING_SLOT_TEX3:
 274          case VARYING_SLOT_TEX4:
 275          case VARYING_SLOT_TEX5:
 276          case VARYING_SLOT_TEX6:
 277          case VARYING_SLOT_TEX7:
 278             if (st->needs_texcoord_semantic) {
 279                stvp->output_semantic_name[slot] = TGSI_SEMANTIC_TEXCOORD;
 280                stvp->output_semantic_index[slot] = attr - VARYING_SLOT_TEX0;
 281                break;
 282             }
 283             /* fall through */
 284          case VARYING_SLOT_VAR0:
 285          default:
 286             assert(attr < VARYING_SLOT_MAX);
 287             stvp->output_semantic_name[slot] = TGSI_SEMANTIC_GENERIC;
 288             stvp->output_semantic_index[slot] =
 289                st_get_generic_varying_index(st, attr);
 290             break;
 291          }
 292       }
 293    }
 294    /* similar hack to above, presetup potentially unused edgeflag output */
 295    stvp->result_to_output[VARYING_SLOT_EDGE] = stvp->num_outputs;
 296    stvp->output_semantic_name[stvp->num_outputs] = TGSI_SEMANTIC_EDGEFLAG;
 297    stvp->output_semantic_index[stvp->num_outputs] = 0;
 298 }
 299
 300
 301 /**
 302  * Translate a vertex program to create a new variant.
 303  */
 304 static struct st_vp_variant *
 305 st_translate_vertex_program(struct st_context *st,
 306                             struct st_vertex_program *stvp,
 307                             const struct st_vp_variant_key *key)
 308 {
 309    struct st_vp_variant *vpv = CALLOC_STRUCT(st_vp_variant);
 310    struct pipe_context *pipe = st->pipe;
 311    struct ureg_program *ureg;
 312    enum pipe_error error;
 313    unsigned num_outputs;
 314
 315    st_prepare_vertex_program(st->ctx, stvp);
 316
 317    if (!stvp->glsl_to_tgsi)
 318    {
 319       _mesa_remove_output_reads(&stvp->Base.Base, PROGRAM_OUTPUT);
 320    }
 321
 322    ureg = ureg_create( TGSI_PROCESSOR_VERTEX );
 323    if (ureg == NULL) {
 324       free(vpv);
 325       return NULL;
 326    }
 327
 328    vpv->key = *key;
 329
 330    vpv->num_inputs = stvp->num_inputs;
 331    num_outputs = stvp->num_outputs;
 332    if (key->passthrough_edgeflags) {
 333       vpv->num_inputs++;
 334       num_outputs++;
 335    }
 336
 337    if (ST_DEBUG & DEBUG_MESA) {
 338       _mesa_print_program(&stvp->Base.Base);
 339       _mesa_print_program_parameters(st->ctx, &stvp->Base.Base);
 340       debug_printf("\n");
 341    }
 342
 343    if (stvp->glsl_to_tgsi)
 344       error = st_translate_program(st->ctx,
 345                                    TGSI_PROCESSOR_VERTEX,
 346                                    ureg,
 347                                    stvp->glsl_to_tgsi,
 348                                    &stvp->Base.Base,
 349                                    /* inputs */
 350                                    vpv->num_inputs,
 351                                    stvp->input_to_index,
 352                                    NULL, /* input semantic name */
 353                                    NULL, /* input semantic index */
 354                                    NULL, /* interp mode */
 355                                    NULL, /* interp location */
 356                                    /* outputs */
 357                                    num_outputs,
 358                                    stvp->result_to_output,
 359                                    stvp->output_semantic_name,
 360                                    stvp->output_semantic_index,
 361                                    key->passthrough_edgeflags,
 362                                    key->clamp_color);
 363    else
 364       error = st_translate_mesa_program(st->ctx,
 365                                         TGSI_PROCESSOR_VERTEX,
 366                                         ureg,
 367                                         &stvp->Base.Base,
 368                                         /* inputs */
 369                                         vpv->num_inputs,
 370                                         stvp->input_to_index,
 371                                         NULL, /* input semantic name */
 372                                         NULL, /* input semantic index */
 373                                         NULL,
 374                                         /* outputs */
 375                                         num_outputs,
 376                                         stvp->result_to_output,
 377                                         stvp->output_semantic_name,
 378                                         stvp->output_semantic_index,
 379                                         key->passthrough_edgeflags,
 380                                         key->clamp_color);
 381
 382    if (error)
 383       goto fail;
 384
 385    vpv->tgsi.tokens = ureg_get_tokens( ureg, NULL );
 386    if (!vpv->tgsi.tokens)
 387       goto fail;
 388
 389    ureg_destroy( ureg );
 390
 391    if (stvp->glsl_to_tgsi) {
 392       st_translate_stream_output_info(stvp->glsl_to_tgsi,
 393                                       stvp->result_to_output,
 394                                       &vpv->tgsi.stream_output);
 395    }
 396
 397    if (ST_DEBUG & DEBUG_TGSI) {
 398       tgsi_dump(vpv->tgsi.tokens, 0);
 399       debug_printf("\n");
 400    }
 401
 402    vpv->driver_shader = pipe->create_vs_state(pipe, &vpv->tgsi);
 403    return vpv;
 404
 405 fail:
 406    debug_printf("%s: failed to translate Mesa program:\n", __FUNCTION__);
 407    _mesa_print_program(&stvp->Base.Base);
 408    debug_assert(0);
 409
 410    ureg_destroy( ureg );
 411    return NULL;
 412 }
 413
 414
 415 /**
 416  * Find/create a vertex program variant.
 417  */
 418 struct st_vp_variant *
 419 st_get_vp_variant(struct st_context *st,
 420                   struct st_vertex_program *stvp,
 421                   const struct st_vp_variant_key *key)
 422 {
 423    struct st_vp_variant *vpv;
 424
 425    /* Search for existing variant */
 426    for (vpv = stvp->variants; vpv; vpv = vpv->next) {
 427       if (memcmp(&vpv->key, key, sizeof(*key)) == 0) {
 428          break;
 429       }
 430    }
 431
 432    if (!vpv) {
 433       /* create now */
 434       vpv = st_translate_vertex_program(st, stvp, key);
 435       if (vpv) {
 436          /* insert into list */
 437          vpv->next = stvp->variants;
 438          stvp->variants = vpv;
 439       }
 440    }
 441
 442    return vpv;
 443 }
 444
 445
 446 static unsigned
 447 st_translate_interp(enum glsl_interp_qualifier glsl_qual, bool is_color)
 448 {
 449    switch (glsl_qual) {
 450    case INTERP_QUALIFIER_NONE:
 451       if (is_color)
 452          return TGSI_INTERPOLATE_COLOR;
 453       return TGSI_INTERPOLATE_PERSPECTIVE;
 454    case INTERP_QUALIFIER_SMOOTH:
 455       return TGSI_INTERPOLATE_PERSPECTIVE;
 456    case INTERP_QUALIFIER_FLAT:
 457       return TGSI_INTERPOLATE_CONSTANT;
 458    case INTERP_QUALIFIER_NOPERSPECTIVE:
 459       return TGSI_INTERPOLATE_LINEAR;
 460    default:
 461       assert(0 && "unexpected interp mode in st_translate_interp()");
 462       return TGSI_INTERPOLATE_PERSPECTIVE;
 463    }
 464 }
 465
 466
 467 /**
 468  * Translate a Mesa fragment shader into a TGSI shader using extra info in
 469  * the key.
 470  * \return  new fragment program variant
 471  */
 472 static struct st_fp_variant *
 473 st_translate_fragment_program(struct st_context *st,
 474                               struct st_fragment_program *stfp,
 475                               const struct st_fp_variant_key *key)
 476 {
 477    struct pipe_context *pipe = st->pipe;
 478    struct st_fp_variant *variant = CALLOC_STRUCT(st_fp_variant);
 479    GLboolean deleteFP = GL_FALSE;
 480
 481    GLuint outputMapping[FRAG_RESULT_MAX];
 482    GLuint inputMapping[VARYING_SLOT_MAX];
 483    GLuint interpMode[PIPE_MAX_SHADER_INPUTS];  /* XXX size? */
 484    GLuint interpLocation[PIPE_MAX_SHADER_INPUTS];
 485    GLuint attr;
 486    GLbitfield64 inputsRead;
 487    struct ureg_program *ureg;
 488
 489    GLboolean write_all = GL_FALSE;
 490
 491    ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS];
 492    ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
 493    uint fs_num_inputs = 0;
 494
 495    ubyte fs_output_semantic_name[PIPE_MAX_SHADER_OUTPUTS];
 496    ubyte fs_output_semantic_index[PIPE_MAX_SHADER_OUTPUTS];
 497    uint fs_num_outputs = 0;
 498
 499    if (!variant)
 500       return NULL;
 501
 502    assert(!(key->bitmap && key->drawpixels));
 503
 504    if (key->bitmap) {
 505       /* glBitmap drawing */
 506       struct gl_fragment_program *fp; /* we free this temp program below */
 507
 508       st_make_bitmap_fragment_program(st, &stfp->Base,
 509                                       &fp, &variant->bitmap_sampler);
 510
 511       variant->parameters = _mesa_clone_parameter_list(fp->Base.Parameters);
 512       stfp = st_fragment_program(fp);
 513       deleteFP = GL_TRUE;
 514    }
 515    else if (key->drawpixels) {
 516       /* glDrawPixels drawing */
 517       struct gl_fragment_program *fp; /* we free this temp program below */
 518
 519       if (key->drawpixels_z || key->drawpixels_stencil) {
 520          fp = st_make_drawpix_z_stencil_program(st, key->drawpixels_z,
 521                                                 key->drawpixels_stencil);
 522       }
 523       else {
 524          /* RGBA */
 525          st_make_drawpix_fragment_program(st, &stfp->Base, &fp);
 526          variant->parameters = _mesa_clone_parameter_list(fp->Base.Parameters);
 527          deleteFP = GL_TRUE;
 528       }
 529       stfp = st_fragment_program(fp);
 530    }
 531
 532    if (!stfp->glsl_to_tgsi)
 533       _mesa_remove_output_reads(&stfp->Base.Base, PROGRAM_OUTPUT);
 534
 535    /*
 536     * Convert Mesa program inputs to TGSI input register semantics.
 537     */
 538    inputsRead = stfp->Base.Base.InputsRead;
 539    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 540       if ((inputsRead & BITFIELD64_BIT(attr)) != 0) {
 541          const GLuint slot = fs_num_inputs++;
 542
 543          inputMapping[attr] = slot;
 544          if (stfp->Base.IsCentroid & BITFIELD64_BIT(attr))
 545             interpLocation[slot] = TGSI_INTERPOLATE_LOC_CENTROID;
 546          else if (stfp->Base.IsSample & BITFIELD64_BIT(attr))
 547             interpLocation[slot] = TGSI_INTERPOLATE_LOC_SAMPLE;
 548          else
 549             interpLocation[slot] = TGSI_INTERPOLATE_LOC_CENTER;
 550
 551          if (key->persample_shading)
 552             interpLocation[slot] = TGSI_INTERPOLATE_LOC_SAMPLE;
 553
 554          switch (attr) {
 555          case VARYING_SLOT_POS:
 556             input_semantic_name[slot] = TGSI_SEMANTIC_POSITION;
 557             input_semantic_index[slot] = 0;
 558             interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
 559             break;
 560          case VARYING_SLOT_COL0:
 561             input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 562             input_semantic_index[slot] = 0;
 563             interpMode[slot] = st_translate_interp(stfp->Base.InterpQualifier[attr],
 564                                                    TRUE);
 565             break;
 566          case VARYING_SLOT_COL1:
 567             input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 568             input_semantic_index[slot] = 1;
 569             interpMode[slot] = st_translate_interp(stfp->Base.InterpQualifier[attr],
 570                                                    TRUE);
 571             break;
 572          case VARYING_SLOT_FOGC:
 573             input_semantic_name[slot] = TGSI_SEMANTIC_FOG;
 574             input_semantic_index[slot] = 0;
 575             interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
 576             break;
 577          case VARYING_SLOT_FACE:
 578             input_semantic_name[slot] = TGSI_SEMANTIC_FACE;
 579             input_semantic_index[slot] = 0;
 580             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 581             break;
 582          case VARYING_SLOT_PRIMITIVE_ID:
 583             input_semantic_name[slot] = TGSI_SEMANTIC_PRIMID;
 584             input_semantic_index[slot] = 0;
 585             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 586             break;
 587          case VARYING_SLOT_LAYER:
 588             input_semantic_name[slot] = TGSI_SEMANTIC_LAYER;
 589             input_semantic_index[slot] = 0;
 590             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 591             break;
 592          case VARYING_SLOT_VIEWPORT:
 593             input_semantic_name[slot] = TGSI_SEMANTIC_VIEWPORT_INDEX;
 594             input_semantic_index[slot] = 0;
 595             interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
 596             break;
 597          case VARYING_SLOT_CLIP_DIST0:
 598             input_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 599             input_semantic_index[slot] = 0;
 600             interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
 601             break;
 602          case VARYING_SLOT_CLIP_DIST1:
 603             input_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 604             input_semantic_index[slot] = 1;
 605             interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
 606             break;
 607             /* In most cases, there is nothing special about these
 608              * inputs, so adopt a convention to use the generic
 609              * semantic name and the mesa VARYING_SLOT_ number as the
 610              * index.
 611              *
 612              * All that is required is that the vertex shader labels
 613              * its own outputs similarly, and that the vertex shader
 614              * generates at least every output required by the
 615              * fragment shader plus fixed-function hardware (such as
 616              * BFC).
 617              *
 618              * However, some drivers may need us to identify the PNTC and TEXi
 619              * varyings if, for example, their capability to replace them with
 620              * sprite coordinates is limited.
 621              */
 622          case VARYING_SLOT_PNTC:
 623             if (st->needs_texcoord_semantic) {
 624                input_semantic_name[slot] = TGSI_SEMANTIC_PCOORD;
 625                input_semantic_index[slot] = 0;
 626                interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
 627                break;
 628             }
 629             /* fall through */
 630          case VARYING_SLOT_TEX0:
 631          case VARYING_SLOT_TEX1:
 632          case VARYING_SLOT_TEX2:
 633          case VARYING_SLOT_TEX3:
 634          case VARYING_SLOT_TEX4:
 635          case VARYING_SLOT_TEX5:
 636          case VARYING_SLOT_TEX6:
 637          case VARYING_SLOT_TEX7:
 638             if (st->needs_texcoord_semantic) {
 639                input_semantic_name[slot] = TGSI_SEMANTIC_TEXCOORD;
 640                input_semantic_index[slot] = attr - VARYING_SLOT_TEX0;
 641                interpMode[slot] =
 642                   st_translate_interp(stfp->Base.InterpQualifier[attr], FALSE);
 643                break;
 644             }
 645             /* fall through */
 646          case VARYING_SLOT_VAR0:
 647          default:
 648             /* Semantic indices should be zero-based because drivers may choose
 649              * to assign a fixed slot determined by that index.
 650              * This is useful because ARB_separate_shader_objects uses location
 651              * qualifiers for linkage, and if the semantic index corresponds to
 652              * these locations, linkage passes in the driver become unecessary.
 653              *
 654              * If needs_texcoord_semantic is true, no semantic indices will be
 655              * consumed for the TEXi varyings, and we can base the locations of
 656              * the user varyings on VAR0.  Otherwise, we use TEX0 as base index.
 657              */
 658             assert(attr >= VARYING_SLOT_TEX0);
 659             input_semantic_name[slot] = TGSI_SEMANTIC_GENERIC;
 660             input_semantic_index[slot] = st_get_generic_varying_index(st, attr);
 661             if (attr == VARYING_SLOT_PNTC)
 662                interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
 663             else
 664                interpMode[slot] = st_translate_interp(stfp->Base.InterpQualifier[attr],
 665                                                       FALSE);
 666             break;
 667          }
 668       }
 669       else {
 670          inputMapping[attr] = -1;
 671       }
 672    }
 673
 674    /*
 675     * Semantics and mapping for outputs
 676     */
 677    {
 678       uint numColors = 0;
 679       GLbitfield64 outputsWritten = stfp->Base.Base.OutputsWritten;
 680
 681       /* if z is written, emit that first */
 682       if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
 683          fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_POSITION;
 684          fs_output_semantic_index[fs_num_outputs] = 0;
 685          outputMapping[FRAG_RESULT_DEPTH] = fs_num_outputs;
 686          fs_num_outputs++;
 687          outputsWritten &= ~(1 << FRAG_RESULT_DEPTH);
 688       }
 689
 690       if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_STENCIL)) {
 691          fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_STENCIL;
 692          fs_output_semantic_index[fs_num_outputs] = 0;
 693          outputMapping[FRAG_RESULT_STENCIL] = fs_num_outputs;
 694          fs_num_outputs++;
 695          outputsWritten &= ~(1 << FRAG_RESULT_STENCIL);
 696       }
 697
 698       if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK)) {
 699          fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_SAMPLEMASK;
 700          fs_output_semantic_index[fs_num_outputs] = 0;
 701          outputMapping[FRAG_RESULT_SAMPLE_MASK] = fs_num_outputs;
 702          fs_num_outputs++;
 703          outputsWritten &= ~(1 << FRAG_RESULT_SAMPLE_MASK);
 704       }
 705
 706       /* handle remaining outputs (color) */
 707       for (attr = 0; attr < FRAG_RESULT_MAX; attr++) {
 708          if (outputsWritten & BITFIELD64_BIT(attr)) {
 709             switch (attr) {
 710             case FRAG_RESULT_DEPTH:
 711             case FRAG_RESULT_STENCIL:
 712             case FRAG_RESULT_SAMPLE_MASK:
 713                /* handled above */
 714                assert(0);
 715                break;
 716             case FRAG_RESULT_COLOR:
 717                write_all = GL_TRUE; /* fallthrough */
 718             default:
 719                assert(attr == FRAG_RESULT_COLOR ||
 720                       (FRAG_RESULT_DATA0 <= attr && attr < FRAG_RESULT_MAX));
 721                fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_COLOR;
 722                fs_output_semantic_index[fs_num_outputs] = numColors;
 723                outputMapping[attr] = fs_num_outputs;
 724                numColors++;
 725                break;
 726             }
 727
 728             fs_num_outputs++;
 729          }
 730       }
 731    }
 732
 733    ureg = ureg_create( TGSI_PROCESSOR_FRAGMENT );
 734    if (ureg == NULL) {
 735       free(variant);
 736       return NULL;
 737    }
 738
 739    if (ST_DEBUG & DEBUG_MESA) {
 740       _mesa_print_program(&stfp->Base.Base);
 741       _mesa_print_program_parameters(st->ctx, &stfp->Base.Base);
 742       debug_printf("\n");
 743    }
 744    if (write_all == GL_TRUE)
 745       ureg_property(ureg, TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS, 1);
 746
 747    if (stfp->Base.FragDepthLayout != FRAG_DEPTH_LAYOUT_NONE) {
 748       switch (stfp->Base.FragDepthLayout) {
 749       case FRAG_DEPTH_LAYOUT_ANY:
 750          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
 751                        TGSI_FS_DEPTH_LAYOUT_ANY);
 752          break;
 753       case FRAG_DEPTH_LAYOUT_GREATER:
 754          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
 755                        TGSI_FS_DEPTH_LAYOUT_GREATER);
 756          break;
 757       case FRAG_DEPTH_LAYOUT_LESS:
 758          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
 759                        TGSI_FS_DEPTH_LAYOUT_LESS);
 760          break;
 761       case FRAG_DEPTH_LAYOUT_UNCHANGED:
 762          ureg_property(ureg, TGSI_PROPERTY_FS_DEPTH_LAYOUT,
 763                        TGSI_FS_DEPTH_LAYOUT_UNCHANGED);
 764          break;
 765       default:
 766          assert(0);
 767       }
 768    }
 769
 770    if (stfp->glsl_to_tgsi)
 771       st_translate_program(st->ctx,
 772                            TGSI_PROCESSOR_FRAGMENT,
 773                            ureg,
 774                            stfp->glsl_to_tgsi,
 775                            &stfp->Base.Base,
 776                            /* inputs */
 777                            fs_num_inputs,
 778                            inputMapping,
 779                            input_semantic_name,
 780                            input_semantic_index,
 781                            interpMode,
 782                            interpLocation,
 783                            /* outputs */
 784                            fs_num_outputs,
 785                            outputMapping,
 786                            fs_output_semantic_name,
 787                            fs_output_semantic_index, FALSE,
 788                            key->clamp_color );
 789    else
 790       st_translate_mesa_program(st->ctx,
 791                                 TGSI_PROCESSOR_FRAGMENT,
 792                                 ureg,
 793                                 &stfp->Base.Base,
 794                                 /* inputs */
 795                                 fs_num_inputs,
 796                                 inputMapping,
 797                                 input_semantic_name,
 798                                 input_semantic_index,
 799                                 interpMode,
 800                                 /* outputs */
 801                                 fs_num_outputs,
 802                                 outputMapping,
 803                                 fs_output_semantic_name,
 804                                 fs_output_semantic_index, FALSE,
 805                                 key->clamp_color);
 806
 807    variant->tgsi.tokens = ureg_get_tokens( ureg, NULL );
 808    ureg_destroy( ureg );
 809
 810    if (ST_DEBUG & DEBUG_TGSI) {
 811       tgsi_dump(variant->tgsi.tokens, 0/*TGSI_DUMP_VERBOSE*/);
 812       debug_printf("\n");
 813    }
 814
 815    /* fill in variant */
 816    variant->driver_shader = pipe->create_fs_state(pipe, &variant->tgsi);
 817    variant->key = *key;
 818
 819    if (deleteFP) {
 820       /* Free the temporary program made above */
 821       struct gl_fragment_program *fp = &stfp->Base;
 822       _mesa_reference_fragprog(st->ctx, &fp, NULL);
 823    }
 824
 825    return variant;
 826 }
 827
 828
 829 /**
 830  * Translate fragment program if needed.
 831  */
 832 struct st_fp_variant *
 833 st_get_fp_variant(struct st_context *st,
 834                   struct st_fragment_program *stfp,
 835                   const struct st_fp_variant_key *key)
 836 {
 837    struct st_fp_variant *fpv;
 838
 839    /* Search for existing variant */
 840    for (fpv = stfp->variants; fpv; fpv = fpv->next) {
 841       if (memcmp(&fpv->key, key, sizeof(*key)) == 0) {
 842          break;
 843       }
 844    }
 845
 846    if (!fpv) {
 847       /* create new */
 848       fpv = st_translate_fragment_program(st, stfp, key);
 849       if (fpv) {
 850          /* insert into list */
 851          fpv->next = stfp->variants;
 852          stfp->variants = fpv;
 853       }
 854    }
 855
 856    return fpv;
 857 }
 858
 859
 860 /**
 861  * Translate a geometry program to create a new variant.
 862  */
 863 static struct st_gp_variant *
 864 st_translate_geometry_program(struct st_context *st,
 865                               struct st_geometry_program *stgp,
 866                               const struct st_gp_variant_key *key)
 867 {
 868    GLuint inputMapping[VARYING_SLOT_MAX];
 869    GLuint outputMapping[VARYING_SLOT_MAX];
 870    struct pipe_context *pipe = st->pipe;
 871    GLuint attr;
 872    GLbitfield64 inputsRead;
 873
 874    uint gs_num_inputs = 0;
 875    uint gs_builtin_inputs = 0;
 876    uint gs_array_offset = 0;
 877
 878    ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS];
 879    ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
 880
 881    ubyte gs_output_semantic_name[PIPE_MAX_SHADER_OUTPUTS];
 882    ubyte gs_output_semantic_index[PIPE_MAX_SHADER_OUTPUTS];
 883    uint gs_num_outputs = 0;
 884
 885    GLint i;
 886    GLuint maxSlot = 0;
 887    struct ureg_program *ureg;
 888    struct pipe_shader_state state = {0};
 889
 890    struct st_gp_variant *gpv;
 891
 892    gpv = CALLOC_STRUCT(st_gp_variant);
 893    if (!gpv)
 894       return NULL;
 895
 896    if (!stgp->glsl_to_tgsi) {
 897       _mesa_remove_output_reads(&stgp->Base.Base, PROGRAM_OUTPUT);
 898    }
 899
 900    ureg = ureg_create( TGSI_PROCESSOR_GEOMETRY );
 901    if (ureg == NULL) {
 902       free(gpv);
 903       return NULL;
 904    }
 905
 906    memset(inputMapping, 0, sizeof(inputMapping));
 907    memset(outputMapping, 0, sizeof(outputMapping));
 908
 909    /*
 910     * Convert Mesa program inputs to TGSI input register semantics.
 911     */
 912    inputsRead = stgp->Base.Base.InputsRead;
 913    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
 914       if ((inputsRead & BITFIELD64_BIT(attr)) != 0) {
 915          const GLuint slot = gs_num_inputs;
 916
 917          gs_num_inputs++;
 918
 919          inputMapping[attr] = slot;
 920
 921          if (attr != VARYING_SLOT_PRIMITIVE_ID) {
 922             gs_array_offset += 2;
 923          } else
 924             ++gs_builtin_inputs;
 925
 926 #if 0
 927          debug_printf("input map at %d = %d\n",
 928                       slot + gs_array_offset, stgp->input_map[slot + gs_array_offset]);
 929 #endif
 930
 931          switch (attr) {
 932          case VARYING_SLOT_PRIMITIVE_ID:
 933             input_semantic_name[slot] = TGSI_SEMANTIC_PRIMID;
 934             input_semantic_index[slot] = 0;
 935             break;
 936          case VARYING_SLOT_POS:
 937             input_semantic_name[slot] = TGSI_SEMANTIC_POSITION;
 938             input_semantic_index[slot] = 0;
 939             break;
 940          case VARYING_SLOT_COL0:
 941             input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 942             input_semantic_index[slot] = 0;
 943             break;
 944          case VARYING_SLOT_COL1:
 945             input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
 946             input_semantic_index[slot] = 1;
 947             break;
 948          case VARYING_SLOT_FOGC:
 949             input_semantic_name[slot] = TGSI_SEMANTIC_FOG;
 950             input_semantic_index[slot] = 0;
 951             break;
 952          case VARYING_SLOT_CLIP_VERTEX:
 953             input_semantic_name[slot] = TGSI_SEMANTIC_CLIPVERTEX;
 954             input_semantic_index[slot] = 0;
 955             break;
 956          case VARYING_SLOT_CLIP_DIST0:
 957             input_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 958             input_semantic_index[slot] = 0;
 959             break;
 960          case VARYING_SLOT_CLIP_DIST1:
 961             input_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
 962             input_semantic_index[slot] = 1;
 963             break;
 964          case VARYING_SLOT_PSIZ:
 965             input_semantic_name[slot] = TGSI_SEMANTIC_PSIZE;
 966             input_semantic_index[slot] = 0;
 967             break;
 968          case VARYING_SLOT_TEX0:
 969          case VARYING_SLOT_TEX1:
 970          case VARYING_SLOT_TEX2:
 971          case VARYING_SLOT_TEX3:
 972          case VARYING_SLOT_TEX4:
 973          case VARYING_SLOT_TEX5:
 974          case VARYING_SLOT_TEX6:
 975          case VARYING_SLOT_TEX7:
 976             if (st->needs_texcoord_semantic) {
 977                input_semantic_name[slot] = TGSI_SEMANTIC_TEXCOORD;
 978                input_semantic_index[slot] = attr - VARYING_SLOT_TEX0;
 979                break;
 980             }
 981             /* fall through */
 982          case VARYING_SLOT_VAR0:
 983          default:
 984             assert(attr >= VARYING_SLOT_VAR0 && attr < VARYING_SLOT_MAX);
 985             input_semantic_name[slot] = TGSI_SEMANTIC_GENERIC;
 986             input_semantic_index[slot] =
 987                st_get_generic_varying_index(st, attr);
 988          break;
 989          }
 990       }
 991    }
 992
 993    /* initialize output semantics to defaults */
 994    for (i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 995       gs_output_semantic_name[i] = TGSI_SEMANTIC_GENERIC;
 996       gs_output_semantic_index[i] = 0;
 997    }
 998
 999    /*
1000     * Determine number of outputs, the (default) output register
1001     * mapping and the semantic information for each output.
1002     */
1003    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
1004       if (stgp->Base.Base.OutputsWritten & BITFIELD64_BIT(attr)) {
1005          GLuint slot;
1006
1007          slot = gs_num_outputs;
1008          gs_num_outputs++;
1009          outputMapping[attr] = slot;
1010
1011          switch (attr) {
1012          case VARYING_SLOT_POS:
1013             assert(slot == 0);
1014             gs_output_semantic_name[slot] = TGSI_SEMANTIC_POSITION;
1015             gs_output_semantic_index[slot] = 0;
1016             break;
1017          case VARYING_SLOT_COL0:
1018             gs_output_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
1019             gs_output_semantic_index[slot] = 0;
1020             break;
1021          case VARYING_SLOT_COL1:
1022             gs_output_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
1023             gs_output_semantic_index[slot] = 1;
1024             break;
1025          case VARYING_SLOT_BFC0:
1026             gs_output_semantic_name[slot] = TGSI_SEMANTIC_BCOLOR;
1027             gs_output_semantic_index[slot] = 0;
1028             break;
1029          case VARYING_SLOT_BFC1:
1030             gs_output_semantic_name[slot] = TGSI_SEMANTIC_BCOLOR;
1031             gs_output_semantic_index[slot] = 1;
1032             break;
1033          case VARYING_SLOT_FOGC:
1034             gs_output_semantic_name[slot] = TGSI_SEMANTIC_FOG;
1035             gs_output_semantic_index[slot] = 0;
1036             break;
1037          case VARYING_SLOT_PSIZ:
1038             gs_output_semantic_name[slot] = TGSI_SEMANTIC_PSIZE;
1039             gs_output_semantic_index[slot] = 0;
1040             break;
1041          case VARYING_SLOT_CLIP_VERTEX:
1042             gs_output_semantic_name[slot] = TGSI_SEMANTIC_CLIPVERTEX;
1043             gs_output_semantic_index[slot] = 0;
1044             break;
1045          case VARYING_SLOT_CLIP_DIST0:
1046             gs_output_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
1047             gs_output_semantic_index[slot] = 0;
1048             break;
1049          case VARYING_SLOT_CLIP_DIST1:
1050             gs_output_semantic_name[slot] = TGSI_SEMANTIC_CLIPDIST;
1051             gs_output_semantic_index[slot] = 1;
1052             break;
1053          case VARYING_SLOT_LAYER:
1054             gs_output_semantic_name[slot] = TGSI_SEMANTIC_LAYER;
1055             gs_output_semantic_index[slot] = 0;
1056             break;
1057          case VARYING_SLOT_PRIMITIVE_ID:
1058             gs_output_semantic_name[slot] = TGSI_SEMANTIC_PRIMID;
1059             gs_output_semantic_index[slot] = 0;
1060             break;
1061          case VARYING_SLOT_VIEWPORT:
1062             gs_output_semantic_name[slot] = TGSI_SEMANTIC_VIEWPORT_INDEX;
1063             gs_output_semantic_index[slot] = 0;
1064             break;
1065          case VARYING_SLOT_TEX0:
1066          case VARYING_SLOT_TEX1:
1067          case VARYING_SLOT_TEX2:
1068          case VARYING_SLOT_TEX3:
1069          case VARYING_SLOT_TEX4:
1070          case VARYING_SLOT_TEX5:
1071          case VARYING_SLOT_TEX6:
1072          case VARYING_SLOT_TEX7:
1073             if (st->needs_texcoord_semantic) {
1074                gs_output_semantic_name[slot] = TGSI_SEMANTIC_TEXCOORD;
1075                gs_output_semantic_index[slot] = attr - VARYING_SLOT_TEX0;
1076                break;
1077             }
1078             /* fall through */
1079          case VARYING_SLOT_VAR0:
1080          default:
1081             assert(slot < Elements(gs_output_semantic_name));
1082             assert(attr >= VARYING_SLOT_VAR0);
1083             gs_output_semantic_name[slot] = TGSI_SEMANTIC_GENERIC;
1084             gs_output_semantic_index[slot] =
1085                st_get_generic_varying_index(st, attr);
1086          break;
1087          }
1088       }
1089    }
1090
1091    /* find max output slot referenced to compute gs_num_outputs */
1092    for (attr = 0; attr < VARYING_SLOT_MAX; attr++) {
1093       if (outputMapping[attr] != ~0U && outputMapping[attr] > maxSlot)
1094          maxSlot = outputMapping[attr];
1095    }
1096    gs_num_outputs = maxSlot + 1;
1097
1098 #if 0 /* debug */
1099    {
1100       GLuint i;
1101       printf("outputMapping? %d\n", outputMapping ? 1 : 0);
1102       if (outputMapping) {
1103          printf("attr -> slot\n");
1104          for (i = 0; i < 16;  i++) {
1105             printf(" %2d       %3d\n", i, outputMapping[i]);
1106          }
1107       }
1108       printf("slot    sem_name  sem_index\n");
1109       for (i = 0; i < gs_num_outputs; i++) {
1110          printf(" %2d         %d         %d\n",
1111                 i,
1112                 gs_output_semantic_name[i],
1113                 gs_output_semantic_index[i]);
1114       }
1115    }
1116 #endif
1117
1118    ureg_property(ureg, TGSI_PROPERTY_GS_INPUT_PRIM, stgp->Base.InputType);
1119    ureg_property(ureg, TGSI_PROPERTY_GS_OUTPUT_PRIM, stgp->Base.OutputType);
1120    ureg_property(ureg, TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES,
1121                  stgp->Base.VerticesOut);
1122    ureg_property(ureg, TGSI_PROPERTY_GS_INVOCATIONS, stgp->Base.Invocations);
1123
1124    if (stgp->glsl_to_tgsi)
1125       st_translate_program(st->ctx,
1126                            TGSI_PROCESSOR_GEOMETRY,
1127                            ureg,
1128                            stgp->glsl_to_tgsi,
1129                            &stgp->Base.Base,
1130                            /* inputs */
1131                            gs_num_inputs,
1132                            inputMapping,
1133                            input_semantic_name,
1134                            input_semantic_index,
1135                            NULL,
1136                            NULL,
1137                            /* outputs */
1138                            gs_num_outputs,
1139                            outputMapping,
1140                            gs_output_semantic_name,
1141                            gs_output_semantic_index,
1142                            FALSE,
1143                            FALSE);
1144    else
1145       st_translate_mesa_program(st->ctx,
1146                                 TGSI_PROCESSOR_GEOMETRY,
1147                                 ureg,
1148                                 &stgp->Base.Base,
1149                                 /* inputs */
1150                                 gs_num_inputs,
1151                                 inputMapping,
1152                                 input_semantic_name,
1153                                 input_semantic_index,
1154                                 NULL,
1155                                 /* outputs */
1156                                 gs_num_outputs,
1157                                 outputMapping,
1158                                 gs_output_semantic_name,
1159                                 gs_output_semantic_index,
1160                                 FALSE,
1161                                 FALSE);
1162
1163    state.tokens = ureg_get_tokens( ureg, NULL );
1164    ureg_destroy( ureg );
1165
1166    if (stgp->glsl_to_tgsi) {
1167       st_translate_stream_output_info(stgp->glsl_to_tgsi,
1168                                       outputMapping,
1169                                       &state.stream_output);
1170    }
1171
1172    if ((ST_DEBUG & DEBUG_TGSI) && (ST_DEBUG & DEBUG_MESA)) {
1173       _mesa_print_program(&stgp->Base.Base);
1174       debug_printf("\n");
1175    }
1176
1177    if (ST_DEBUG & DEBUG_TGSI) {
1178       tgsi_dump(state.tokens, 0);
1179       debug_printf("\n");
1180    }
1181
1182    /* fill in new variant */
1183    gpv->driver_shader = pipe->create_gs_state(pipe, &state);
1184    gpv->key = *key;
1185
1186    ureg_free_tokens(state.tokens);
1187    return gpv;
1188 }
1189
1190
1191 /**
1192  * Get/create geometry program variant.
1193  */
1194 struct st_gp_variant *
1195 st_get_gp_variant(struct st_context *st,
1196                   struct st_geometry_program *stgp,
1197                   const struct st_gp_variant_key *key)
1198 {
1199    struct st_gp_variant *gpv;
1200
1201    /* Search for existing variant */
1202    for (gpv = stgp->variants; gpv; gpv = gpv->next) {
1203       if (memcmp(&gpv->key, key, sizeof(*key)) == 0) {
1204          break;
1205       }
1206    }
1207
1208    if (!gpv) {
1209       /* create new */
1210       gpv = st_translate_geometry_program(st, stgp, key);
1211       if (gpv) {
1212          /* insert into list */
1213          gpv->next = stgp->variants;
1214          stgp->variants = gpv;
1215       }
1216    }
1217
1218    return gpv;
1219 }
1220
1221
1222
1223
1224 /**
1225  * Debug- print current shader text
1226  */
1227 void
1228 st_print_shaders(struct gl_context *ctx)
1229 {
1230    struct gl_shader_program **shProg = ctx->_Shader->CurrentProgram;
1231    unsigned j;
1232
1233    for (j = 0; j < 3; j++) {
1234       unsigned i;
1235
1236       if (shProg[j] == NULL)
1237          continue;
1238
1239       for (i = 0; i < shProg[j]->NumShaders; i++) {
1240          struct gl_shader *sh;
1241
1242          switch (shProg[j]->Shaders[i]->Type) {
1243          case GL_VERTEX_SHADER:
1244             sh = (i != 0) ? NULL : shProg[j]->Shaders[i];
1245             break;
1246          case GL_GEOMETRY_SHADER_ARB:
1247             sh = (i != 1) ? NULL : shProg[j]->Shaders[i];
1248             break;
1249          case GL_FRAGMENT_SHADER:
1250             sh = (i != 2) ? NULL : shProg[j]->Shaders[i];
1251             break;
1252          default:
1253             assert(0);
1254             sh = NULL;
1255             break;
1256          }
1257
1258          if (sh != NULL) {
1259             printf("GLSL shader %u of %u:\n", i, shProg[j]->NumShaders);
1260             printf("%s\n", sh->Source);
1261          }
1262       }
1263    }
1264 }
1265
1266
1267 /**
1268  * Vert/Geom/Frag programs have per-context variants.  Free all the
1269  * variants attached to the given program which match the given context.
1270  */
1271 static void
1272 destroy_program_variants(struct st_context *st, struct gl_program *program)
1273 {
1274    if (!program || program == &_mesa_DummyProgram)
1275       return;
1276
1277    switch (program->Target) {
1278    case GL_VERTEX_PROGRAM_ARB:
1279       {
1280          struct st_vertex_program *stvp = (struct st_vertex_program *) program;
1281          struct st_vp_variant *vpv, **prevPtr = &stvp->variants;
1282
1283          for (vpv = stvp->variants; vpv; ) {
1284             struct st_vp_variant *next = vpv->next;
1285             if (vpv->key.st == st) {
1286                /* unlink from list */
1287                *prevPtr = next;
1288                /* destroy this variant */
1289                delete_vp_variant(st, vpv);
1290             }
1291             else {
1292                prevPtr = &vpv->next;
1293             }
1294             vpv = next;
1295          }
1296       }
1297       break;
1298    case GL_FRAGMENT_PROGRAM_ARB:
1299       {
1300          struct st_fragment_program *stfp =
1301             (struct st_fragment_program *) program;
1302          struct st_fp_variant *fpv, **prevPtr = &stfp->variants;
1303
1304          for (fpv = stfp->variants; fpv; ) {
1305             struct st_fp_variant *next = fpv->next;
1306             if (fpv->key.st == st) {
1307                /* unlink from list */
1308                *prevPtr = next;
1309                /* destroy this variant */
1310                delete_fp_variant(st, fpv);
1311             }
1312             else {
1313                prevPtr = &fpv->next;
1314             }
1315             fpv = next;
1316          }
1317       }
1318       break;
1319    case MESA_GEOMETRY_PROGRAM:
1320       {
1321          struct st_geometry_program *stgp =
1322             (struct st_geometry_program *) program;
1323          struct st_gp_variant *gpv, **prevPtr = &stgp->variants;
1324
1325          for (gpv = stgp->variants; gpv; ) {
1326             struct st_gp_variant *next = gpv->next;
1327             if (gpv->key.st == st) {
1328                /* unlink from list */
1329                *prevPtr = next;
1330                /* destroy this variant */
1331                delete_gp_variant(st, gpv);
1332             }
1333             else {
1334                prevPtr = &gpv->next;
1335             }
1336             gpv = next;
1337          }
1338       }
1339       break;
1340    default:
1341       _mesa_problem(NULL, "Unexpected program target 0x%x in "
1342                     "destroy_program_variants_cb()", program->Target);
1343    }
1344 }
1345
1346
1347 /**
1348  * Callback for _mesa_HashWalk.  Free all the shader's program variants
1349  * which match the given context.
1350  */
1351 static void
1352 destroy_shader_program_variants_cb(GLuint key, void *data, void *userData)
1353 {
1354    struct st_context *st = (struct st_context *) userData;
1355    struct gl_shader *shader = (struct gl_shader *) data;
1356
1357    switch (shader->Type) {
1358    case GL_SHADER_PROGRAM_MESA:
1359       {
1360          struct gl_shader_program *shProg = (struct gl_shader_program *) data;
1361          GLuint i;
1362
1363          for (i = 0; i < shProg->NumShaders; i++) {
1364             destroy_program_variants(st, shProg->Shaders[i]->Program);
1365          }
1366
1367          for (i = 0; i < Elements(shProg->_LinkedShaders); i++) {
1368             if (shProg->_LinkedShaders[i])
1369                destroy_program_variants(st, shProg->_LinkedShaders[i]->Program);
1370          }
1371       }
1372       break;
1373    case GL_VERTEX_SHADER:
1374    case GL_FRAGMENT_SHADER:
1375    case GL_GEOMETRY_SHADER:
1376       {
1377          destroy_program_variants(st, shader->Program);
1378       }
1379       break;
1380    default:
1381       assert(0);
1382    }
1383 }
1384
1385
1386 /**
1387  * Callback for _mesa_HashWalk.  Free all the program variants which match
1388  * the given context.
1389  */
1390 static void
1391 destroy_program_variants_cb(GLuint key, void *data, void *userData)
1392 {
1393    struct st_context *st = (struct st_context *) userData;
1394    struct gl_program *program = (struct gl_program *) data;
1395    destroy_program_variants(st, program);
1396 }
1397
1398
1399 /**
1400  * Walk over all shaders and programs to delete any variants which
1401  * belong to the given context.
1402  * This is called during context tear-down.
1403  */
1404 void
1405 st_destroy_program_variants(struct st_context *st)
1406 {
1407    /* ARB vert/frag program */
1408    _mesa_HashWalk(st->ctx->Shared->Programs,
1409                   destroy_program_variants_cb, st);
1410
1411    /* GLSL vert/frag/geom shaders */
1412    _mesa_HashWalk(st->ctx->Shared->ShaderObjects,
1413                   destroy_shader_program_variants_cb, st);
1414 }
1415
1416
1417 /**
1418  * For debugging, print/dump the current vertex program.
1419  */
1420 void
1421 st_print_current_vertex_program(void)
1422 {
1423    GET_CURRENT_CONTEXT(ctx);
1424
1425    if (ctx->VertexProgram._Current) {
1426       struct st_vertex_program *stvp =
1427          (struct st_vertex_program *) ctx->VertexProgram._Current;
1428       struct st_vp_variant *stv;
1429
1430       debug_printf("Vertex program %u\n", stvp->Base.Base.Id);
1431
1432       for (stv = stvp->variants; stv; stv = stv->next) {
1433          debug_printf("variant %p\n", stv);
1434          tgsi_dump(stv->tgsi.tokens, 0);
1435       }
1436    }
1437 }