src/mesa/program/program.c

   1 /*
   2  * Mesa 3-D graphics library
   3  * Version:  6.5.3
   4  *
   5  * Copyright (C) 1999-2007  Brian Paul   All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  11  * and/or sell copies of the Software, and to permit persons to whom the
  12  * Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included
  15  * in all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  18  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  21  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  22  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 /**
  26  * \file program.c
  27  * Vertex and fragment program support functions.
  28  * \author Brian Paul
  29  */
  30
  31
  32 #include "main/glheader.h"
  33 #include "main/context.h"
  34 #include "main/hash.h"
  35 #include "main/mfeatures.h"
  36 #include "program.h"
  37 #include "prog_cache.h"
  38 #include "prog_parameter.h"
  39 #include "prog_instruction.h"
  40
  41
  42 /**
  43  * A pointer to this dummy program is put into the hash table when
  44  * glGenPrograms is called.
  45  */
  46 struct gl_program _mesa_DummyProgram;
  47
  48
  49 /**
  50  * Init context's vertex/fragment program state
  51  */
  52 void
  53 _mesa_init_program(struct gl_context *ctx)
  54 {
  55    GLuint i;
  56
  57    /*
  58     * If this assertion fails, we need to increase the field
  59     * size for register indexes (see INST_INDEX_BITS).
  60     */
  61    ASSERT(ctx->Const.VertexProgram.MaxUniformComponents / 4
  62           <= (1 << INST_INDEX_BITS));
  63    ASSERT(ctx->Const.FragmentProgram.MaxUniformComponents / 4
  64           <= (1 << INST_INDEX_BITS));
  65
  66    ASSERT(ctx->Const.VertexProgram.MaxTemps <= (1 << INST_INDEX_BITS));
  67    ASSERT(ctx->Const.VertexProgram.MaxLocalParams <= (1 << INST_INDEX_BITS));
  68    ASSERT(ctx->Const.FragmentProgram.MaxTemps <= (1 << INST_INDEX_BITS));
  69    ASSERT(ctx->Const.FragmentProgram.MaxLocalParams <= (1 << INST_INDEX_BITS));
  70
  71    ASSERT(ctx->Const.VertexProgram.MaxUniformComponents <= 4 * MAX_UNIFORMS);
  72    ASSERT(ctx->Const.FragmentProgram.MaxUniformComponents <= 4 * MAX_UNIFORMS);
  73
  74    ASSERT(ctx->Const.VertexProgram.MaxAddressOffset <= (1 << INST_INDEX_BITS));
  75    ASSERT(ctx->Const.FragmentProgram.MaxAddressOffset <= (1 << INST_INDEX_BITS));
  76
  77    /* If this fails, increase prog_instruction::TexSrcUnit size */
  78    ASSERT(MAX_TEXTURE_UNITS <= (1 << 5));
  79
  80    /* If this fails, increase prog_instruction::TexSrcTarget size */
  81    ASSERT(NUM_TEXTURE_TARGETS <= (1 << 3));
  82
  83    ctx->Program.ErrorPos = -1;
  84    ctx->Program.ErrorString = _mesa_strdup("");
  85
  86 #if FEATURE_NV_vertex_program || FEATURE_ARB_vertex_program
  87    ctx->VertexProgram.Enabled = GL_FALSE;
  88 #if FEATURE_es2_glsl
  89    ctx->VertexProgram.PointSizeEnabled =
  90       (ctx->API == API_OPENGLES2) ? GL_TRUE : GL_FALSE;
  91 #else
  92    ctx->VertexProgram.PointSizeEnabled = GL_FALSE;
  93 #endif
  94    ctx->VertexProgram.TwoSideEnabled = GL_FALSE;
  95    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
  96                             ctx->Shared->DefaultVertexProgram);
  97    assert(ctx->VertexProgram.Current);
  98    for (i = 0; i < MAX_NV_VERTEX_PROGRAM_PARAMS / 4; i++) {
  99       ctx->VertexProgram.TrackMatrix[i] = GL_NONE;
 100       ctx->VertexProgram.TrackMatrixTransform[i] = GL_IDENTITY_NV;
 101    }
 102    ctx->VertexProgram.Cache = _mesa_new_program_cache();
 103 #endif
 104
 105 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
 106    ctx->FragmentProgram.Enabled = GL_FALSE;
 107    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
 108                             ctx->Shared->DefaultFragmentProgram);
 109    assert(ctx->FragmentProgram.Current);
 110    ctx->FragmentProgram.Cache = _mesa_new_program_cache();
 111 #endif
 112
 113 #if FEATURE_ARB_geometry_shader4
 114    ctx->GeometryProgram.Enabled = GL_FALSE;
 115    /* right now by default we don't have a geometry program */
 116    _mesa_reference_geomprog(ctx, &ctx->GeometryProgram.Current,
 117                             NULL);
 118    ctx->GeometryProgram.Cache = _mesa_new_program_cache();
 119 #endif
 120
 121    /* XXX probably move this stuff */
 122 #if FEATURE_ATI_fragment_shader
 123    ctx->ATIFragmentShader.Enabled = GL_FALSE;
 124    ctx->ATIFragmentShader.Current = ctx->Shared->DefaultFragmentShader;
 125    assert(ctx->ATIFragmentShader.Current);
 126    ctx->ATIFragmentShader.Current->RefCount++;
 127 #endif
 128 }
 129
 130
 131 /**
 132  * Free a context's vertex/fragment program state
 133  */
 134 void
 135 _mesa_free_program_data(struct gl_context *ctx)
 136 {
 137 #if FEATURE_NV_vertex_program || FEATURE_ARB_vertex_program
 138    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current, NULL);
 139    _mesa_delete_program_cache(ctx, ctx->VertexProgram.Cache);
 140 #endif
 141 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
 142    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current, NULL);
 143    _mesa_delete_program_cache(ctx, ctx->FragmentProgram.Cache);
 144 #endif
 145 #if FEATURE_ARB_geometry_shader4
 146    _mesa_reference_geomprog(ctx, &ctx->GeometryProgram.Current, NULL);
 147    _mesa_delete_program_cache(ctx, ctx->GeometryProgram.Cache);
 148 #endif
 149    /* XXX probably move this stuff */
 150 #if FEATURE_ATI_fragment_shader
 151    if (ctx->ATIFragmentShader.Current) {
 152       ctx->ATIFragmentShader.Current->RefCount--;
 153       if (ctx->ATIFragmentShader.Current->RefCount <= 0) {
 154          free(ctx->ATIFragmentShader.Current);
 155       }
 156    }
 157 #endif
 158    free((void *) ctx->Program.ErrorString);
 159 }
 160
 161
 162 /**
 163  * Update the default program objects in the given context to reference those
 164  * specified in the shared state and release those referencing the old
 165  * shared state.
 166  */
 167 void
 168 _mesa_update_default_objects_program(struct gl_context *ctx)
 169 {
 170 #if FEATURE_NV_vertex_program || FEATURE_ARB_vertex_program
 171    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
 172                             (struct gl_vertex_program *)
 173                             ctx->Shared->DefaultVertexProgram);
 174    assert(ctx->VertexProgram.Current);
 175 #endif
 176
 177 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
 178    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
 179                             (struct gl_fragment_program *)
 180                             ctx->Shared->DefaultFragmentProgram);
 181    assert(ctx->FragmentProgram.Current);
 182 #endif
 183
 184 #if FEATURE_ARB_geometry_shader4
 185    _mesa_reference_geomprog(ctx, &ctx->GeometryProgram.Current,
 186                             (struct gl_geometry_program *)
 187                             ctx->Shared->DefaultGeometryProgram);
 188 #endif
 189
 190    /* XXX probably move this stuff */
 191 #if FEATURE_ATI_fragment_shader
 192    if (ctx->ATIFragmentShader.Current) {
 193       ctx->ATIFragmentShader.Current->RefCount--;
 194       if (ctx->ATIFragmentShader.Current->RefCount <= 0) {
 195          free(ctx->ATIFragmentShader.Current);
 196       }
 197    }
 198    ctx->ATIFragmentShader.Current = (struct ati_fragment_shader *) ctx->Shared->DefaultFragmentShader;
 199    assert(ctx->ATIFragmentShader.Current);
 200    ctx->ATIFragmentShader.Current->RefCount++;
 201 #endif
 202 }
 203
 204
 205 /**
 206  * Set the vertex/fragment program error state (position and error string).
 207  * This is generally called from within the parsers.
 208  */
 209 void
 210 _mesa_set_program_error(struct gl_context *ctx, GLint pos, const char *string)
 211 {
 212    ctx->Program.ErrorPos = pos;
 213    free((void *) ctx->Program.ErrorString);
 214    if (!string)
 215       string = "";
 216    ctx->Program.ErrorString = _mesa_strdup(string);
 217 }
 218
 219
 220 /**
 221  * Find the line number and column for 'pos' within 'string'.
 222  * Return a copy of the line which contains 'pos'.  Free the line with
 223  * free().
 224  * \param string  the program string
 225  * \param pos     the position within the string
 226  * \param line    returns the line number corresponding to 'pos'.
 227  * \param col     returns the column number corresponding to 'pos'.
 228  * \return copy of the line containing 'pos'.
 229  */
 230 const GLubyte *
 231 _mesa_find_line_column(const GLubyte *string, const GLubyte *pos,
 232                        GLint *line, GLint *col)
 233 {
 234    const GLubyte *lineStart = string;
 235    const GLubyte *p = string;
 236    GLubyte *s;
 237    int len;
 238
 239    *line = 1;
 240
 241    while (p != pos) {
 242       if (*p == (GLubyte) '\n') {
 243          (*line)++;
 244          lineStart = p + 1;
 245       }
 246       p++;
 247    }
 248
 249    *col = (pos - lineStart) + 1;
 250
 251    /* return copy of this line */
 252    while (*p != 0 && *p != '\n')
 253       p++;
 254    len = p - lineStart;
 255    s = (GLubyte *) malloc(len + 1);
 256    memcpy(s, lineStart, len);
 257    s[len] = 0;
 258
 259    return s;
 260 }
 261
 262
 263 /**
 264  * Initialize a new vertex/fragment program object.
 265  */
 266 static struct gl_program *
 267 _mesa_init_program_struct( struct gl_context *ctx, struct gl_program *prog,
 268                            GLenum target, GLuint id)
 269 {
 270    (void) ctx;
 271    if (prog) {
 272       GLuint i;
 273       memset(prog, 0, sizeof(*prog));
 274       prog->Id = id;
 275       prog->Target = target;
 276       prog->Resident = GL_TRUE;
 277       prog->RefCount = 1;
 278       prog->Format = GL_PROGRAM_FORMAT_ASCII_ARB;
 279
 280       /* default mapping from samplers to texture units */
 281       for (i = 0; i < MAX_SAMPLERS; i++)
 282          prog->SamplerUnits[i] = i;
 283    }
 284
 285    return prog;
 286 }
 287
 288
 289 /**
 290  * Initialize a new fragment program object.
 291  */
 292 struct gl_program *
 293 _mesa_init_fragment_program( struct gl_context *ctx, struct gl_fragment_program *prog,
 294                              GLenum target, GLuint id)
 295 {
 296    if (prog)
 297       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 298    else
 299       return NULL;
 300 }
 301
 302
 303 /**
 304  * Initialize a new vertex program object.
 305  */
 306 struct gl_program *
 307 _mesa_init_vertex_program( struct gl_context *ctx, struct gl_vertex_program *prog,
 308                            GLenum target, GLuint id)
 309 {
 310    if (prog)
 311       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 312    else
 313       return NULL;
 314 }
 315
 316
 317 /**
 318  * Initialize a new geometry program object.
 319  */
 320 struct gl_program *
 321 _mesa_init_geometry_program( struct gl_context *ctx, struct gl_geometry_program *prog,
 322                              GLenum target, GLuint id)
 323 {
 324    if (prog)
 325       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 326    else
 327       return NULL;
 328 }
 329
 330
 331 /**
 332  * Allocate and initialize a new fragment/vertex program object but
 333  * don't put it into the program hash table.  Called via
 334  * ctx->Driver.NewProgram.  May be overridden (ie. replaced) by a
 335  * device driver function to implement OO deriviation with additional
 336  * types not understood by this function.
 337  *
 338  * \param ctx  context
 339  * \param id   program id/number
 340  * \param target  program target/type
 341  * \return  pointer to new program object
 342  */
 343 struct gl_program *
 344 _mesa_new_program(struct gl_context *ctx, GLenum target, GLuint id)
 345 {
 346    struct gl_program *prog;
 347    switch (target) {
 348    case GL_VERTEX_PROGRAM_ARB: /* == GL_VERTEX_PROGRAM_NV */
 349    case GL_VERTEX_STATE_PROGRAM_NV:
 350       prog = _mesa_init_vertex_program(ctx, CALLOC_STRUCT(gl_vertex_program),
 351                                        target, id );
 352       break;
 353    case GL_FRAGMENT_PROGRAM_NV:
 354    case GL_FRAGMENT_PROGRAM_ARB:
 355       prog =_mesa_init_fragment_program(ctx,
 356                                          CALLOC_STRUCT(gl_fragment_program),
 357                                          target, id );
 358       break;
 359    case MESA_GEOMETRY_PROGRAM:
 360       prog = _mesa_init_geometry_program(ctx,
 361                                          CALLOC_STRUCT(gl_geometry_program),
 362                                          target, id);
 363       break;
 364    default:
 365       _mesa_problem(ctx, "bad target in _mesa_new_program");
 366       prog = NULL;
 367    }
 368    return prog;
 369 }
 370
 371
 372 /**
 373  * Delete a program and remove it from the hash table, ignoring the
 374  * reference count.
 375  * Called via ctx->Driver.DeleteProgram.  May be wrapped (OO deriviation)
 376  * by a device driver function.
 377  */
 378 void
 379 _mesa_delete_program(struct gl_context *ctx, struct gl_program *prog)
 380 {
 381    (void) ctx;
 382    ASSERT(prog);
 383    ASSERT(prog->RefCount==0);
 384
 385    if (prog == &_mesa_DummyProgram)
 386       return;
 387
 388    if (prog->String)
 389       free(prog->String);
 390
 391    _mesa_free_instructions(prog->Instructions, prog->NumInstructions);
 392
 393    if (prog->Parameters) {
 394       _mesa_free_parameter_list(prog->Parameters);
 395    }
 396    if (prog->Varying) {
 397       _mesa_free_parameter_list(prog->Varying);
 398    }
 399    if (prog->Attributes) {
 400       _mesa_free_parameter_list(prog->Attributes);
 401    }
 402
 403    free(prog);
 404 }
 405
 406
 407 /**
 408  * Return the gl_program object for a given ID.
 409  * Basically just a wrapper for _mesa_HashLookup() to avoid a lot of
 410  * casts elsewhere.
 411  */
 412 struct gl_program *
 413 _mesa_lookup_program(struct gl_context *ctx, GLuint id)
 414 {
 415    if (id)
 416       return (struct gl_program *) _mesa_HashLookup(ctx->Shared->Programs, id);
 417    else
 418       return NULL;
 419 }
 420
 421
 422 /**
 423  * Reference counting for vertex/fragment programs
 424  * This is normally only called from the _mesa_reference_program() macro
 425  * when there's a real pointer change.
 426  */
 427 void
 428 _mesa_reference_program_(struct gl_context *ctx,
 429                          struct gl_program **ptr,
 430                          struct gl_program *prog)
 431 {
 432 #ifndef NDEBUG
 433    assert(ptr);
 434    if (*ptr && prog) {
 435       /* sanity check */
 436       if ((*ptr)->Target == GL_VERTEX_PROGRAM_ARB)
 437          ASSERT(prog->Target == GL_VERTEX_PROGRAM_ARB);
 438       else if ((*ptr)->Target == GL_FRAGMENT_PROGRAM_ARB)
 439          ASSERT(prog->Target == GL_FRAGMENT_PROGRAM_ARB ||
 440                 prog->Target == GL_FRAGMENT_PROGRAM_NV);
 441       else if ((*ptr)->Target == MESA_GEOMETRY_PROGRAM)
 442          ASSERT(prog->Target == MESA_GEOMETRY_PROGRAM);
 443    }
 444 #endif
 445
 446    if (*ptr) {
 447       GLboolean deleteFlag;
 448
 449       /*_glthread_LOCK_MUTEX((*ptr)->Mutex);*/
 450 #if 0
 451       printf("Program %p ID=%u Target=%s  Refcount-- to %d\n",
 452              *ptr, (*ptr)->Id,
 453              ((*ptr)->Target == GL_VERTEX_PROGRAM_ARB ? "VP" :
 454               ((*ptr)->Target == MESA_GEOMETRY_PROGRAM ? "GP" : "FP")),
 455              (*ptr)->RefCount - 1);
 456 #endif
 457       ASSERT((*ptr)->RefCount > 0);
 458       (*ptr)->RefCount--;
 459
 460       deleteFlag = ((*ptr)->RefCount == 0);
 461       /*_glthread_UNLOCK_MUTEX((*ptr)->Mutex);*/
 462
 463       if (deleteFlag) {
 464          ASSERT(ctx);
 465          ctx->Driver.DeleteProgram(ctx, *ptr);
 466       }
 467
 468       *ptr = NULL;
 469    }
 470
 471    assert(!*ptr);
 472    if (prog) {
 473       /*_glthread_LOCK_MUTEX(prog->Mutex);*/
 474       prog->RefCount++;
 475 #if 0
 476       printf("Program %p ID=%u Target=%s  Refcount++ to %d\n",
 477              prog, prog->Id,
 478              (prog->Target == GL_VERTEX_PROGRAM_ARB ? "VP" :
 479               (prog->Target == MESA_GEOMETRY_PROGRAM ? "GP" : "FP")),
 480              prog->RefCount);
 481 #endif
 482       /*_glthread_UNLOCK_MUTEX(prog->Mutex);*/
 483    }
 484
 485    *ptr = prog;
 486 }
 487
 488
 489 /**
 490  * Return a copy of a program.
 491  * XXX Problem here if the program object is actually OO-derivation
 492  * made by a device driver.
 493  */
 494 struct gl_program *
 495 _mesa_clone_program(struct gl_context *ctx, const struct gl_program *prog)
 496 {
 497    struct gl_program *clone;
 498
 499    clone = ctx->Driver.NewProgram(ctx, prog->Target, prog->Id);
 500    if (!clone)
 501       return NULL;
 502
 503    assert(clone->Target == prog->Target);
 504    assert(clone->RefCount == 1);
 505
 506    clone->String = (GLubyte *) _mesa_strdup((char *) prog->String);
 507    clone->Format = prog->Format;
 508    clone->Instructions = _mesa_alloc_instructions(prog->NumInstructions);
 509    if (!clone->Instructions) {
 510       _mesa_reference_program(ctx, &clone, NULL);
 511       return NULL;
 512    }
 513    _mesa_copy_instructions(clone->Instructions, prog->Instructions,
 514                            prog->NumInstructions);
 515    clone->InputsRead = prog->InputsRead;
 516    clone->OutputsWritten = prog->OutputsWritten;
 517    clone->SamplersUsed = prog->SamplersUsed;
 518    clone->ShadowSamplers = prog->ShadowSamplers;
 519    memcpy(clone->TexturesUsed, prog->TexturesUsed, sizeof(prog->TexturesUsed));
 520
 521    if (prog->Parameters)
 522       clone->Parameters = _mesa_clone_parameter_list(prog->Parameters);
 523    memcpy(clone->LocalParams, prog->LocalParams, sizeof(clone->LocalParams));
 524    if (prog->Varying)
 525       clone->Varying = _mesa_clone_parameter_list(prog->Varying);
 526    if (prog->Attributes)
 527       clone->Attributes = _mesa_clone_parameter_list(prog->Attributes);
 528    memcpy(clone->LocalParams, prog->LocalParams, sizeof(clone->LocalParams));
 529    clone->IndirectRegisterFiles = prog->IndirectRegisterFiles;
 530    clone->NumInstructions = prog->NumInstructions;
 531    clone->NumTemporaries = prog->NumTemporaries;
 532    clone->NumParameters = prog->NumParameters;
 533    clone->NumAttributes = prog->NumAttributes;
 534    clone->NumAddressRegs = prog->NumAddressRegs;
 535    clone->NumNativeInstructions = prog->NumNativeInstructions;
 536    clone->NumNativeTemporaries = prog->NumNativeTemporaries;
 537    clone->NumNativeParameters = prog->NumNativeParameters;
 538    clone->NumNativeAttributes = prog->NumNativeAttributes;
 539    clone->NumNativeAddressRegs = prog->NumNativeAddressRegs;
 540    clone->NumAluInstructions = prog->NumAluInstructions;
 541    clone->NumTexInstructions = prog->NumTexInstructions;
 542    clone->NumTexIndirections = prog->NumTexIndirections;
 543    clone->NumNativeAluInstructions = prog->NumNativeAluInstructions;
 544    clone->NumNativeTexInstructions = prog->NumNativeTexInstructions;
 545    clone->NumNativeTexIndirections = prog->NumNativeTexIndirections;
 546
 547    switch (prog->Target) {
 548    case GL_VERTEX_PROGRAM_ARB:
 549       {
 550          const struct gl_vertex_program *vp
 551             = (const struct gl_vertex_program *) prog;
 552          struct gl_vertex_program *vpc = (struct gl_vertex_program *) clone;
 553          vpc->IsPositionInvariant = vp->IsPositionInvariant;
 554          vpc->IsNVProgram = vp->IsNVProgram;
 555       }
 556       break;
 557    case GL_FRAGMENT_PROGRAM_ARB:
 558       {
 559          const struct gl_fragment_program *fp
 560             = (const struct gl_fragment_program *) prog;
 561          struct gl_fragment_program *fpc = (struct gl_fragment_program *) clone;
 562          fpc->UsesKill = fp->UsesKill;
 563          fpc->OriginUpperLeft = fp->OriginUpperLeft;
 564          fpc->PixelCenterInteger = fp->PixelCenterInteger;
 565       }
 566       break;
 567    case MESA_GEOMETRY_PROGRAM:
 568       {
 569          const struct gl_geometry_program *gp
 570             = (const struct gl_geometry_program *) prog;
 571          struct gl_geometry_program *gpc = (struct gl_geometry_program *) clone;
 572          gpc->VerticesOut = gp->VerticesOut;
 573          gpc->InputType = gp->InputType;
 574          gpc->OutputType = gp->OutputType;
 575       }
 576       break;
 577    default:
 578       _mesa_problem(NULL, "Unexpected target in _mesa_clone_program");
 579    }
 580
 581    return clone;
 582 }
 583
 584
 585 /**
 586  * Insert 'count' NOP instructions at 'start' in the given program.
 587  * Adjust branch targets accordingly.
 588  */
 589 GLboolean
 590 _mesa_insert_instructions(struct gl_program *prog, GLuint start, GLuint count)
 591 {
 592    const GLuint origLen = prog->NumInstructions;
 593    const GLuint newLen = origLen + count;
 594    struct prog_instruction *newInst;
 595    GLuint i;
 596
 597    /* adjust branches */
 598    for (i = 0; i < prog->NumInstructions; i++) {
 599       struct prog_instruction *inst = prog->Instructions + i;
 600       if (inst->BranchTarget > 0) {
 601          if ((GLuint)inst->BranchTarget >= start) {
 602             inst->BranchTarget += count;
 603          }
 604       }
 605    }
 606
 607    /* Alloc storage for new instructions */
 608    newInst = _mesa_alloc_instructions(newLen);
 609    if (!newInst) {
 610       return GL_FALSE;
 611    }
 612
 613    /* Copy 'start' instructions into new instruction buffer */
 614    _mesa_copy_instructions(newInst, prog->Instructions, start);
 615
 616    /* init the new instructions */
 617    _mesa_init_instructions(newInst + start, count);
 618
 619    /* Copy the remaining/tail instructions to new inst buffer */
 620    _mesa_copy_instructions(newInst + start + count,
 621                            prog->Instructions + start,
 622                            origLen - start);
 623
 624    /* free old instructions */
 625    _mesa_free_instructions(prog->Instructions, origLen);
 626
 627    /* install new instructions */
 628    prog->Instructions = newInst;
 629    prog->NumInstructions = newLen;
 630
 631    return GL_TRUE;
 632 }
 633
 634 /**
 635  * Delete 'count' instructions at 'start' in the given program.
 636  * Adjust branch targets accordingly.
 637  */
 638 GLboolean
 639 _mesa_delete_instructions(struct gl_program *prog, GLuint start, GLuint count)
 640 {
 641    const GLuint origLen = prog->NumInstructions;
 642    const GLuint newLen = origLen - count;
 643    struct prog_instruction *newInst;
 644    GLuint i;
 645
 646    /* adjust branches */
 647    for (i = 0; i < prog->NumInstructions; i++) {
 648       struct prog_instruction *inst = prog->Instructions + i;
 649       if (inst->BranchTarget > 0) {
 650          if (inst->BranchTarget > (GLint) start) {
 651             inst->BranchTarget -= count;
 652          }
 653       }
 654    }
 655
 656    /* Alloc storage for new instructions */
 657    newInst = _mesa_alloc_instructions(newLen);
 658    if (!newInst) {
 659       return GL_FALSE;
 660    }
 661
 662    /* Copy 'start' instructions into new instruction buffer */
 663    _mesa_copy_instructions(newInst, prog->Instructions, start);
 664
 665    /* Copy the remaining/tail instructions to new inst buffer */
 666    _mesa_copy_instructions(newInst + start,
 667                            prog->Instructions + start + count,
 668                            newLen - start);
 669
 670    /* free old instructions */
 671    _mesa_free_instructions(prog->Instructions, origLen);
 672
 673    /* install new instructions */
 674    prog->Instructions = newInst;
 675    prog->NumInstructions = newLen;
 676
 677    return GL_TRUE;
 678 }
 679
 680
 681 /**
 682  * Search instructions for registers that match (oldFile, oldIndex),
 683  * replacing them with (newFile, newIndex).
 684  */
 685 static void
 686 replace_registers(struct prog_instruction *inst, GLuint numInst,
 687                   GLuint oldFile, GLuint oldIndex,
 688                   GLuint newFile, GLuint newIndex)
 689 {
 690    GLuint i, j;
 691    for (i = 0; i < numInst; i++) {
 692       /* src regs */
 693       for (j = 0; j < _mesa_num_inst_src_regs(inst[i].Opcode); j++) {
 694          if (inst[i].SrcReg[j].File == oldFile &&
 695              inst[i].SrcReg[j].Index == oldIndex) {
 696             inst[i].SrcReg[j].File = newFile;
 697             inst[i].SrcReg[j].Index = newIndex;
 698          }
 699       }
 700       /* dst reg */
 701       if (inst[i].DstReg.File == oldFile && inst[i].DstReg.Index == oldIndex) {
 702          inst[i].DstReg.File = newFile;
 703          inst[i].DstReg.Index = newIndex;
 704       }
 705    }
 706 }
 707
 708
 709 /**
 710  * Search instructions for references to program parameters.  When found,
 711  * increment the parameter index by 'offset'.
 712  * Used when combining programs.
 713  */
 714 static void
 715 adjust_param_indexes(struct prog_instruction *inst, GLuint numInst,
 716                      GLuint offset)
 717 {
 718    GLuint i, j;
 719    for (i = 0; i < numInst; i++) {
 720       for (j = 0; j < _mesa_num_inst_src_regs(inst[i].Opcode); j++) {
 721          GLuint f = inst[i].SrcReg[j].File;
 722          if (f == PROGRAM_CONSTANT ||
 723              f == PROGRAM_UNIFORM ||
 724              f == PROGRAM_STATE_VAR) {
 725             inst[i].SrcReg[j].Index += offset;
 726          }
 727       }
 728    }
 729 }
 730
 731
 732 /**
 733  * Combine two programs into one.  Fix instructions so the outputs of
 734  * the first program go to the inputs of the second program.
 735  */
 736 struct gl_program *
 737 _mesa_combine_programs(struct gl_context *ctx,
 738                        const struct gl_program *progA,
 739                        const struct gl_program *progB)
 740 {
 741    struct prog_instruction *newInst;
 742    struct gl_program *newProg;
 743    const GLuint lenA = progA->NumInstructions - 1; /* omit END instr */
 744    const GLuint lenB = progB->NumInstructions;
 745    const GLuint numParamsA = _mesa_num_parameters(progA->Parameters);
 746    const GLuint newLength = lenA + lenB;
 747    GLboolean usedTemps[MAX_PROGRAM_TEMPS];
 748    GLuint firstTemp = 0;
 749    GLbitfield inputsB;
 750    GLuint i;
 751
 752    ASSERT(progA->Target == progB->Target);
 753
 754    newInst = _mesa_alloc_instructions(newLength);
 755    if (!newInst)
 756       return GL_FALSE;
 757
 758    _mesa_copy_instructions(newInst, progA->Instructions, lenA);
 759    _mesa_copy_instructions(newInst + lenA, progB->Instructions, lenB);
 760
 761    /* adjust branch / instruction addresses for B's instructions */
 762    for (i = 0; i < lenB; i++) {
 763       newInst[lenA + i].BranchTarget += lenA;
 764    }
 765
 766    newProg = ctx->Driver.NewProgram(ctx, progA->Target, 0);
 767    newProg->Instructions = newInst;
 768    newProg->NumInstructions = newLength;
 769
 770    /* find used temp regs (we may need new temps below) */
 771    _mesa_find_used_registers(newProg, PROGRAM_TEMPORARY,
 772                              usedTemps, MAX_PROGRAM_TEMPS);
 773
 774    if (newProg->Target == GL_FRAGMENT_PROGRAM_ARB) {
 775       struct gl_fragment_program *fprogA, *fprogB, *newFprog;
 776       GLbitfield progB_inputsRead = progB->InputsRead;
 777       GLint progB_colorFile, progB_colorIndex;
 778
 779       fprogA = (struct gl_fragment_program *) progA;
 780       fprogB = (struct gl_fragment_program *) progB;
 781       newFprog = (struct gl_fragment_program *) newProg;
 782
 783       newFprog->UsesKill = fprogA->UsesKill || fprogB->UsesKill;
 784
 785       /* We'll do a search and replace for instances
 786        * of progB_colorFile/progB_colorIndex below...
 787        */
 788       progB_colorFile = PROGRAM_INPUT;
 789       progB_colorIndex = FRAG_ATTRIB_COL0;
 790
 791       /*
 792        * The fragment program may get color from a state var rather than
 793        * a fragment input (vertex output) if it's constant.
 794        * See the texenvprogram.c code.
 795        * So, search the program's parameter list now to see if the program
 796        * gets color from a state var instead of a conventional fragment
 797        * input register.
 798        */
 799       for (i = 0; i < progB->Parameters->NumParameters; i++) {
 800          struct gl_program_parameter *p = &progB->Parameters->Parameters[i];
 801          if (p->Type == PROGRAM_STATE_VAR &&
 802              p->StateIndexes[0] == STATE_INTERNAL &&
 803              p->StateIndexes[1] == STATE_CURRENT_ATTRIB &&
 804              (int) p->StateIndexes[2] == (int) VERT_ATTRIB_COLOR0) {
 805             progB_inputsRead |= FRAG_BIT_COL0;
 806             progB_colorFile = PROGRAM_STATE_VAR;
 807             progB_colorIndex = i;
 808             break;
 809          }
 810       }
 811
 812       /* Connect color outputs of fprogA to color inputs of fprogB, via a
 813        * new temporary register.
 814        */
 815       if ((progA->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_COLOR)) &&
 816           (progB_inputsRead & FRAG_BIT_COL0)) {
 817          GLint tempReg = _mesa_find_free_register(usedTemps, MAX_PROGRAM_TEMPS,
 818                                                   firstTemp);
 819          if (tempReg < 0) {
 820             _mesa_problem(ctx, "No free temp regs found in "
 821                           "_mesa_combine_programs(), using 31");
 822             tempReg = 31;
 823          }
 824          firstTemp = tempReg + 1;
 825
 826          /* replace writes to result.color[0] with tempReg */
 827          replace_registers(newInst, lenA,
 828                            PROGRAM_OUTPUT, FRAG_RESULT_COLOR,
 829                            PROGRAM_TEMPORARY, tempReg);
 830          /* replace reads from the input color with tempReg */
 831          replace_registers(newInst + lenA, lenB,
 832                            progB_colorFile, progB_colorIndex, /* search for */
 833                            PROGRAM_TEMPORARY, tempReg  /* replace with */ );
 834       }
 835
 836       /* compute combined program's InputsRead */
 837       inputsB = progB_inputsRead;
 838       if (progA->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_COLOR)) {
 839          inputsB &= ~(1 << FRAG_ATTRIB_COL0);
 840       }
 841       newProg->InputsRead = progA->InputsRead | inputsB;
 842       newProg->OutputsWritten = progB->OutputsWritten;
 843       newProg->SamplersUsed = progA->SamplersUsed | progB->SamplersUsed;
 844    }
 845    else {
 846       /* vertex program */
 847       assert(0);      /* XXX todo */
 848    }
 849
 850    /*
 851     * Merge parameters (uniforms, constants, etc)
 852     */
 853    newProg->Parameters = _mesa_combine_parameter_lists(progA->Parameters,
 854                                                        progB->Parameters);
 855
 856    adjust_param_indexes(newInst + lenA, lenB, numParamsA);
 857
 858
 859    return newProg;
 860 }
 861
 862
 863 /**
 864  * Populate the 'used' array with flags indicating which registers (TEMPs,
 865  * INPUTs, OUTPUTs, etc, are used by the given program.
 866  * \param file  type of register to scan for
 867  * \param used  returns true/false flags for in use / free
 868  * \param usedSize  size of the 'used' array
 869  */
 870 void
 871 _mesa_find_used_registers(const struct gl_program *prog,
 872                           gl_register_file file,
 873                           GLboolean used[], GLuint usedSize)
 874 {
 875    GLuint i, j;
 876
 877    memset(used, 0, usedSize);
 878
 879    for (i = 0; i < prog->NumInstructions; i++) {
 880       const struct prog_instruction *inst = prog->Instructions + i;
 881       const GLuint n = _mesa_num_inst_src_regs(inst->Opcode);
 882
 883       if (inst->DstReg.File == file) {
 884          ASSERT(inst->DstReg.Index < usedSize);
 885          if(inst->DstReg.Index < usedSize)
 886             used[inst->DstReg.Index] = GL_TRUE;
 887       }
 888
 889       for (j = 0; j < n; j++) {
 890          if (inst->SrcReg[j].File == file) {
 891             ASSERT(inst->SrcReg[j].Index < usedSize);
 892             if(inst->SrcReg[j].Index < usedSize)
 893                used[inst->SrcReg[j].Index] = GL_TRUE;
 894          }
 895       }
 896    }
 897 }
 898
 899
 900 /**
 901  * Scan the given 'used' register flag array for the first entry
 902  * that's >= firstReg.
 903  * \param used  vector of flags indicating registers in use (as returned
 904  *              by _mesa_find_used_registers())
 905  * \param usedSize  size of the 'used' array
 906  * \param firstReg  first register to start searching at
 907  * \return index of unused register, or -1 if none.
 908  */
 909 GLint
 910 _mesa_find_free_register(const GLboolean used[],
 911                          GLuint usedSize, GLuint firstReg)
 912 {
 913    GLuint i;
 914
 915    assert(firstReg < usedSize);
 916
 917    for (i = firstReg; i < usedSize; i++)
 918       if (!used[i])
 919          return i;
 920
 921    return -1;
 922 }
 923
 924
 925
 926 /**
 927  * Check if the given register index is valid (doesn't exceed implementation-
 928  * dependent limits).
 929  * \return GL_TRUE if OK, GL_FALSE if bad index
 930  */
 931 GLboolean
 932 _mesa_valid_register_index(const struct gl_context *ctx,
 933                            gl_shader_type shaderType,
 934                            gl_register_file file, GLint index)
 935 {
 936    const struct gl_program_constants *c;
 937
 938    switch (shaderType) {
 939    case MESA_SHADER_VERTEX:
 940       c = &ctx->Const.VertexProgram;
 941       break;
 942    case MESA_SHADER_FRAGMENT:
 943       c = &ctx->Const.FragmentProgram;
 944       break;
 945    case MESA_SHADER_GEOMETRY:
 946       c = &ctx->Const.GeometryProgram;
 947       break;
 948    default:
 949       _mesa_problem(ctx,
 950                     "unexpected shader type in _mesa_valid_register_index()");
 951       return GL_FALSE;
 952    }
 953
 954    switch (file) {
 955    case PROGRAM_UNDEFINED:
 956       return GL_TRUE;  /* XXX or maybe false? */
 957
 958    case PROGRAM_TEMPORARY:
 959       return index >= 0 && index < c->MaxTemps;
 960
 961    case PROGRAM_ENV_PARAM:
 962       return index >= 0 && index < c->MaxEnvParams;
 963
 964    case PROGRAM_LOCAL_PARAM:
 965       return index >= 0 && index < c->MaxLocalParams;
 966
 967    case PROGRAM_NAMED_PARAM:
 968       return index >= 0 && index < c->MaxParameters;
 969
 970    case PROGRAM_UNIFORM:
 971    case PROGRAM_STATE_VAR:
 972       /* aka constant buffer */
 973       return index >= 0 && index < c->MaxUniformComponents / 4;
 974
 975    case PROGRAM_CONSTANT:
 976       /* constant buffer w/ possible relative negative addressing */
 977       return (index > (int) c->MaxUniformComponents / -4 &&
 978               index < c->MaxUniformComponents / 4);
 979
 980    case PROGRAM_INPUT:
 981       if (index < 0)
 982          return GL_FALSE;
 983
 984       switch (shaderType) {
 985       case MESA_SHADER_VERTEX:
 986          return index < VERT_ATTRIB_GENERIC0 + c->MaxAttribs;
 987       case MESA_SHADER_FRAGMENT:
 988          return index < FRAG_ATTRIB_VAR0 + ctx->Const.MaxVarying;
 989       case MESA_SHADER_GEOMETRY:
 990          return index < GEOM_ATTRIB_VAR0 + ctx->Const.MaxVarying;
 991       default:
 992          return GL_FALSE;
 993       }
 994
 995    case PROGRAM_OUTPUT:
 996       if (index < 0)
 997          return GL_FALSE;
 998
 999       switch (shaderType) {
1000       case MESA_SHADER_VERTEX:
1001          return index < VERT_RESULT_VAR0 + ctx->Const.MaxVarying;
1002       case MESA_SHADER_FRAGMENT:
1003          return index < FRAG_RESULT_DATA0 + ctx->Const.MaxDrawBuffers;
1004       case MESA_SHADER_GEOMETRY:
1005          return index < GEOM_RESULT_VAR0 + ctx->Const.MaxVarying;
1006       default:
1007          return GL_FALSE;
1008       }
1009
1010    case PROGRAM_ADDRESS:
1011       return index >= 0 && index < c->MaxAddressRegs;
1012
1013    default:
1014       _mesa_problem(ctx,
1015                     "unexpected register file in _mesa_valid_register_index()");
1016       return GL_FALSE;
1017    }
1018 }
1019
1020
1021
1022 /**
1023  * "Post-process" a GPU program.  This is intended to be used for debugging.
1024  * Example actions include no-op'ing instructions or changing instruction
1025  * behaviour.
1026  */
1027 void
1028 _mesa_postprocess_program(struct gl_context *ctx, struct gl_program *prog)
1029 {
1030    static const GLfloat white[4] = { 0.5, 0.5, 0.5, 0.5 };
1031    GLuint i;
1032    GLuint whiteSwizzle;
1033    GLint whiteIndex = _mesa_add_unnamed_constant(prog->Parameters,
1034                                                  white, 4, &whiteSwizzle);
1035
1036    (void) whiteIndex;
1037
1038    for (i = 0; i < prog->NumInstructions; i++) {
1039       struct prog_instruction *inst = prog->Instructions + i;
1040       const GLuint n = _mesa_num_inst_src_regs(inst->Opcode);
1041
1042       (void) n;
1043
1044       if (_mesa_is_tex_instruction(inst->Opcode)) {
1045 #if 0
1046          /* replace TEX/TXP/TXB with MOV */
1047          inst->Opcode = OPCODE_MOV;
1048          inst->DstReg.WriteMask = WRITEMASK_XYZW;
1049          inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
1050          inst->SrcReg[0].Negate = NEGATE_NONE;
1051 #endif
1052
1053 #if 0
1054          /* disable shadow texture mode */
1055          inst->TexShadow = 0;
1056 #endif
1057       }
1058
1059       if (inst->Opcode == OPCODE_TXP) {
1060 #if 0
1061          inst->Opcode = OPCODE_MOV;
1062          inst->DstReg.WriteMask = WRITEMASK_XYZW;
1063          inst->SrcReg[0].File = PROGRAM_CONSTANT;
1064          inst->SrcReg[0].Index = whiteIndex;
1065          inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
1066          inst->SrcReg[0].Negate = NEGATE_NONE;
1067 #endif
1068 #if 0
1069          inst->TexShadow = 0;
1070 #endif
1071 #if 0
1072          inst->Opcode = OPCODE_TEX;
1073          inst->TexShadow = 0;
1074 #endif
1075       }
1076
1077    }
1078 }