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