src/mesa/shader/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 "program.h"
  36 #include "prog_cache.h"
  37 #include "prog_parameter.h"
  38 #include "prog_instruction.h"
  39
  40
  41 /**
  42  * A pointer to this dummy program is put into the hash table when
  43  * glGenPrograms is called.
  44  */
  45 struct gl_program _mesa_DummyProgram;
  46
  47
  48 /**
  49  * Init context's vertex/fragment program state
  50  */
  51 void
  52 _mesa_init_program(GLcontext *ctx)
  53 {
  54    GLuint i;
  55
  56    /*
  57     * If this assertion fails, we need to increase the field
  58     * size for register indexes.
  59     */
  60    ASSERT(ctx->Const.VertexProgram.MaxUniformComponents / 4
  61           <= (1 << INST_INDEX_BITS));
  62    ASSERT(ctx->Const.FragmentProgram.MaxUniformComponents / 4
  63           <= (1 << INST_INDEX_BITS));
  64
  65    /* If this fails, increase prog_instruction::TexSrcUnit size */
  66    ASSERT(MAX_TEXTURE_UNITS < (1 << 5));
  67
  68    /* If this fails, increase prog_instruction::TexSrcTarget size */
  69    ASSERT(NUM_TEXTURE_TARGETS < (1 << 3));
  70
  71    ctx->Program.ErrorPos = -1;
  72    ctx->Program.ErrorString = _mesa_strdup("");
  73
  74 #if FEATURE_NV_vertex_program || FEATURE_ARB_vertex_program
  75    ctx->VertexProgram.Enabled = GL_FALSE;
  76 #if FEATURE_es2_glsl
  77    ctx->VertexProgram.PointSizeEnabled = GL_TRUE;
  78 #else
  79    ctx->VertexProgram.PointSizeEnabled = GL_FALSE;
  80 #endif
  81    ctx->VertexProgram.TwoSideEnabled = GL_FALSE;
  82    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
  83                             ctx->Shared->DefaultVertexProgram);
  84    assert(ctx->VertexProgram.Current);
  85    for (i = 0; i < MAX_NV_VERTEX_PROGRAM_PARAMS / 4; i++) {
  86       ctx->VertexProgram.TrackMatrix[i] = GL_NONE;
  87       ctx->VertexProgram.TrackMatrixTransform[i] = GL_IDENTITY_NV;
  88    }
  89    ctx->VertexProgram.Cache = _mesa_new_program_cache();
  90 #endif
  91
  92 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
  93    ctx->FragmentProgram.Enabled = GL_FALSE;
  94    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
  95                             ctx->Shared->DefaultFragmentProgram);
  96    assert(ctx->FragmentProgram.Current);
  97    ctx->FragmentProgram.Cache = _mesa_new_program_cache();
  98 #endif
  99
 100
 101    /* XXX probably move this stuff */
 102 #if FEATURE_ATI_fragment_shader
 103    ctx->ATIFragmentShader.Enabled = GL_FALSE;
 104    ctx->ATIFragmentShader.Current = ctx->Shared->DefaultFragmentShader;
 105    assert(ctx->ATIFragmentShader.Current);
 106    ctx->ATIFragmentShader.Current->RefCount++;
 107 #endif
 108 }
 109
 110
 111 /**
 112  * Free a context's vertex/fragment program state
 113  */
 114 void
 115 _mesa_free_program_data(GLcontext *ctx)
 116 {
 117 #if FEATURE_NV_vertex_program || FEATURE_ARB_vertex_program
 118    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current, NULL);
 119    _mesa_delete_program_cache(ctx, ctx->VertexProgram.Cache);
 120 #endif
 121 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
 122    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current, NULL);
 123    _mesa_delete_program_cache(ctx, ctx->FragmentProgram.Cache);
 124 #endif
 125    /* XXX probably move this stuff */
 126 #if FEATURE_ATI_fragment_shader
 127    if (ctx->ATIFragmentShader.Current) {
 128       ctx->ATIFragmentShader.Current->RefCount--;
 129       if (ctx->ATIFragmentShader.Current->RefCount <= 0) {
 130          _mesa_free(ctx->ATIFragmentShader.Current);
 131       }
 132    }
 133 #endif
 134    _mesa_free((void *) ctx->Program.ErrorString);
 135 }
 136
 137
 138 /**
 139  * Update the default program objects in the given context to reference those
 140  * specified in the shared state and release those referencing the old
 141  * shared state.
 142  */
 143 void
 144 _mesa_update_default_objects_program(GLcontext *ctx)
 145 {
 146 #if FEATURE_NV_vertex_program || FEATURE_ARB_vertex_program
 147    _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
 148                             (struct gl_vertex_program *)
 149                             ctx->Shared->DefaultVertexProgram);
 150    assert(ctx->VertexProgram.Current);
 151 #endif
 152
 153 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
 154    _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
 155                             (struct gl_fragment_program *)
 156                             ctx->Shared->DefaultFragmentProgram);
 157    assert(ctx->FragmentProgram.Current);
 158 #endif
 159
 160    /* XXX probably move this stuff */
 161 #if FEATURE_ATI_fragment_shader
 162    if (ctx->ATIFragmentShader.Current) {
 163       ctx->ATIFragmentShader.Current->RefCount--;
 164       if (ctx->ATIFragmentShader.Current->RefCount <= 0) {
 165          _mesa_free(ctx->ATIFragmentShader.Current);
 166       }
 167    }
 168    ctx->ATIFragmentShader.Current = (struct ati_fragment_shader *) ctx->Shared->DefaultFragmentShader;
 169    assert(ctx->ATIFragmentShader.Current);
 170    ctx->ATIFragmentShader.Current->RefCount++;
 171 #endif
 172 }
 173
 174
 175 /**
 176  * Set the vertex/fragment program error state (position and error string).
 177  * This is generally called from within the parsers.
 178  */
 179 void
 180 _mesa_set_program_error(GLcontext *ctx, GLint pos, const char *string)
 181 {
 182    ctx->Program.ErrorPos = pos;
 183    _mesa_free((void *) ctx->Program.ErrorString);
 184    if (!string)
 185       string = "";
 186    ctx->Program.ErrorString = _mesa_strdup(string);
 187 }
 188
 189
 190 /**
 191  * Find the line number and column for 'pos' within 'string'.
 192  * Return a copy of the line which contains 'pos'.  Free the line with
 193  * _mesa_free().
 194  * \param string  the program string
 195  * \param pos     the position within the string
 196  * \param line    returns the line number corresponding to 'pos'.
 197  * \param col     returns the column number corresponding to 'pos'.
 198  * \return copy of the line containing 'pos'.
 199  */
 200 const GLubyte *
 201 _mesa_find_line_column(const GLubyte *string, const GLubyte *pos,
 202                        GLint *line, GLint *col)
 203 {
 204    const GLubyte *lineStart = string;
 205    const GLubyte *p = string;
 206    GLubyte *s;
 207    int len;
 208
 209    *line = 1;
 210
 211    while (p != pos) {
 212       if (*p == (GLubyte) '\n') {
 213          (*line)++;
 214          lineStart = p + 1;
 215       }
 216       p++;
 217    }
 218
 219    *col = (pos - lineStart) + 1;
 220
 221    /* return copy of this line */
 222    while (*p != 0 && *p != '\n')
 223       p++;
 224    len = p - lineStart;
 225    s = (GLubyte *) _mesa_malloc(len + 1);
 226    _mesa_memcpy(s, lineStart, len);
 227    s[len] = 0;
 228
 229    return s;
 230 }
 231
 232
 233 /**
 234  * Initialize a new vertex/fragment program object.
 235  */
 236 static struct gl_program *
 237 _mesa_init_program_struct( GLcontext *ctx, struct gl_program *prog,
 238                            GLenum target, GLuint id)
 239 {
 240    (void) ctx;
 241    if (prog) {
 242       GLuint i;
 243       _mesa_bzero(prog, sizeof(*prog));
 244       prog->Id = id;
 245       prog->Target = target;
 246       prog->Resident = GL_TRUE;
 247       prog->RefCount = 1;
 248       prog->Format = GL_PROGRAM_FORMAT_ASCII_ARB;
 249
 250       /* default mapping from samplers to texture units */
 251       for (i = 0; i < MAX_SAMPLERS; i++)
 252          prog->SamplerUnits[i] = i;
 253    }
 254
 255    return prog;
 256 }
 257
 258
 259 /**
 260  * Initialize a new fragment program object.
 261  */
 262 struct gl_program *
 263 _mesa_init_fragment_program( GLcontext *ctx, struct gl_fragment_program *prog,
 264                              GLenum target, GLuint id)
 265 {
 266    if (prog)
 267       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 268    else
 269       return NULL;
 270 }
 271
 272
 273 /**
 274  * Initialize a new vertex program object.
 275  */
 276 struct gl_program *
 277 _mesa_init_vertex_program( GLcontext *ctx, struct gl_vertex_program *prog,
 278                            GLenum target, GLuint id)
 279 {
 280    if (prog)
 281       return _mesa_init_program_struct( ctx, &prog->Base, target, id );
 282    else
 283       return NULL;
 284 }
 285
 286
 287 /**
 288  * Allocate and initialize a new fragment/vertex program object but
 289  * don't put it into the program hash table.  Called via
 290  * ctx->Driver.NewProgram.  May be overridden (ie. replaced) by a
 291  * device driver function to implement OO deriviation with additional
 292  * types not understood by this function.
 293  *
 294  * \param ctx  context
 295  * \param id   program id/number
 296  * \param target  program target/type
 297  * \return  pointer to new program object
 298  */
 299 struct gl_program *
 300 _mesa_new_program(GLcontext *ctx, GLenum target, GLuint id)
 301 {
 302    struct gl_program *prog;
 303    switch (target) {
 304    case GL_VERTEX_PROGRAM_ARB: /* == GL_VERTEX_PROGRAM_NV */
 305    case GL_VERTEX_STATE_PROGRAM_NV:
 306       prog = _mesa_init_vertex_program(ctx, CALLOC_STRUCT(gl_vertex_program),
 307                                        target, id );
 308       break;
 309    case GL_FRAGMENT_PROGRAM_NV:
 310    case GL_FRAGMENT_PROGRAM_ARB:
 311       prog =_mesa_init_fragment_program(ctx,
 312                                          CALLOC_STRUCT(gl_fragment_program),
 313                                          target, id );
 314       break;
 315    default:
 316       _mesa_problem(ctx, "bad target in _mesa_new_program");
 317       prog = NULL;
 318    }
 319    return prog;
 320 }
 321
 322
 323 /**
 324  * Delete a program and remove it from the hash table, ignoring the
 325  * reference count.
 326  * Called via ctx->Driver.DeleteProgram.  May be wrapped (OO deriviation)
 327  * by a device driver function.
 328  */
 329 void
 330 _mesa_delete_program(GLcontext *ctx, struct gl_program *prog)
 331 {
 332    (void) ctx;
 333    ASSERT(prog);
 334    ASSERT(prog->RefCount==0);
 335
 336    if (prog == &_mesa_DummyProgram)
 337       return;
 338
 339    if (prog->String)
 340       _mesa_free(prog->String);
 341
 342    _mesa_free_instructions(prog->Instructions, prog->NumInstructions);
 343
 344    if (prog->Parameters) {
 345       _mesa_free_parameter_list(prog->Parameters);
 346    }
 347    if (prog->Varying) {
 348       _mesa_free_parameter_list(prog->Varying);
 349    }
 350    if (prog->Attributes) {
 351       _mesa_free_parameter_list(prog->Attributes);
 352    }
 353
 354    _mesa_free(prog);
 355 }
 356
 357
 358 /**
 359  * Return the gl_program object for a given ID.
 360  * Basically just a wrapper for _mesa_HashLookup() to avoid a lot of
 361  * casts elsewhere.
 362  */
 363 struct gl_program *
 364 _mesa_lookup_program(GLcontext *ctx, GLuint id)
 365 {
 366    if (id)
 367       return (struct gl_program *) _mesa_HashLookup(ctx->Shared->Programs, id);
 368    else
 369       return NULL;
 370 }
 371
 372
 373 /**
 374  * Reference counting for vertex/fragment programs
 375  */
 376 void
 377 _mesa_reference_program(GLcontext *ctx,
 378                         struct gl_program **ptr,
 379                         struct gl_program *prog)
 380 {
 381    assert(ptr);
 382    if (*ptr && prog) {
 383       /* sanity check */
 384       if ((*ptr)->Target == GL_VERTEX_PROGRAM_ARB)
 385          ASSERT(prog->Target == GL_VERTEX_PROGRAM_ARB);
 386       else if ((*ptr)->Target == GL_FRAGMENT_PROGRAM_ARB)
 387          ASSERT(prog->Target == GL_FRAGMENT_PROGRAM_ARB ||
 388                 prog->Target == GL_FRAGMENT_PROGRAM_NV);
 389    }
 390    if (*ptr == prog) {
 391       return;  /* no change */
 392    }
 393    if (*ptr) {
 394       GLboolean deleteFlag;
 395
 396       /*_glthread_LOCK_MUTEX((*ptr)->Mutex);*/
 397 #if 0
 398       printf("Program %p ID=%u Target=%s  Refcount-- to %d\n",
 399              *ptr, (*ptr)->Id,
 400              ((*ptr)->Target == GL_VERTEX_PROGRAM_ARB ? "VP" : "FP"),
 401              (*ptr)->RefCount - 1);
 402 #endif
 403       ASSERT((*ptr)->RefCount > 0);
 404       (*ptr)->RefCount--;
 405
 406       deleteFlag = ((*ptr)->RefCount == 0);
 407       /*_glthread_UNLOCK_MUTEX((*ptr)->Mutex);*/
 408
 409       if (deleteFlag) {
 410          ASSERT(ctx);
 411          ctx->Driver.DeleteProgram(ctx, *ptr);
 412       }
 413
 414       *ptr = NULL;
 415    }
 416
 417    assert(!*ptr);
 418    if (prog) {
 419       /*_glthread_LOCK_MUTEX(prog->Mutex);*/
 420       prog->RefCount++;
 421 #if 0
 422       printf("Program %p ID=%u Target=%s  Refcount++ to %d\n",
 423              prog, prog->Id,
 424              (prog->Target == GL_VERTEX_PROGRAM_ARB ? "VP" : "FP"),
 425              prog->RefCount);
 426 #endif
 427       /*_glthread_UNLOCK_MUTEX(prog->Mutex);*/
 428    }
 429
 430    *ptr = prog;
 431 }
 432
 433
 434 /**
 435  * Return a copy of a program.
 436  * XXX Problem here if the program object is actually OO-derivation
 437  * made by a device driver.
 438  */
 439 struct gl_program *
 440 _mesa_clone_program(GLcontext *ctx, const struct gl_program *prog)
 441 {
 442    struct gl_program *clone;
 443
 444    clone = ctx->Driver.NewProgram(ctx, prog->Target, prog->Id);
 445    if (!clone)
 446       return NULL;
 447
 448    assert(clone->Target == prog->Target);
 449    assert(clone->RefCount == 1);
 450
 451    clone->String = (GLubyte *) _mesa_strdup((char *) prog->String);
 452    clone->Format = prog->Format;
 453    clone->Instructions = _mesa_alloc_instructions(prog->NumInstructions);
 454    if (!clone->Instructions) {
 455       _mesa_reference_program(ctx, &clone, NULL);
 456       return NULL;
 457    }
 458    _mesa_copy_instructions(clone->Instructions, prog->Instructions,
 459                            prog->NumInstructions);
 460    clone->InputsRead = prog->InputsRead;
 461    clone->OutputsWritten = prog->OutputsWritten;
 462    clone->SamplersUsed = prog->SamplersUsed;
 463    clone->ShadowSamplers = prog->ShadowSamplers;
 464    memcpy(clone->TexturesUsed, prog->TexturesUsed, sizeof(prog->TexturesUsed));
 465
 466    if (prog->Parameters)
 467       clone->Parameters = _mesa_clone_parameter_list(prog->Parameters);
 468    memcpy(clone->LocalParams, prog->LocalParams, sizeof(clone->LocalParams));
 469    if (prog->Varying)
 470       clone->Varying = _mesa_clone_parameter_list(prog->Varying);
 471    if (prog->Attributes)
 472       clone->Attributes = _mesa_clone_parameter_list(prog->Attributes);
 473    memcpy(clone->LocalParams, prog->LocalParams, sizeof(clone->LocalParams));
 474    clone->NumInstructions = prog->NumInstructions;
 475    clone->NumTemporaries = prog->NumTemporaries;
 476    clone->NumParameters = prog->NumParameters;
 477    clone->NumAttributes = prog->NumAttributes;
 478    clone->NumAddressRegs = prog->NumAddressRegs;
 479    clone->NumNativeInstructions = prog->NumNativeInstructions;
 480    clone->NumNativeTemporaries = prog->NumNativeTemporaries;
 481    clone->NumNativeParameters = prog->NumNativeParameters;
 482    clone->NumNativeAttributes = prog->NumNativeAttributes;
 483    clone->NumNativeAddressRegs = prog->NumNativeAddressRegs;
 484    clone->NumAluInstructions = prog->NumAluInstructions;
 485    clone->NumTexInstructions = prog->NumTexInstructions;
 486    clone->NumTexIndirections = prog->NumTexIndirections;
 487    clone->NumNativeAluInstructions = prog->NumNativeAluInstructions;
 488    clone->NumNativeTexInstructions = prog->NumNativeTexInstructions;
 489    clone->NumNativeTexIndirections = prog->NumNativeTexIndirections;
 490
 491    switch (prog->Target) {
 492    case GL_VERTEX_PROGRAM_ARB:
 493       {
 494          const struct gl_vertex_program *vp
 495             = (const struct gl_vertex_program *) prog;
 496          struct gl_vertex_program *vpc = (struct gl_vertex_program *) clone;
 497          vpc->IsPositionInvariant = vp->IsPositionInvariant;
 498          vpc->IsNVProgram = vp->IsNVProgram;
 499       }
 500       break;
 501    case GL_FRAGMENT_PROGRAM_ARB:
 502       {
 503          const struct gl_fragment_program *fp
 504             = (const struct gl_fragment_program *) prog;
 505          struct gl_fragment_program *fpc = (struct gl_fragment_program *) clone;
 506          fpc->FogOption = fp->FogOption;
 507          fpc->UsesKill = fp->UsesKill;
 508       }
 509       break;
 510    default:
 511       _mesa_problem(NULL, "Unexpected target in _mesa_clone_program");
 512    }
 513
 514    return clone;
 515 }
 516
 517
 518 /**
 519  * Insert 'count' NOP instructions at 'start' in the given program.
 520  * Adjust branch targets accordingly.
 521  */
 522 GLboolean
 523 _mesa_insert_instructions(struct gl_program *prog, GLuint start, GLuint count)
 524 {
 525    const GLuint origLen = prog->NumInstructions;
 526    const GLuint newLen = origLen + count;
 527    struct prog_instruction *newInst;
 528    GLuint i;
 529
 530    /* adjust branches */
 531    for (i = 0; i < prog->NumInstructions; i++) {
 532       struct prog_instruction *inst = prog->Instructions + i;
 533       if (inst->BranchTarget > 0) {
 534          if ((GLuint)inst->BranchTarget >= start) {
 535             inst->BranchTarget += count;
 536          }
 537       }
 538    }
 539
 540    /* Alloc storage for new instructions */
 541    newInst = _mesa_alloc_instructions(newLen);
 542    if (!newInst) {
 543       return GL_FALSE;
 544    }
 545
 546    /* Copy 'start' instructions into new instruction buffer */
 547    _mesa_copy_instructions(newInst, prog->Instructions, start);
 548
 549    /* init the new instructions */
 550    _mesa_init_instructions(newInst + start, count);
 551
 552    /* Copy the remaining/tail instructions to new inst buffer */
 553    _mesa_copy_instructions(newInst + start + count,
 554                            prog->Instructions + start,
 555                            origLen - start);
 556
 557    /* free old instructions */
 558    _mesa_free_instructions(prog->Instructions, origLen);
 559
 560    /* install new instructions */
 561    prog->Instructions = newInst;
 562    prog->NumInstructions = newLen;
 563
 564    return GL_TRUE;
 565 }
 566
 567 /**
 568  * Delete 'count' instructions at 'start' in the given program.
 569  * Adjust branch targets accordingly.
 570  */
 571 GLboolean
 572 _mesa_delete_instructions(struct gl_program *prog, GLuint start, GLuint count)
 573 {
 574    const GLuint origLen = prog->NumInstructions;
 575    const GLuint newLen = origLen - count;
 576    struct prog_instruction *newInst;
 577    GLuint i;
 578
 579    /* adjust branches */
 580    for (i = 0; i < prog->NumInstructions; i++) {
 581       struct prog_instruction *inst = prog->Instructions + i;
 582       if (inst->BranchTarget > 0) {
 583          if (inst->BranchTarget > (GLint) start) {
 584             inst->BranchTarget -= count;
 585          }
 586       }
 587    }
 588
 589    /* Alloc storage for new instructions */
 590    newInst = _mesa_alloc_instructions(newLen);
 591    if (!newInst) {
 592       return GL_FALSE;
 593    }
 594
 595    /* Copy 'start' instructions into new instruction buffer */
 596    _mesa_copy_instructions(newInst, prog->Instructions, start);
 597
 598    /* Copy the remaining/tail instructions to new inst buffer */
 599    _mesa_copy_instructions(newInst + start,
 600                            prog->Instructions + start + count,
 601                            newLen - start);
 602
 603    /* free old instructions */
 604    _mesa_free_instructions(prog->Instructions, origLen);
 605
 606    /* install new instructions */
 607    prog->Instructions = newInst;
 608    prog->NumInstructions = newLen;
 609
 610    return GL_TRUE;
 611 }
 612
 613
 614 /**
 615  * Search instructions for registers that match (oldFile, oldIndex),
 616  * replacing them with (newFile, newIndex).
 617  */
 618 static void
 619 replace_registers(struct prog_instruction *inst, GLuint numInst,
 620                   GLuint oldFile, GLuint oldIndex,
 621                   GLuint newFile, GLuint newIndex)
 622 {
 623    GLuint i, j;
 624    for (i = 0; i < numInst; i++) {
 625       /* src regs */
 626       for (j = 0; j < _mesa_num_inst_src_regs(inst->Opcode); j++) {
 627          if (inst[i].SrcReg[j].File == oldFile &&
 628              inst[i].SrcReg[j].Index == oldIndex) {
 629             inst[i].SrcReg[j].File = newFile;
 630             inst[i].SrcReg[j].Index = newIndex;
 631          }
 632       }
 633       /* dst reg */
 634       if (inst[i].DstReg.File == oldFile && inst[i].DstReg.Index == oldIndex) {
 635          inst[i].DstReg.File = newFile;
 636          inst[i].DstReg.Index = newIndex;
 637       }
 638    }
 639 }
 640
 641
 642 /**
 643  * Search instructions for references to program parameters.  When found,
 644  * increment the parameter index by 'offset'.
 645  * Used when combining programs.
 646  */
 647 static void
 648 adjust_param_indexes(struct prog_instruction *inst, GLuint numInst,
 649                      GLuint offset)
 650 {
 651    GLuint i, j;
 652    for (i = 0; i < numInst; i++) {
 653       for (j = 0; j < _mesa_num_inst_src_regs(inst->Opcode); j++) {
 654          GLuint f = inst[i].SrcReg[j].File;
 655          if (f == PROGRAM_CONSTANT ||
 656              f == PROGRAM_UNIFORM ||
 657              f == PROGRAM_STATE_VAR) {
 658             inst[i].SrcReg[j].Index += offset;
 659          }
 660       }
 661    }
 662 }
 663
 664
 665 /**
 666  * Combine two programs into one.  Fix instructions so the outputs of
 667  * the first program go to the inputs of the second program.
 668  */
 669 struct gl_program *
 670 _mesa_combine_programs(GLcontext *ctx,
 671                        const struct gl_program *progA,
 672                        const struct gl_program *progB)
 673 {
 674    struct prog_instruction *newInst;
 675    struct gl_program *newProg;
 676    const GLuint lenA = progA->NumInstructions - 1; /* omit END instr */
 677    const GLuint lenB = progB->NumInstructions;
 678    const GLuint numParamsA = _mesa_num_parameters(progA->Parameters);
 679    const GLuint newLength = lenA + lenB;
 680    GLboolean usedTemps[MAX_PROGRAM_TEMPS];
 681    GLuint firstTemp = 0;
 682    GLbitfield inputsB;
 683    GLuint i;
 684
 685    ASSERT(progA->Target == progB->Target);
 686
 687    newInst = _mesa_alloc_instructions(newLength);
 688    if (!newInst)
 689       return GL_FALSE;
 690
 691    _mesa_copy_instructions(newInst, progA->Instructions, lenA);
 692    _mesa_copy_instructions(newInst + lenA, progB->Instructions, lenB);
 693
 694    /* adjust branch / instruction addresses for B's instructions */
 695    for (i = 0; i < lenB; i++) {
 696       newInst[lenA + i].BranchTarget += lenA;
 697    }
 698
 699    newProg = ctx->Driver.NewProgram(ctx, progA->Target, 0);
 700    newProg->Instructions = newInst;
 701    newProg->NumInstructions = newLength;
 702
 703    /* find used temp regs (we may need new temps below) */
 704    _mesa_find_used_registers(newProg, PROGRAM_TEMPORARY,
 705                              usedTemps, MAX_PROGRAM_TEMPS);
 706
 707    if (newProg->Target == GL_FRAGMENT_PROGRAM_ARB) {
 708       struct gl_fragment_program *fprogA, *fprogB, *newFprog;
 709       GLbitfield progB_inputsRead = progB->InputsRead;
 710       GLint progB_colorFile, progB_colorIndex;
 711
 712       fprogA = (struct gl_fragment_program *) progA;
 713       fprogB = (struct gl_fragment_program *) progB;
 714       newFprog = (struct gl_fragment_program *) newProg;
 715
 716       newFprog->UsesKill = fprogA->UsesKill || fprogB->UsesKill;
 717
 718       /* We'll do a search and replace for instances
 719        * of progB_colorFile/progB_colorIndex below...
 720        */
 721       progB_colorFile = PROGRAM_INPUT;
 722       progB_colorIndex = FRAG_ATTRIB_COL0;
 723
 724       /*
 725        * The fragment program may get color from a state var rather than
 726        * a fragment input (vertex output) if it's constant.
 727        * See the texenvprogram.c code.
 728        * So, search the program's parameter list now to see if the program
 729        * gets color from a state var instead of a conventional fragment
 730        * input register.
 731        */
 732       for (i = 0; i < progB->Parameters->NumParameters; i++) {
 733          struct gl_program_parameter *p = &progB->Parameters->Parameters[i];
 734          if (p->Type == PROGRAM_STATE_VAR &&
 735              p->StateIndexes[0] == STATE_INTERNAL &&
 736              p->StateIndexes[1] == STATE_CURRENT_ATTRIB &&
 737              p->StateIndexes[2] == VERT_ATTRIB_COLOR0) {
 738             progB_inputsRead |= FRAG_BIT_COL0;
 739             progB_colorFile = PROGRAM_STATE_VAR;
 740             progB_colorIndex = i;
 741             break;
 742          }
 743       }
 744
 745       /* Connect color outputs of fprogA to color inputs of fprogB, via a
 746        * new temporary register.
 747        */
 748       if ((progA->OutputsWritten & (1 << FRAG_RESULT_COLOR)) &&
 749           (progB_inputsRead & FRAG_BIT_COL0)) {
 750          GLint tempReg = _mesa_find_free_register(usedTemps, MAX_PROGRAM_TEMPS,
 751                                                   firstTemp);
 752          if (tempReg < 0) {
 753             _mesa_problem(ctx, "No free temp regs found in "
 754                           "_mesa_combine_programs(), using 31");
 755             tempReg = 31;
 756          }
 757          firstTemp = tempReg + 1;
 758
 759          /* replace writes to result.color[0] with tempReg */
 760          replace_registers(newInst, lenA,
 761                            PROGRAM_OUTPUT, FRAG_RESULT_COLOR,
 762                            PROGRAM_TEMPORARY, tempReg);
 763          /* replace reads from the input color with tempReg */
 764          replace_registers(newInst + lenA, lenB,
 765                            progB_colorFile, progB_colorIndex, /* search for */
 766                            PROGRAM_TEMPORARY, tempReg  /* replace with */ );
 767       }
 768
 769       /* compute combined program's InputsRead */
 770       inputsB = progB_inputsRead;
 771       if (progA->OutputsWritten & (1 << FRAG_RESULT_COLOR)) {
 772          inputsB &= ~(1 << FRAG_ATTRIB_COL0);
 773       }
 774       newProg->InputsRead = progA->InputsRead | inputsB;
 775       newProg->OutputsWritten = progB->OutputsWritten;
 776       newProg->SamplersUsed = progA->SamplersUsed | progB->SamplersUsed;
 777    }
 778    else {
 779       /* vertex program */
 780       assert(0);      /* XXX todo */
 781    }
 782
 783    /*
 784     * Merge parameters (uniforms, constants, etc)
 785     */
 786    newProg->Parameters = _mesa_combine_parameter_lists(progA->Parameters,
 787                                                        progB->Parameters);
 788
 789    adjust_param_indexes(newInst + lenA, lenB, numParamsA);
 790
 791
 792    return newProg;
 793 }
 794
 795
 796 /**
 797  * Populate the 'used' array with flags indicating which registers (TEMPs,
 798  * INPUTs, OUTPUTs, etc, are used by the given program.
 799  * \param file  type of register to scan for
 800  * \param used  returns true/false flags for in use / free
 801  * \param usedSize  size of the 'used' array
 802  */
 803 void
 804 _mesa_find_used_registers(const struct gl_program *prog,
 805                           gl_register_file file,
 806                           GLboolean used[], GLuint usedSize)
 807 {
 808    GLuint i, j;
 809
 810    _mesa_memset(used, 0, usedSize);
 811
 812    for (i = 0; i < prog->NumInstructions; i++) {
 813       const struct prog_instruction *inst = prog->Instructions + i;
 814       const GLuint n = _mesa_num_inst_src_regs(inst->Opcode);
 815
 816       if (inst->DstReg.File == file) {
 817          used[inst->DstReg.Index] = GL_TRUE;
 818       }
 819
 820       for (j = 0; j < n; j++) {
 821          if (inst->SrcReg[j].File == file) {
 822             used[inst->SrcReg[j].Index] = GL_TRUE;
 823          }
 824       }
 825    }
 826 }
 827
 828
 829 /**
 830  * Scan the given 'used' register flag array for the first entry
 831  * that's >= firstReg.
 832  * \param used  vector of flags indicating registers in use (as returned
 833  *              by _mesa_find_used_registers())
 834  * \param usedSize  size of the 'used' array
 835  * \param firstReg  first register to start searching at
 836  * \return index of unused register, or -1 if none.
 837  */
 838 GLint
 839 _mesa_find_free_register(const GLboolean used[],
 840                          GLuint usedSize, GLuint firstReg)
 841 {
 842    GLuint i;
 843
 844    assert(firstReg < usedSize);
 845
 846    for (i = firstReg; i < usedSize; i++)
 847       if (!used[i])
 848          return i;
 849
 850    return -1;
 851 }
 852
 853
 854 /**
 855  * "Post-process" a GPU program.  This is intended to be used for debugging.
 856  * Example actions include no-op'ing instructions or changing instruction
 857  * behaviour.
 858  */
 859 void
 860 _mesa_postprocess_program(GLcontext *ctx, struct gl_program *prog)
 861 {
 862    static const GLfloat white[4] = { 0.5, 0.5, 0.5, 0.5 };
 863    GLuint i;
 864    GLuint whiteSwizzle;
 865    GLint whiteIndex = _mesa_add_unnamed_constant(prog->Parameters,
 866                                                  white, 4, &whiteSwizzle);
 867
 868    (void) whiteIndex;
 869
 870    for (i = 0; i < prog->NumInstructions; i++) {
 871       struct prog_instruction *inst = prog->Instructions + i;
 872       const GLuint n = _mesa_num_inst_src_regs(inst->Opcode);
 873
 874       (void) n;
 875
 876       if (_mesa_is_tex_instruction(inst->Opcode)) {
 877 #if 0
 878          /* replace TEX/TXP/TXB with MOV */
 879          inst->Opcode = OPCODE_MOV;
 880          inst->DstReg.WriteMask = WRITEMASK_XYZW;
 881          inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
 882          inst->SrcReg[0].Negate = NEGATE_NONE;
 883 #endif
 884
 885 #if 0
 886          /* disable shadow texture mode */
 887          inst->TexShadow = 0;
 888 #endif
 889       }
 890
 891       if (inst->Opcode == OPCODE_TXP) {
 892 #if 0
 893          inst->Opcode = OPCODE_MOV;
 894          inst->DstReg.WriteMask = WRITEMASK_XYZW;
 895          inst->SrcReg[0].File = PROGRAM_CONSTANT;
 896          inst->SrcReg[0].Index = whiteIndex;
 897          inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
 898          inst->SrcReg[0].Negate = NEGATE_NONE;
 899 #endif
 900 #if 0
 901          inst->TexShadow = 0;
 902 #endif
 903 #if 0
 904          inst->Opcode = OPCODE_TEX;
 905          inst->TexShadow = 0;
 906 #endif
 907       }
 908
 909    }
 910 }